CAFE 0.9.7 leaderboard (earlier release)

0.9.7 did not pass the test fixed beforehand for CAFE 1.0 (Amendment 13 item 9: 61 resolved pairs with the datasets resampled, 56 needed; A signal-to-noise 4.04, 6 needed), so 0.9.5 stays the CAFE 1.0 candidate. The test · all versions compared.

Main board

pj_better_10m ◊ is first and alone in tier 1 under the tier rule. It is not resolved against Mantis. Its leads over pj_better_3p7m ◊, pjgood-warp-xview-combo-10M ◊, Chronos-2 are within seed noise. With the datasets resampled, 6 rows share tier 1. On A, the no-encoder ZERO row (0.638) scores above 11 of the 15 ranked rows; only pj_better_10m ◊, pj_better_3p7m ◊, pjgood-warp-xview-combo-10M ◊, pj_better_1m ◊ score above it, all of them Forgis rows. Without A (C and FE only), tier 1 is pj_better_10m ◊, Chronos-2, Mantis and the ◊ ranks move: pj_better_3p7m ◊ 2 to 4; pjgood-warp-xview-combo-10M ◊ 3 to 5. The board rests on 24 C, 10 A and 15 FE datasets. Main board: one readout per model.

Forgis runs CAFE and developed 5 of the 15 ranked models, marked ◊, with CAFE test scores visible during their development. They get no special colour here. You can take them out of every table with the switch above the board. Their weights are not public, so outsiders cannot rerun them. Forgis staff chose the 0.9.7 drops and additions with the 0.9.6 board known; six drops were informed by scores (C 3W, bosch, soredd_abc; FE ngafid_hastall, substorm_omni, epa_aqs_pm25), and the two C additions were accepted after a screening that scored a ◊ row (pjgood-warp-xview-combo-10M) against its random-init twin on them (see the card). ◊ rows: Forgis weights, not public checkpoints. Other variants scored on CAFE or PrecursorBench Eval data before each was submitted (lower bounds, see Known limitations): LeSaturn-warppool-factorized-7.4M 43, pjgood-warp-xview-combo-10M 59, pj_better_3p7m and pj_better_10m 64, and pj_better_1m 65.

Task

Readout

Rows Baselines and references Tabular rows Forgis rows ◊

Dataset groups

◊ Forgis row. Kind tells unranked rows apart: baseline, reference, tabular. Click a column to sort. The glossary near the end explains every term.

Is 0.9.7 more separable than 0.9.6?

0.9.7 resolves 61 of 105 pairs with the datasets resampled, against 51 for 0.9.6 and 49 for 0.9.5 and its A signal-to-noise is 4.04. The test of Amendment 13 item 9 needs at least 56 pairs and an A signal-to-noise of at least 6. 0.9.7 fails it (an A signal-to-noise of 4.04 is below 6), so 0.9.5 stays the CAFE 1.0 candidate. In either case CAFE 1.0 also needs two fresh audits in a row with no blocker, and Jonas Petersen's explicit go.

Read the candidacy with this: 0.9.5 contains 5 A datasets that the admission rules of Amendment 13 reject (CARE_A, CARE_C, RIGS, W3 and SKAB), and it counts ESA_M1/ESA_M2 and SMAP/MSL as separate datasets, which item 2 now pools. Whether 1.0 may be a version that fails its successor's admission rules is for Jonas Petersen to decide before any 1.0 go.

The goal for the A signal-to-noise was 9; it was not met (4.04).

The pair count was fixed before any 0.9.7 cell existed (Amendment 13 item 9); the reading of the A signal-to-noise used here was fixed in the PREREG09 note of 2026-11-05 19:02Z, after the first 0.9.7 A cells existed and before any was read (the amendment's literal wording, spread SD over pair noise, is in the table below; both readings fail): the 105 pairs of the 15 ranked rows of 0.9.5, on the CAFE score, with the datasets resampled within each task as well as the test material (10,000 draws, Holm), by the same code for every version. Holding the datasets fixed, as the main board does, would narrow the intervals of a bigger version even if its new datasets added nothing; with the datasets resampled, a dataset that orders the models differently from the others widens them. Repeating the resampling with other seeds moves the count by about 1.0 pairs (SD over 6 seeds); the published count is from the preregistered seed, and the 5 other seeds give 58 to 60 (mean over all 6: 59.3), all at or above the pass mark of the count. The A signal-to-noise is the mean absolute gap between two ranked rows on the A score divided by the median bootstrap SD of that gap (separability096.py, as on the 0.9.6 card; PREREG09 note of 2026-11-05 19:02Z); spread SD over pair noise is shown beside it.

pairs resolved, of 105, and A signal-to-noise	0.9.5	0.9.6	0.9.7
CAFE score, datasets resampled (the measure)	49	51	61
CAFE score, datasets fixed (main board)	73	73	85
of which within seed noise	1	1	10
C board, datasets fixed	70	74	80
A board, datasets fixed	18	37	60
FE board, datasets fixed	87	87	87
C board, datasets resampled	15	21	25
A board, datasets resampled	4	6	9
FE board, datasets resampled	67	66	67
A signal-to-noise (the test's measure)	2.48	3.00	4.04
A spread SD over pair noise	2.23	2.65	3.54

Sensitivities of the test raised by the audits (reporting only; the board above is the result):

- With every A dataset resampled by recording (the blocks of one recording form one cluster): 60 pairs resolved and an A signal-to-noise of 4.08; the test fails (test097_cr.json). No rank changes. - With KUKA with its first 100 Eval series in hash order instead of 255 (the cap of 100 for several series per asset): 61 pairs resolved and an A signal-to-noise of 4.05; the test fails (test097_k100.json). No rank changes. - With ipshocks_fluxes on the FE board (check 3 lowered to 900 events; PREREG09 notes of 2026-11-06): 58 pairs resolved and an A signal-to-noise of 4.04; the test fails (test097_ips.json). Ranks that change: pj_better_3p7m ◊ 2 to 3; pjgood-warp-xview-combo-10M ◊ 3 to 2; Timestral-28M 13 to 14; Timestral-5M 14 to 13.

On the datasets-resampled boards of 0.9.5 and 0.9.7 no pair is resolved in opposite directions.

On the datasets-resampled boards of 0.9.6 and 0.9.7 no pair is resolved in opposite directions.

No pair of ranked rows is resolved in opposite directions on the main boards of 0.9.6 and 0.9.7, on the CAFE score or on any task board.

Read this beside the conflict-of-interest disclosure in What changed in 0.9.7: Forgis staff chose the drops and adds with the 0.9.6 board known, and Forgis develops the ◊ rows.

What changed in 0.9.7

Amendment 13 was written into PREREG09 before any 0.9.7 download, build or cell, with the whole 0.9.6 board known. The dated notes after it each give their own time and what had been computed or read when they were written; the admission notes came before any 0.9.7 cell on the datasets they admit (the two C additions had screening scores before that, see New datasets), and some later notes came after 0.9.7 cells or the 0.9.7 board existed (see the provenance table under Rules). Each model keeps the readout chosen for it in 0.9.5; the C head is the one of Amendment 12; A is scored at w = 256, k = 5, centre assignment, as in 0.9.6.

15 datasets of 0.9.6 are not on the 0.9.7 board and 6 are new, plus 2 pooled from 0.9.6 datasets. C has 24 datasets, A 10 and FE 15 (0.9.6: 25, 14, 17).

task	dataset	change	reason, as recorded
C	mv06/3w	removed or pooled	Amendment 13 item 6
C	mv06/bosch	removed or pooled	Amendment 13 item 6
C	nicola15/soredd_abc	removed or pooled	Amendment 13 item 6
C	nicola15/bdg2_meter6	added	rebuilt from the exp-nicola-15 builder; X sha256 equals LOG.md (item 6)
C	nicola15/robo6_rb6	added	rebuilt from the exp-nicola-15 builder; X sha256 equals LOG.md (item 6)
A	ESA_M1	removed or pooled	pooled into ESA (Amendment 13 item 2)
A	ESA_M2	removed or pooled	pooled into ESA (Amendment 13 item 2)
A	CARE_A	removed or pooled	Amendment 13 item 3: labels cover most of each series, fewer than 15 Eval series (rule 2(a), (c))
A	SMAP	removed or pooled	pooled into NASA (Amendment 13 item 2)
A	MSL	removed or pooled	pooled into NASA (Amendment 13 item 2)
A	CARE_C	removed or pooled	Amendment 13 item 3: labels cover most of each series, fewer than 15 Eval series (rule 2(a), (c))
A	RIGS	removed or pooled	Amendment 13 item 3: 5 Eval series, 7 segments (rule 2(a), (b))
A	W3	removed or pooled	Amendment 13 item 3: median anomaly share 0.39 (rule 2(c)); a time-only score reaches 1.000
A	SKAB	removed or pooled	Amendment 13 item 3: 8 Eval series, label to end of file (rule 2(a))
A	ESA	pooled from 0.9.6 datasets	items 2(a)-(c) hold: 95 Eval, 97 segments, median share 0.032; elapsed-time best 0.611 (gate applies to new datasets only)
A	NASA	pooled from 0.9.6 datasets	items 2(a)-(c) hold: 37 Eval, 47 segments, median share 0.040; elapsed-time best 0.516 (gate applies to new datasets only)
A	UAVSEAD	added	items 2(a)-(c) hold: 101 Eval, 147 segments, median share 0.083; elapsed-time best 0.585 (gate 0.65 holds)
A	UR5E	added	items 2(a)-(c) hold: 63 Eval, 74 segments, median share 0.191; elapsed-time best 0.616 (gate 0.65 holds)
A	KUKA	added	items 2(a)-(c) hold: 255 Eval, 534 segments, median share 0.108; elapsed-time best 0.519 (gate 0.65 holds)
FE	epa_aqs_pm25	removed or pooled	Amendment 13 item 5
FE	substorm_omni	removed or pooled	Amendment 13 item 5
FE	ngafid_hastall	removed or pooled	Amendment 13 item 5
FE	heatpump_defrost	added	item 5 gates hold: 53,978 events >= 300; 113 systems >= 30; time since last event 0.659 <= 0.796; time of day 0.566 <= 0.585; best free forecast 0.729 <= 0.85

Datasets of one release or paper are one dataset (Amendment 13 item 2); their series are pooled and the dataset score is the mean over all pooled Eval series: ESA = ESA_M1 + ESA_M2 (95 Eval series); NASA = SMAP + MSL (37 Eval series).

The cap on series per A dataset was raised from 35 (Amendment 13 item 2): 300 for sources whose series are independent physical runs, 100 for sources with several series per asset; Tuning is the first ceil(0.15 m) in hash order, as before.

A dataset	cap	Eval series	Tuning series	kind
ESA	100	95	18	existing, pooled ESA_M1 + ESA_M2 (item 2)
NASA	100	37	4	existing, pooled SMAP + MSL (item 2)
UKWIND	100	17	4	existing, unchanged (never limited by the cap of 35)
HOT	100	50	9	existing, re-selected under cap 100
PREDIST	100	41	8	existing, re-selected under cap 100
NOBOOM	300	20	4	existing, re-selected under cap 300 (lists unchanged)
HAI	100	58	11	existing, pooled with the other HAI releases (item 4)
UAVSEAD	300	101	18	new, firm
UR5E	300	63	12	new, firm
KUKA	300	255	45	new, second wave

NOBOOM is recorded with cap 300, although one distillation plant would fall under the cap of 100 for several series per asset; with 20 Eval series neither cap binds, so its list is the same either way.

A candidates not admitted, with the reason recorded before any cell on them (full measurements under Admission):

candidate	status	kind	reason
RFLYMAD	out	new, conditional	item 4 elapsed-time gate: 0.731 > 0.65
PMU	out	new, conditional	the continuous 1 s TransnetBW record (Zenodo 15784548) is one channel (CE frequency) and ends 2024-04-25, covering 23 of 88 cases; P4 needs >= 2 channels and 23 cases give at most 19 Eval series and 19 segments (rule 2(b)); not built
ONBOARDEU	out	new, conditional	item 4 elapsed-time gate: 0.708 > 0.65
TCABR	out	new, conditional	item 4 elapsed-time gate: 1.000 > 0.65
BASIC	out	new, conditional	check 1 / item 4 condition: the flights are software-in-the-loop, not real (Ahmad and Akram, Data in Brief 2024, doi:10.1016/j.dib.2024.110069: 'The BASiC dataset is exclusively derived from simulations conducted in a SITL environment using the ArduPilot platform'); not built
EAWAG	out	new, second wave	only 4 ponds (2D, 3D, 5A, 5B), phase 1, carry the expert 'Anomaly' column, with 5 segments in total: rules 2(a) and 2(b) cannot hold; the licence field is empty (portal: public domain for data outside Swiss copyright)
CAFUC2	out	new, second wave	check 2 (physical label; PREREG09 note of 2026-10-31 ~20:50Z: injected or staged faults pass when the onset is observed in the system): CAFUC2's faults are synthetic signals written into five copies of each recorded flight after the flight; no aircraft experienced them, so the label is not a state of or an event in the physical system. Decided before any cell; reversible by Jonas (needs-you)
GREEKWIND	out	new, second wave	2(a): 7 Eval series < 15; 2(b): 8 anomaly segments < 20
CRANFIELD	out	new, second wave	the official download (MATLAB File Exchange #50938) needs a MathWorks sign-in; the other copies (Kaggle, IEEE DataPort) need a login; item 4 / Amendment 13: sets that need a registration are not added
MAST	out	new, second wave	no precursor-only cut without an elapsed-time leak: a scored block that ends before the current quench carries the precursor label at its end (a rising elapsed-time score reaches 0.933 on synthetic series of that shape); TokaMark ships no disruption labels or quench times

The C candidate nicola15/sns_hvcm_b_end had to show a pretraining effect on its own first cells (Amendment 13 item 6): pjgood-warp-xview-combo-10M minus the random-init LeSaturn row (3.7 M parameters, another architecture; not a matched twin of the 10.6 M combo) positive in at least 4 of 5 repeats and a mean of at least +0.02. It was positive in 3 and the mean was +0.0169, so it is out. The gate was not set blind: the exp-nicola-15 screening of the same morning had already scored combo against its random-init twin on this dataset (+0.026 ± 0.021, 4 of 5 repeats; LOG.md), a result that would have passed it. The rule is score-informed; it is computed for every C dataset under Admission.

nicola15/bdg2_meter6 and nicola15/robo6_rb6 were admitted on the sha256 check of item 6 (the rebuilt X equals the one recorded in the screening). That check reads no score, but the screening that accepted them earlier on 2026-11-05 did: it scored pjgood-warp-xview-combo-10M ◊ against its random-init twin on these builds: nicola15/bdg2_meter6 0.649 against 0.597, difference +0.053 ± 0.010, 5 of 5 repeats; nicola15/robo6_rb6 0.603 against 0.501, difference +0.101 ± 0.013, 5 of 5 repeats (exp-nicola-15 LOG.md), so these additions were score-informed too. The combo-minus-twin numbers of 0.9.7 are shown under Admission.

CAFUC2 (general-aviation sorties) is out (2026-11-05T18:56Z): check 2 (physical label; PREREG09 note of 2026-10-31 ~20:50Z: injected or staged faults pass when the onset is observed in the system): CAFUC2's faults are synthetic signals written into five copies of each recorded flight after the flight; no aircraft experienced them, so the label is not a state of or an event in the physical system. Decided before any cell; reversible by Jonas (needs-you). It would otherwise pass items 2 and 4 (96 Eval series, 286 segments, median share 0.021, elapsed-time best 0.508). Only Jonas Petersen can reverse it, in a dated note.

FE admission is model-free (Amendment 13 item 5): at least 300 events and 30 independent units or days, a time-since-last-event row at most 0.7956 (the highest on the 14 kept FE datasets, nasa_battery_rw), a time-of-day row at most 0.5851 (towie_genbearing_hot), and the best free forecast at most 0.85. The cap on lab-friction sets is three.

Everything else is as in 0.9.6.

Conflict of interest (Amendment 13 item 1, as written before the run): Forgis owns the ◊ rows and Forgis staff chose these rules and drops with the board known. The drops of FE substorm_omni, epa_aqs_pm25 and ngafid_hastall and of C bosch and soredd_abc were informed by scores (separability and combo minus twin) and are score-informed selection in the sense of Amendment 3. The A and 3W drops follow the structural rules of item 2 and the 0.9.5 evidence (3W: Amendment 12 era decision of exp-nicola-15). The thresholds of item 2 (15 series, 20 segments, share 0.3) were placed in gaps seen in the 0.9.6 data (kept datasets reach a share of 0.21; the five removed ones start at 0.35), so they are post-hoc cut-offs, not independent ones. The card shows the 0.9.6 board beside 0.9.7, the effect of every drop and add on every row, and the ◊ lead under each.

Note added 2026-11-06: the 3W drop called structural here was made on combo minus twin (item 6 and the exp-nicola-15 screening), so six drops were score-informed, and the two C additions were accepted after a screening that scored a ◊ row on them.

The mean lead of the ◊ rows over the other ranked rows, on the CAFE score, under each board (reported, part of no rule):

board	mean ◊ row minus mean other ranked row (CAFE)	best ◊ row minus best other row (CAFE)
0.9.6 board	+0.0262	+0.0034
0.9.6 with the drops of 0.9.7 and no additions (sensitivity `drops_only`)	+0.0308	+0.0073
0.9.7 board	+0.0284	+0.0057

Every ranked row on the 0.9.6 datasets (computed by the 0.9.7 code), on those datasets with the drops of 0.9.7 only (no addition, no pooling), and on 0.9.7: CAFE score, rank and tier. The 5 largest gains from the drops alone are exactly the ◊ rows.

model	0.9.6 datasets	rank	tier	drops only	change	rank	tier	0.9.7	rank	tier
pj_better_10m ◊	0.7363	1	1	0.7480	+0.0117	1	1	0.7384	1	1
pj_better_3p7m ◊	0.7347	2	1	0.7450	+0.0103	2	1	0.7341	2	2
pjgood-warp-xview-combo-10M ◊	0.7326	4	1	0.7438	+0.0112	3	1	0.7340	3	2
Mantis	0.7329	3	1	0.7407	+0.0078	4	1	0.7326	4	2
Chronos-2	0.7301	5	1	0.7389	+0.0087	5	1	0.7301	5	2
LeSaturn-warppool-factorized-7.4M ◊	0.7270	6	2	0.7386	+0.0116	6	1	0.7275	6	3
pj_better_1m ◊	0.7224	7	2	0.7315	+0.0091	7	2	0.7241	7	3
MantisV2	0.7192	8	2	0.7275	+0.0083	8	2	0.7181	8	4
NuTime	0.7067	9	3	0.7108	+0.0041	9	3	0.7050	9	5
Timestral-209M	0.7004	10	3	0.7062	+0.0057	10	3	0.6989	10	5
Timestral-123M	0.6990	11	3	0.7045	+0.0055	11	3	0.6969	11	6
Toto-Open-Base-1.0	0.6972	12	3	0.7031	+0.0058	12	3	0.6954	12	6
Timestral-28M	0.6965	13	3	0.7023	+0.0057	14	3	0.6943	13	6
Timestral-5M	0.6965	14	3	0.7024	+0.0059	13	3	0.6943	14	6
TimesFM-2.5-200M	0.6657	15	4	0.6699	+0.0042	15	4	0.6662	15	7

On the 1.0 measure (pairs resolved with the datasets resampled), the 0.9.6 datasets give 51, the drops alone 58 and 0.9.7 61: taking the drops first, of the +10 pairs +7 come from the drops, which were chosen with the 0.9.6 board and its separability table known, and +3 from the additions, the pooling and the regrown A lists (another order could split the gain differently).

The dated notes of 0.9.7 (PREREG09, after Amendment 13)

Note 2026-11-05 18:46Z (0.9.7 admission, wave 0: HOT; Amendment 13 item 8 step 2). Written before any 0.9.7 cell on HOT's new series exists. HOT (Hill of Towie, existing A dataset) re-selected under cap 100 (item 2; several series per turbine) with data097/src/select097.py (= select096 with the per-source cap): 59 candidates after construction, chance computed for all 59, none excluded, 59 accepted, 50 Eval and 9 Tuning (0.9.6: 29 and 6). 24 series are new; 3 old Eval series move to Tuning under the unchanged hash rule. Measured with work/r097/measure097.py (validated: reproduces nulls095 HOT to 1.1e-16): 107 anomaly segments in Eval, median scored anomaly share 0.0142, 50 Eval series with prefix >= 512, elapsed-time best 0.529 (centre). Items 2(a), (b), (c) hold: HOT is in. The board uses the stored 0.9.6 cells for the 26 + 3 old series and new cells for the 24 new ones; the old series are also recomputed by the 0.9.7 tree as a reproduction check. Record: data097/admit097.json (wave 0). No 0.9.7 score has been read.

Note 2026-11-05 18:57Z (0.9.7 admission, wave 1; Amendment 13 item 8 step 2). Written before any 0.9.7 cell on the units it admits. Record: data097/admit097.json, generated by work/r097/admit097.py from the measure097.py files (no model read). In (rules 2(a)-(c) hold; the elapsed-time gate holds for the new ones): ESA (ESA_M1 + ESA_M2 pooled, 95 Eval, 97 segments, median share 0.032, elapsed 0.611), NASA (SMAP + MSL pooled, 37, 47, 0.040, 0.516), UKWIND (17, 41, 0.057, 0.558; unchanged), HOT (50, 107, 0.014, 0.529; cap 100), PREDIST (cap 100: 41 Eval, 8 Tuning, was 29 / 6; 41 segments, 0.051, 0.512; 14 new series, PREDIST__m1_s17_e29 and __m2_s19_e29 move to Tuning), NOBOOM (cap 300; lists unchanged: 20, 37, 0.191, 0.561), UAV-SEAD (new, firm: 101 Eval, 147 segments, 0.083, elapsed 0.585 <= 0.65; real PX4 flights; 25 Hz grid chosen on a label-free count of flights reaching a 512-row prefix), UR5e (new, firm, CC BY-NC 4.0: 63 Eval, 74 segments, 0.191, elapsed 0.616; status, safety, program-node, program-loop, digital IO (bit 5 = the labelling button), speed-scaling and time registers dropped). Gates on the existing lists: measure097 equals nulls095/nulls096 per series (max |diff| 2.2e-16); PREDIST and NOBOOM rebuilt from raw reproduce every 0.9.6 series (X, y, tr, channels; PREDIST's unread ts array differs in unit only, pandas version). Out: Eawag (only 4 ponds carry the expert label, 5 segments: rules 2(a), (b) cannot hold), FAIR-MAST (a precursor-only cut puts the label at the end of every scored block, so elapsed time finds it; TokaMark ships no disruption labels), CAFUC2 (check 2: its faults are synthetic signals written into copies of recorded flights after the flight, so the label is not an event in the physical system; 96 Eval would otherwise pass items 2 and 4; Jonas may reverse this in a dated note). Pending: HAI pooling, RflyMAD, BASiC, TCABR, PMU, OnboardEU, KUKA, Cranfield, Greek wind. No 0.9.7 score has been read.

Note 2026-11-05 19:02Z (reading of Amendment 13 item 9, A signal-to-noise; written before any 0.9.7 score exists). Item 9 defines the A signal-to-noise as "spread SD of the ranked rows over their pair noise, as the 0.9.6 card computes it" and anchors it at "the 3.0 of 0.9.6". The 0.9.6 card's number is the snr field of results/separability096.json (separability096.py: mean absolute point difference over all ranked-row pairs of the task score divided by the median bootstrap SD of the paired difference), which is 2.997 for A; spread SD over pair noise is 0.02110 / 0.00795 = 2.65. The test uses the card's quantity, the one that gives 3.0 (threshold 6, goal 9), computed by separability096.py unchanged on board097; spread SD over pair noise is reported beside it for 0.9.5, 0.9.6 and 0.9.7. No rule is changed.

Note 2026-11-05 19:05Z (deviation from Amendment 13 item 6, order of the SNS HVCM gate cells; disclosed, no rule changed). Item 6 says the gate cells of nicola15/sns_hvcm_b_end (combo-10M and its random twin, 5 repeats) are "computed before any other 0.9.7 cell". They were not: A cells of the units admitted in waves 0 and 1 (HOT, PREDIST, UAV-SEAD, UR5e; Clariden jobs 3867625, 3867627, 3867656) were computed first, because the C unit had not been rebuilt yet (nicola-l4 unreachable). No 0.9.7 score of any kind has been read: the only checks run on those cells are equality checks against stored 0.9.6 cells. The SNS rule is mechanical (combo-10M minus twin positive in at least 4 of 5 repeats and mean at least +0.02) and reads only SNS cells, so the order cannot have informed it. Its gate cells are computed as soon as the unit is rebuilt and accepted by the sha256 check, and its decision is written in a dated note before any other 0.9.7 C cell is read. The card states this deviation.

Note 2026-11-05 19:10Z (0.9.7 C admission; Amendment 13 item 6). VM nicola-l4 was unreachable (SSH host key verification failed; not bypassed), so the three exp-nicola-15 C units were rebuilt from their builders (branch dev/c-candidate-checks, builders byte-identical to the logged commits 885ab0cd0, 99988a280, 28ae362c4). X sha256 (raw float32 bytes of prepared.npz X) equals LOG.md for all three: bdg2_meter6 82e03cf1..., robo6_rb6 935ae4ca..., sns_hvcm_b_end b8dcbe28.... In: nicola15/bdg2_meter6 (CC BY-SA 4.0; 75,409 windows, 1,615 groups, 6 classes), nicola15/robo6_rb6 (MIT; 6,869 windows, 5,352 groups, 6 classes). Pending: nicola15/sns_hvcm_b_end (CC BY 4.0; 3,323 windows, 15 groups, 2 classes) until its combo-minus-twin gate (cgate097.py) is computed; its first gate run lost the twin cell to a missing result directory and is being rerun. No 0.9.7 C cell has been read. Record: data097/admit097.json (C section).

Note 2026-11-05 19:12Z (0.9.7 admission, wave 2; Amendment 13 item 8 step 2). Out: BASiC (item 4 condition: its own paper, Ahmad and Akram, Data in Brief 2024, doi:10.1016/j.dib.2024.110069, states the flights are software-in-the-loop simulations; not built) and RflyMAD real flights (built from the real-flight subset only, non-commercial, 255 Eval series, 405 segments, median share 0.289; fails the elapsed-time gate: rising 0.731 > 0.65, the faults come late in the flights). No 0.9.7 score has been read. Record: data097/admit097.json.

Note 2026-11-05 19:13Z (Amendment 13 item 6: SNS HVCM gate). nicola15/sns_hvcm_b_end gate cells (head H1, 5 repeats; combo-10M from Clariden job 3867712, its random twin from 3867716 after the first twin task failed on a missing result directory): combo minus twin per repeat -0.0187, +0.0161, +0.0473, -0.0146, +0.0546; positive in 3 of 5 (needs 4), mean +0.0169 (needs +0.02). SNS HVCM is out of 0.9.7. C has 24 datasets (25 of 0.9.6 minus 3 plus bdg2_meter6 and robo6_rb6). No other 0.9.7 C cell has been read. Record: results/cgate097.json, data097/admit097.json.

Note 2026-11-05 19:15Z (Amendment 13 item 5, F3 thresholds; computed 18:53:31Z, before any new FE download or build). On the 14 kept FE datasets of 0.9.6, from the stored PrecursorBench v1.4 reference curves (no model involved): the time-since-last-event row (PB v1.4 "event timing") reaches at most 0.7956 (nasa_battery_rw); a pure time-of-day row (hour-of-day harmonics 1 and 2 of the last observed sample; defined on the 9 datasets with a calendar) reaches at most 0.5851 (towie_genbearing_hot). A new FE dataset is admitted only if both rows stay at or below these values, with at least 300 events, at least 30 independent units or days, and the best free-forecast row at most 0.85. File: data097/F3_thresholds097.json (copied from the FE builder's output, written 18:53:31Z; the first FE download started 18:54Z per results/timing097.json).

Note 2026-11-05 19:21Z (0.9.7 admission, wave 3; Amendment 13 item 8 step 2). HAI in (pooled, item 4): HAI 22.04 plus 20.07, 21.03 and 23.05 of the same testbed (CC BY-SA 4.0), built with the 22.04 rules; build_hai.py reproduces all 21 old 22.04 series byte for byte. Two run-level decisions, both made before any HAI cell: (i) "hash order unchanged" (item 2) is read as each source keeping its own hash tag, so pooled HAI is ordered with HAI's tag A08|, which puts the 21 old series exactly in their 0.9.5 Eval/Tuning places (with A083| five old series would have moved); select097.py reads an optional hash_tag per source (default A083|); (ii) HAIEnd 23.05 is left out: it carries the same two 23.05 test recordings and the same label file through a different set of tags, so its series repeat physical attacks already in HAI 23.05 (needs-you item; with it HAI would have 72 Eval series, 134 segments, elapsed 0.527, also passing). Result: 58 Eval, 11 Tuning, 103 segments, median scored share 0.030, elapsed best 0.533. OnboardEU out: clean single-defect subset chosen by the label text before any build (10 of 25 locations), 85 Eval series, but the elapsed-time gate fails (rising 0.708 > 0.65). No 0.9.7 score has been read.

Note 2026-11-05 19:23Z (0.9.7 admission, wave 4, the final A list, and the A gate threshold; Amendment 13 items 4 and 8; before any 0.9.7 A cell is read). In: KUKA LWR4+ collisions (second wave; Zenodo CC BY 4.0, description MIT): 255 Eval series from 148 recordings, 534 segments, median share 0.108, elapsed best 0.519. Its condition holds as written: the labels are the authors' hand-entered collision onsets (y = 1 on [m - 20, m + 492) as in their example.m), whose counts match the experimenter's independent log in 200 of 206 recordings; the external torque at the marked onset varies (coefficient of variation 0.90) and a threshold is crossed a median 5.5 samples after it, so the labels are not thresholds on the torque. Disclosed: a single threshold of about 1.75 Nm on the external-torque channels finds the collisions with F1 0.92, so this dataset is easy for a level detector (the max-|z| reference row will show it). Out: TCABR (the time-matched precursor build gives 3 series because disruptive and normal shots come from different days; the disruptive-only build has elapsed best 1.000 > 0.65), PMU (the continuous 1 s record is one channel and covers 23 of 88 cases: P4 and rule 2(b) fail; not built), Greek wind farm (7 Eval series, 8 segments: rules 2(a), (b)), Cranfield (the official download needs a MathWorks sign-in). Final 0.9.7 A list (10 datasets, 737 Eval series): ESA 95, NASA 37, UKWIND 17, HOT 50, PREDIST 41, NOBOOM 20, HAI 58, UAV-SEAD 101, UR5e 63, KUKA 255. One flight set (UAV-SEAD) of 10, within the cap of a third. A gate threshold (Amendment 5 rule 5(b), recomputed over the final list as item 4 requires): 0.5540 = mean over the 10 datasets of the highest data-free null (centre assignment; per dataset the largest of the Eval means of elapsed time rising, falling and the random window score, from measure097.py): ESA 0.6109, NASA 0.5162, UKWIND 0.5578, HOT 0.5293, PREDIST 0.5124, NOBOOM 0.5605, HAI 0.5328, UAV-SEAD 0.5851, UR5e 0.6156, KUKA 0.5194. A enters the CAFE score if at least half the ranked rows score above it. File: results/agate097.json. No 0.9.7 A cell has been read.

Note 2026-11-05 19:44Z (0.9.7 FE admission, wave 1; Amendment 13 item 5). In: heatpump_defrost (HeatpumpMonitor.org, CC BY 4.0; event = defrost cycle; inputs = measured temperatures and power; built by data097/src/fe_build_heatpump.py with the PrecursorBench v1.4 rule). Model-free gates (reference rows scored as the other FE reference rows, data097/fe_gates_heatpump_defrost.json): 53,978 events (>= 300), 113 independent installations (>= 30), time since last event 0.659 (<= 0.7956), time of day 0.566 (<= 0.5851), best free-forecast row 0.729, curvature (<= 0.85). IPShocks and PyrNet are still building; each is out if not built and gated by 21:30Z. The FE head for the new unit is the torch head of the dated note 011865766 (0.9.6); its reproduction of the 0.9.5 numpy cells holds through the 0.9.7 tree (ZERO on nasa_battery_rw, all draws equal, DECISIONS097.md), so the FT tasks run without a new gate file. No 0.9.7 FE cell has been read.

Note 2026-11-05 21:59Z (0.9.7 FE admission, wave 2, final FE list; Amendment 13 item 5). In: ipshocks_fluxes (IPShocks catalogue, CC BY 4.0, events = shock arrivals at L1; inputs = ACE and Wind field, plasma and particle fluxes from NASA CDAWeb, 1998-2021, 5-min bins). Gates (computed 19:47Z, data097/fe_gates_ipshocks_fluxes.json): 914 events, 629 distinct shock days (independence counted by day because both spacecraft sit at L1), time since last event 0.490, time of day 0.513, best free forecast 0.695 (current value). Builder choices made after a structural test but before any gate score: 1-hour events, lead times 5 min to 4 h, row thinning as build_ngafid. For heatpump_defrost one builder choice was made after a build and before any score: feed stretches shorter than 1,024 min are set missing. Out: PyrNet (not built and gated by the 21:30Z cutoff). Final FE list: the 14 kept datasets plus heatpump_defrost and ipshocks_fluxes = 16. Disclosed: on two kept FE datasets the best free-forecast row is already above the 0.85 gate that new FE datasets must meet (memristor_jerk 0.882, nasa_battery_rw 0.873; data097/F3_thresholds097.json); the gate applies to new datasets only, so they stay. No 0.9.7 FE cell has been read.

Note 2026-11-05 22:32Z (responses to audits 1 and 2; disclosures; no rule changed). Two fresh audits (work/r097/audits/audit1_rules.md, audit2_code.md) found no blocker. Context: at about 22:10Z a provisional board (the two new FE datasets and two HAI rows still without cells) was built to check the code end to end, so 0.9.7 scores have been visible since then; for that reason no admission rule or decision below is changed now, and every response is a disclosure or a reporting-only sensitivity. 1. KUKA cap (audit 1 R1, audit 2 F8). KUKA was admitted at 19:23Z (before any A cell was read) with cap 300, reading its 206 recordings as independent runs of a robot program, as UR5e's runs were read; item 2 gives 100 to "several series per asset", and KUKA is one robot arm with several blocks per recording (255 Eval series from 148 recordings). The cap stays as admitted. Reported beside the result: the board and the item-9 test with KUKA at its first 100 Eval series in hash order (results/test097_k100.json). 2. Within-dataset independence (audit 1 R2, audit 2 F9). The A bootstrap resamples series, as in 0.9.6 (item 7); blocks of one recording (HAI, KUKA, ESA, UAV-SEAD, HOT, UKWIND) are resampled as if independent except the near-duplicate clusters of 0.9.5. Reported beside the result: the board and the item-9 test with every A dataset resampled by recording (results/test097_cr.json). 3. PyrNet cutoff (audit 1 R3). The 21:30Z cutoff for the FE builds was set by the orchestrator at 19:45Z in a message to the builder (data097/fe_status097.md); it is not in Amendment 13 (whose item 8 has no deadline beyond the plan's budget) and is earlier than the plan's suggested 4-hour phase-1 cut-off (~22:42Z). PyrNet's 4.8 GB download finished at 21:58Z and its builder never ran. PyrNet is not in 0.9.7; it is a 0.9.7.1 candidate. 4. Wording of the 19:02Z note (audit 1 R4). That note says "before any 0.9.7 score exists"; A cells of waves 0 and 1 existed from 18:49Z but none had been read. Read it as "before any 0.9.7 score was read". 5. A_series gate tolerance (audit 1 R6). The gate comparing recomputed old series with their stored cells was exact until PREDIST and HAI were recomputed; it now allows 1e-15 (written 22:16Z, after the 1.1e-16 differences appeared) because select097 recomputed those datasets' chance values (difference about 1e-17). The board uses the stored cells for every old series, so no score depends on it; the bit-for-bit gate of one unchanged unit per task (A_repro, 49/49 exact) is unchanged. 6. FE caveats (audit 1 R8, R10). heatpump_defrost's event (heat output below 0 W while the compressor draws more than 50 W) is defined partly on electrical power, which is also an input; the F3 thresholds were computed from 10 head seeds on the kept datasets and the new datasets' gate rows from 3. Both are stated on the card. 7. Other disclosures. UAV-SEAD keeps battery voltage, which falls through a flight (its elapsed-time score is 0.585, inside the gate) (audit 1 R13). The SNS gate compared combo-10M with the board's random-init LeSaturn row (3.7M parameters), the twin used for "combo minus twin" since 0.9.6 (audit 1 R16). A at w = 512 (item 7) is not computed in this run unless time remains (audit 1 R11). The leave-one-out boards hold every non-admitted dataset out so that a campaign sibling cannot enter (fixed in sens097.py at 22:06Z, before the final board).

Note 2026-11-06 10:02Z (ipshocks_fluxes and standing check 3; written after the 0.9.7 board was built, before any ipshocks_fluxes score was read). The board leaves ipshocks_fluxes out under the standing admission check 3 (section "Admission", item 3: "FE: at least 1,000 events"), which Amendment 13 item 7 keeps ("everything else is 0.9.6"): it has 914 events. Amendment 13 item 5 sets the gates for new FE datasets (at least 300 events, at least 30 independent units or days, F3, best free forecast at most 0.85), and the note of 21:59Z admitted it under those gates as the 16th FE dataset; neither text says whether the 300-event gate replaces check 3 for new FE datasets. The conflict was not seen before: the provisional runs of 22:00Z to 22:30Z had no cells on the new FE units and left them out for that reason. The board code (board097.py, unchanged since before any 0.9.7 score) applies check 3, and the run's escalation rule takes the default "leave the unit out" when a decision needs Jonas Petersen. 0.9.7 therefore has 15 FE datasets; ipshocks_fluxes is listed as admitted by item 5 and left out by check 3. Its cells exist (Clariden job 3868625) and are kept. Only Jonas Petersen can reverse this, in a dated note; the board would then be rebuilt with 16 FE datasets and every result file regenerated. The 0.9.7 test of item 9, the tiers and every reported number are those of the 15-dataset board. No ipshocks_fluxes score has been read.

Note 2026-11-06 10:10Z (addendum to the note of 10:02Z, after audit 3 finding F-A1; no rule changed). The 10:02Z note says no ipshocks_fluxes score had been read. The writer had not looked at one, but that cannot be shown: the cells of every row on ipshocks_fluxes were in ~/.hepa09/r097/cells097/F.json from the build of about 09:25Z and in the card's data/F.json. Audit 3 then built the board with ipshocks_fluxes (check 3 lowered to 900 events) as a check. That board is now a reported sensitivity (results/board097_ips.json, test097_ips.json; release097.sh), shown on the card beside the test. Any reversal of the 10:02Z default would therefore be made with those scores known, and the card would say so. The published board stays the 15-FE-dataset board.

Note 2026-11-06 11:05Z (decision on ipshocks_fluxes by Jonas Petersen; closes the notes of 10:02Z and 10:10Z). Jonas Petersen decided: ipshocks_fluxes stays out of 0.9.7. Standing check 3 (FE at least 1,000 events) holds for every FE dataset, new ones included; Amendment 13 item 5's 300-event gate is an additional gate for new FE datasets and does not replace check 3. The decision was made with the board with ipshocks_fluxes known (sensitivity results/board097_ips.json: 58 pairs resolved against 61, pj_better_3p7m and pjgood-warp-xview-combo-10M swap ranks 2 and 3; the item-9 test fails either way), and is the default the run took at 10:02Z. 0.9.7 has 15 FE datasets.

Comparison across versions since 0.9.4

Rank of each ranked model on the CAFE score in each version. A diamond is a ◊ row. Ranks inside tier 1 are no claim; the tables give the tiers.

The versions differ in datasets, heads and, for 0.9.4, in the ranked rows (TiRex-2 is ranked in 0.9.4 only), so scores of different versions are means over different material; ranks, tiers and resolved pairs are what can be compared. 0.9.4 is the released 0.9.4 board, unchanged in 0.9.4.1 (board094_frozen.json).

Datasets and pairs resolved. Main board: datasets fixed, test material resampled. Datasets resampled: the measure of the 1.0 test.

version	C datasets	A datasets	FE datasets	ranked models	tasks in CAFE	CAFE pairs resolved, main	of	CAFE pairs resolved, datasets resampled	of
0.9.4 (0.9.4.1)	16	7	16	16	C, A, FE	39 (1 within seed noise)	120	16	120
0.9.5	17	12	16	15	C, A, FE	73 (1 within seed noise)	105	49	105
0.9.6	25	14	17	15	C, A, FE	73 (1 within seed noise)	105	51	105
0.9.7	24	10	15	15	C, A, FE	85 (10 within seed noise)	105	61	105

Per task. Pairs resolved on the task board, main / datasets resampled, and signal-to-noise (mean absolute gap between two ranked rows over the median bootstrap SD of that gap; separability096.py). A dash: no such file for that version.

version	C pairs	A pairs	FE pairs	C signal-to-noise	A signal-to-noise	FE signal-to-noise	CAFE signal-to-noise
0.9.4 (0.9.4.1)	73 / –	24 / –	69 / –	–	–	–	–
0.9.5	70 / 15	18 / 4	87 / 67	8.39	2.48	15.10	7.32
0.9.6	74 / 21	37 / 6	87 / 66	8.98	3.00	14.93	7.89
0.9.7	80 / 25	60 / 9	87 / 67	10.06	4.04	16.26	10.88

No separability file exists for 0.9.4.

First place. The first row of each board; inside tier 1 the order is no claim.

version	CAFE	C	A	FE	models in CAFE tier 1
0.9.4 (0.9.4.1)	MantisV2	Timestral-209M	Timestral-123M	pj_better_1m ◊	13
0.9.5	pj_better_10m ◊	Chronos-2	pj_better_3p7m ◊	Mantis	6
0.9.6	pj_better_10m ◊	Chronos-2	pj_better_3p7m ◊	Mantis	5
0.9.7	pj_better_10m ◊	Chronos-2	pj_better_10m ◊	Mantis	1

Rank and tier of each ranked model on the CAFE score, rank (tier):

model	0.9.4 (0.9.4.1)	0.9.5	0.9.6	0.9.7
pj_better_10m ◊	6 (1)	1 (1)	1 (1)	1 (1)
pj_better_3p7m ◊	7 (1)	2 (1)	2 (1)	2 (2)
pjgood-warp-xview-combo-10M ◊	5 (1)	4 (1)	4 (1)	3 (2)
Mantis	4 (1)	3 (1)	3 (1)	4 (2)
Chronos-2	12 (1)	5 (1)	5 (1)	5 (2)
LeSaturn-warppool-factorized-7.4M ◊	10 (1)	6 (1)	6 (2)	6 (3)
pj_better_1m ◊	9 (1)	7 (2)	7 (2)	7 (3)
MantisV2	1 (1)	8 (2)	8 (2)	8 (4)
NuTime	14 (2)	9 (3)	9 (3)	9 (5)
Timestral-209M	3 (1)	10 (3)	10 (3)	10 (5)
Timestral-123M	2 (1)	11 (3)	11 (3)	11 (6)
Toto-Open-Base-1.0	15 (3)	14 (3)	12 (3)	12 (6)
Timestral-28M	11 (1)	13 (3)	13 (3)	13 (6)
Timestral-5M	13 (1)	12 (3)	14 (3)	14 (6)
TimesFM-2.5-200M	16 (3)	15 (4)	15 (4)	15 (7)
TiRex-2	8 (1)	–	–	–

What changed, by the amendments of each version and the datasets added (+) or removed (-) against the version before:

version	rules	datasets
0.9.4 (0.9.4.1)	Amendment 5	first version compared
0.9.5	Amendments 6, 7, 8, 9 and 10	C: +2 (aursad095, voraus095) -1 (voraus); A: +5 (ESA_M1, CARE_A, MSL, W3, SKAB)
0.9.6	Amendments 11 and 12	C: +8 (chillan5, haptic_texture8, openrainer_m, poseit_ft, px4_airframe3, soredd_abc, swiss_load10, ved_powertrain3); A: +2 (PREDIST, NOBOOM); FE: +1 (ngafid_hastall)
0.9.7	Amendment 13	C: +2 (bdg2_meter6, robo6_rb6) -3 (3w, bosch, soredd_abc); A: +5 (ESA, NASA, UAVSEAD, UR5E, KUKA) -9 (ESA_M1, ESA_M2, CARE_A, SMAP, MSL, CARE_C, RIGS, W3, SKAB); FE: +1 (heatpump_defrost) -3 (epa_aqs_pm25, substorm_omni, ngafid_hastall)

Each version's own changes are in the change log.

0.9.6 and 0.9.7 side by side

Scores of the two versions are means over different datasets, so a change in score is not a change in the model; ranks and tiers are what can be compared.

Ranked rows of both versions, in the 0.9.7 order.

model	rank 0.9.6	rank 0.9.7	move	tier 0.9.6	tier 0.9.7	score 0.9.6	score 0.9.7	change
pj_better_10m ◊	1	1	=	1	1	0.7363	0.7384	+0.0021
pj_better_3p7m ◊	2	2	=	1	2	0.7347	0.7341	−0.0006
pjgood-warp-xview-combo-10M ◊	4	3	↑1	1	2	0.7326	0.7340	+0.0013
Mantis	3	4	↓1	1	2	0.7329	0.7326	−0.0003
Chronos-2	5	5	=	1	2	0.7301	0.7301	−0.0000
LeSaturn-warppool-factorized-7.4M ◊	6	6	=	2	3	0.7270	0.7275	+0.0005
pj_better_1m ◊	7	7	=	2	3	0.7224	0.7241	+0.0017
MantisV2	8	8	=	2	4	0.7192	0.7181	−0.0011
NuTime	9	9	=	3	5	0.7067	0.7050	−0.0017
Timestral-209M	10	10	=	3	5	0.7004	0.6989	−0.0015
Timestral-123M	11	11	=	3	6	0.6990	0.6969	−0.0021
Toto-Open-Base-1.0	12	12	=	3	6	0.6972	0.6954	−0.0018
Timestral-28M	13	13	=	3	6	0.6965	0.6943	−0.0022
Timestral-5M	14	14	=	3	6	0.6965	0.6943	−0.0022
TimesFM-2.5-200M	15	15	=	4	7	0.6657	0.6662	+0.0006

Ranked rows of both versions, in the 0.9.7 order. 25 datasets in 0.9.6, 24 in 0.9.7.

model	rank 0.9.6	rank 0.9.7	move	tier 0.9.6	tier 0.9.7	score 0.9.6	score 0.9.7	change
Chronos-2	1	1	=	1	1	0.8365	0.8442	+0.0077
pjgood-warp-xview-combo-10M ◊	3	2	↑1	1	1	0.8337	0.8429	+0.0092
Timestral-209M	2	3	↓1	1	1	0.8347	0.8411	+0.0064
Timestral-123M	4	4	=	1	2	0.8327	0.8389	+0.0062
pj_better_10m ◊	5	5	=	2	2	0.8297	0.8380	+0.0083
pj_better_3p7m ◊	7	6	↑1	2	2	0.8275	0.8348	+0.0072
Timestral-28M	6	7	↓1	2	3	0.8290	0.8347	+0.0057
Mantis	9	8	↑1	3	4	0.8225	0.8294	+0.0069
LeSaturn-warppool-factorized-7.4M ◊	8	9	↓1	3	4	0.8232	0.8291	+0.0059
NuTime	10	10	=	3	4	0.8215	0.8266	+0.0052
Timestral-5M	11	11	=	3	4	0.8213	0.8258	+0.0044
MantisV2	12	12	=	3	4	0.8186	0.8257	+0.0071
pj_better_1m ◊	13	13	=	4	5	0.8150	0.8206	+0.0057
TimesFM-2.5-200M	14	14	=	5	6	0.7952	0.8021	+0.0068
Toto-Open-Base-1.0	15	15	=	6	7	0.7678	0.7726	+0.0048

Ranked rows of both versions, in the 0.9.7 order. 14 datasets in 0.9.6, 10 in 0.9.7.

model	rank 0.9.6	rank 0.9.7	move	tier 0.9.6	tier 0.9.7	score 0.9.6	score 0.9.7	change
pj_better_10m ◊	3	1	↑2	1	1	0.6718	0.6483	−0.0235
pj_better_3p7m ◊	1	2	↓1	1	1	0.6762	0.6466	−0.0295
pjgood-warp-xview-combo-10M ◊	2	3	↓1	1	1	0.6727	0.6463	−0.0264
pj_better_1m ◊	6	4	↑2	1	1	0.6628	0.6416	−0.0212
Mantis	5	5	=	1	1	0.6634	0.6376	−0.0258
MantisV2	4	6	↓2	1	1	0.6649	0.6334	−0.0315
NuTime	7	7	=	1	1	0.6603	0.6331	−0.0272
LeSaturn-warppool-factorized-7.4M ◊	8	8	=	2	2	0.6518	0.6293	−0.0225
Chronos-2	9	9	=	2	2	0.6508	0.6252	−0.0255
Timestral-209M	10	10	=	2	2	0.6420	0.6162	−0.0258
Timestral-28M	12	11	↑1	2	3	0.6402	0.6143	−0.0259
Timestral-123M	11	12	↓1	2	3	0.6405	0.6134	−0.0270
Timestral-5M	13	13	=	2	3	0.6381	0.6134	−0.0247
TimesFM-2.5-200M	15	14	↑1	3	4	0.6031	0.5886	−0.0146
Toto-Open-Base-1.0	14	15	↓1	2	4	0.6141	0.5855	−0.0286

Ranked rows of both versions, in the 0.9.7 order. 17 datasets in 0.9.6, 15 in 0.9.7.

model	rank 0.9.6	rank 0.9.7	move	tier 0.9.6	tier 0.9.7	score 0.9.6	score 0.9.7	change
Mantis	1	1	=	1	1	0.7129	0.7310	+0.0181
pj_better_10m ◊	3	2	↑1	1	1	0.7073	0.7288	+0.0215
Toto-Open-Base-1.0	2	3	↓1	1	1	0.7098	0.7282	+0.0184
LeSaturn-warppool-factorized-7.4M ◊	4	4	=	1	1	0.7061	0.7240	+0.0179
pj_better_3p7m ◊	6	5	↑1	2	2	0.7003	0.7210	+0.0207
Chronos-2	5	6	↓1	2	2	0.7031	0.7208	+0.0177
pjgood-warp-xview-combo-10M ◊	7	7	=	3	3	0.6915	0.7127	+0.0212
pj_better_1m ◊	8	8	=	3	3	0.6896	0.7102	+0.0206
MantisV2	9	9	=	4	4	0.6741	0.6951	+0.0209
NuTime	10	10	=	5	5	0.6385	0.6553	+0.0168
Timestral-5M	11	11	=	5	6	0.6300	0.6437	+0.0137
Timestral-209M	12	12	=	6	6	0.6246	0.6396	+0.0150
Timestral-123M	13	13	=	6	6	0.6240	0.6384	+0.0145
Timestral-28M	14	14	=	6	7	0.6204	0.6340	+0.0136
TimesFM-2.5-200M	15	15	=	7	8	0.5987	0.6081	+0.0094

Separability by task, dataset and model

Computed by separability096.py on the 0.9.7 board, over the 15 ranked rows only. Spread SD: SD of the rows' scores. Pair noise: median bootstrap SD of the paired difference of two rows. Signal-to-noise: mean absolute gap between two rows over the pair noise. Resolved alone: pairs resolved on a board of that dataset alone. Tau with the task: Kendall tau between a dataset's order and the mean of the task's other datasets. Board effect: CAFE pairs resolved without the dataset minus with it (negative: the dataset helps).

By task.

task	datasets	spread SD	range	pair noise	signal-to-noise	in 0.9.6	pairs resolved	datasets resampled	tau with the other tasks	share of CAFE variance	CAFE pairs without it	same, resampled
C	24	0.0184	0.0716	0.0018	10.06	8.98	80	25	0.181	0.143	+0	-4
A	10	0.0196	0.0628	0.0056	4.04	3.00	60	9	0.600	0.271	+6	-1
FE	15	0.0438	0.1229	0.0030	16.26	14.93	87	67	0.448	0.587	-20	-31
CAFE	49	0.0212	0.0721	0.0022	10.88	7.89	85	61	–	–	–	–

By dataset (● new in 0.9.7; ○ pooled from 0.9.6 datasets whose scores were published), highest signal-to-noise first within each task.

All 49 datasets

task	dataset	spread SD	range	pair noise	signal-to-noise	pairs resolved alone	tau with the task	task pairs without it	CAFE pairs without it, resampled
C	phys/lendb	0.1152	0.3918	0.0046	24.45	97	0.429	-7	+1
C	nicola15/bdg2_meter6 ●	0.0279	0.0952	0.0037	8.53	89	0.219	+0	-1
C	nicola15/robo6_rb6 ●	0.0239	0.0741	0.0038	7.40	81	0.295	+1	+0
C	mv06/stead	0.0169	0.0642	0.0027	7.03	80	0.371	-4	-1
C	mv06/hydraulic	0.0959	0.3650	0.0125	6.43	62	0.314	-5	+3
C	nicola15/ved_powertrain3	0.0191	0.0625	0.0033	6.16	73	0.524	-3	-1
C	nicola15/haptic_texture8	0.0161	0.0677	0.0031	5.96	75	0.162	+0	+0
C	ucr/ElectricDevices	0.0126	0.0406	0.0031	4.74	63	0.486	-1	-1
C	ucr/SonyAIBORobotSurface2	0.0592	0.2405	0.0148	4.20	47	0.019	-3	-2
C	nicola15/openrainer_m	0.0148	0.0472	0.0042	4.12	61	0.105	+1	+0
C	ucr/FreezerRegularTrain	0.0117	0.0456	0.0030	4.07	46	0.452	+0	-2
C	mv06/voraus095	0.0344	0.1055	0.0095	3.97	54	0.200	-2	-1
C	nicola15/chillan5	0.0080	0.0318	0.0025	3.60	54	0.600	+0	+0
C	mv06/aursad095	0.0201	0.0624	0.0066	3.59	57	0.200	-1	+0
C	nicola15/px4_airframe3	0.0211	0.0633	0.0068	3.50	60	0.295	-5	+0
C	ucr/MoteStrain	0.0281	0.1079	0.0088	3.41	38	0.410	-2	-4
C	ucr/Wafer	0.0108	0.0383	0.0033	3.33	33	0.421	-1	-2
C	uea/AsphaltPavementTypeCoordinates	0.0233	0.0897	0.0084	3.16	47	0.143	+0	+0
C	ucr/PLAID	0.0383	0.1572	0.0141	2.94	35	0.238	-1	-3
C	nicola15/swiss_load10	0.0134	0.0473	0.0052	2.91	51	0.077	-1	-1
C	ucr/FordA	0.0207	0.0747	0.0082	2.89	31	0.524	-5	+0
C	nicola15/poseit_ft	0.0160	0.0573	0.0065	2.84	50	0.562	-3	+0
C	mv06/gas251	0.0155	0.0548	0.0061	2.73	39	0.257	+1	+0
C	ucr/ACSF1	0.0394	0.1222	0.0237	1.94	15	0.010	+1	+0
A	KUKA ●	0.0199	0.0640	0.0038	6.08	81	0.562	-2	-3
A	HAI	0.0368	0.1207	0.0125	3.46	58	0.543	-14	-2
A	UR5E ●	0.0293	0.1196	0.0087	3.44	39	0.048	+1	-2
A	ESA ○	0.0287	0.0744	0.0101	3.22	42	0.638	-6	-2
A	HOT	0.0344	0.1507	0.0127	2.85	38	0.467	-2	+0
A	UAVSEAD ●	0.0170	0.0589	0.0091	2.09	15	0.219	-3	-2
A	PREDIST	0.0323	0.1198	0.0176	2.04	19	0.619	-3	-3
A	NOBOOM	0.0401	0.1293	0.0263	1.76	25	0.657	-10	-3
A	NASA ○	0.0328	0.1255	0.0245	1.48	11	0.257	+5	-2
A	UKWIND	0.0217	0.0852	0.0229	1.02	1	0.162	+9	+0
FE	megathrust_sync_v2	0.0646	0.1925	0.0059	12.37	79	0.486	+1	-2
FE	rotary_utrecht	0.0371	0.1058	0.0041	10.18	75	0.733	-1	-1
FE	minati_circuits	0.0577	0.1722	0.0079	8.37	79	0.657	+0	+0
FE	memristor_jerk	0.0646	0.2532	0.0078	8.30	72	0.733	+0	-1
FE	asos_lows_b	0.0611	0.1531	0.0089	7.77	66	0.733	+1	-2
FE	stickslip_poincloux	0.0415	0.1443	0.0071	6.88	81	0.638	+0	+2
FE	ndbc_buoy_v2_b	0.0480	0.1285	0.0098	5.72	71	0.810	-1	-1
FE	geyser_strokkur	0.0372	0.1079	0.0078	5.66	73	0.676	+0	+2
FE	cdip_waves_c	0.0600	0.1592	0.0132	5.35	74	0.848	-2	-2
FE	noaa_coops_surge_b	0.0648	0.1664	0.0155	4.83	56	0.829	+0	-2
FE	heatpump_defrost ●	0.0361	0.1018	0.0118	3.60	54	0.638	+1	+0
FE	nasa_battery_rw	0.0254	0.0866	0.0083	3.41	44	0.410	+1	-1
FE	usgs_reach_v2_b	0.0419	0.1052	0.0154	3.10	40	0.600	+1	-1
FE	wind_genbearing_hot	0.0276	0.0769	0.0105	3.08	35	0.676	-1	-1
FE	towie_genbearing_hot	0.0472	0.1269	0.0195	2.82	39	0.600	+3	-2

By model. Pairs resolved: pairs on the CAFE main board in which the model is resolved above or below the other. Datasets first: datasets where it has the highest score among the ranked rows.

model	rank	tier	pairs resolved	tier, datasets resampled	datasets first	median rank over datasets
pj_better_10m ◊	1	1	13 of 14	1	4	4
pj_better_3p7m ◊	2	2	11 of 14	1	3	5
pjgood-warp-xview-combo-10M ◊	3	2	11 of 14	1	5	5
Mantis	4	2	9 of 14	1	4	6
Chronos-2	5	2	10 of 14	1	5	7
LeSaturn-warppool-factorized-7.4M ◊	6	3	11 of 14	1	1	7
pj_better_1m ◊	7	3	12 of 14	2	2	8
MantisV2	8	4	13 of 14	2	3	8
NuTime	9	5	13 of 14	3	4	9
Timestral-209M	10	5	11 of 14	3	3	9
Timestral-123M	11	6	10 of 14	3	3	10
Toto-Open-Base-1.0	12	6	10 of 14	3	9	13
Timestral-28M	13	6	11 of 14	3	3	10
Timestral-5M	14	6	11 of 14	3	0	11
TimesFM-2.5-200M	15	7	14 of 14	4	0	14

Every dataset's effect on the board

Each row is the 0.9.7 board rebuilt without one dataset, every other rule unchanged (49 of 49 datasets). Tau: Kendall tau of its order with the main order. Main board: pj_better_10m ◊ first, 1 model in tier 1, 85 pairs resolved (10 within seed noise), 61 with the datasets resampled.

task	dataset	first	its CAFE	tier 1	tier 1 change	pairs resolved	datasets resampled	tau
C	ucr/MoteStrain	pj_better_10m ◊	0.7369	1	same	85	57	0.962
C	ucr/SonyAIBORobotSurface2	pj_better_10m ◊	0.7380	1	same	83	59	0.981
C	ucr/Wafer	pj_better_10m ◊	0.7362	1	same	85	59	0.981
C	ucr/ACSF1	pj_better_10m ◊	0.7389	1	same	86	61	1.000
C	ucr/PLAID	pj_better_10m ◊	0.7376	1	same	85	58	0.981
C	ucr/FreezerRegularTrain	pj_better_10m ◊	0.7362	1	same	85	59	0.962
C	ucr/FordA	pj_better_10m ◊	0.7372	1	same	85	61	1.000
C	ucr/ElectricDevices	pj_better_10m ◊	0.7390	1	same	85	60	1.000
C	phys/lendb	pj_better_10m ◊	0.7375	1	same	84	62	0.943
C	mv06/hydraulic	pj_better_10m ◊	0.7366	1	same	84	64	0.962
C	mv06/gas251	pj_better_10m ◊	0.7377	1	same	85	61	1.000
C	mv06/stead	pj_better_10m ◊	0.7367	1	same	85	60	0.981
C	uea/AsphaltPavementTypeCoordinates	pj_better_10m ◊	0.7368	1	same	85	61	1.000
C	mv06/aursad095	pj_better_10m ◊	0.7405	1	same	85	61	1.000
C	mv06/voraus095	pj_better_10m ◊	0.7397	1	same	85	60	0.962
C	nicola15/chillan5	pj_better_10m ◊	0.7376	1	same	85	61	1.000
C	nicola15/haptic_texture8	pj_better_10m ◊	0.7421	1	same	85	61	1.000
C	nicola15/openrainer_m	pj_better_10m ◊	0.7378	1	same	85	61	0.943
C	nicola15/poseit_ft	pj_better_10m ◊	0.7415	1	same	85	61	1.000
C	nicola15/px4_airframe3	pj_better_10m ◊	0.7393	1	same	85	61	0.981
C	nicola15/swiss_load10	pj_better_10m ◊	0.7413	1	same	85	60	0.981
C	nicola15/ved_powertrain3	pj_better_10m ◊	0.7372	1	same	85	60	1.000
C	nicola15/bdg2_meter6 ●	pj_better_10m ◊	0.7391	1	same	84	60	0.981
C	nicola15/robo6_rb6 ●	pj_better_10m ◊	0.7393	1	same	85	61	0.981
A	HAI	pj_better_10m ◊	0.7395	1	same	80	59	0.981
A	UKWIND	pj_better_10m ◊	0.7383	1	same	86	61	0.962
A	HOT	pj_better_10m ◊	0.7395	1	same	82	61	0.981
A	PREDIST	pj_better_10m ◊	0.7370	1	same	84	58	0.962
A	NOBOOM	pj_better_10m ◊	0.7329	1	same	84	58	0.943
A	ESA ○	pj_better_10m ◊	0.7372	1	same	86	59	0.981
A	NASA ○	pj_better_10m ◊	0.7376	2	+Mantis	88	59	0.962
A	UAVSEAD ●	pj_better_10m ◊	0.7401	1	same	85	59	0.962
A	UR5E ●	pj_better_10m ◊	0.7404	1	same	87	59	0.981
A	KUKA ●	pj_better_10m ◊	0.7412	1	same	85	58	0.962
FE	stickslip_poincloux	pj_better_10m ◊	0.7390	1	same	85	63	0.981
FE	minati_circuits	pj_better_10m ◊	0.7367	1	same	87	61	0.962
FE	geyser_strokkur	pj_better_10m ◊	0.7390	1	same	84	63	0.962
FE	ndbc_buoy_v2_b	pj_better_10m ◊	0.7400	1	same	85	60	0.981
FE	nasa_battery_rw	pj_better_10m ◊	0.7350	1	same	84	60	0.962
FE	wind_genbearing_hot	pj_better_10m ◊	0.7404	1	same	85	60	0.962
FE	towie_genbearing_hot	pj_better_10m ◊	0.7387	1	same	84	59	0.981
FE	memristor_jerk	pj_better_10m ◊	0.7340	1	same	85	60	0.981
FE	noaa_coops_surge_b	pj_better_10m ◊	0.7384	1	same	85	59	0.981
FE	asos_lows_b	pj_better_10m ◊	0.7392	1	same	85	59	0.981
FE	usgs_reach_v2_b	pj_better_10m ◊	0.7390	1	same	85	60	1.000
FE	rotary_utrecht	pj_better_10m ◊	0.7400	1	same	85	60	0.981
FE	megathrust_sync_v2	pj_better_10m ◊	0.7381	1	same	85	59	0.981
FE	cdip_waves_c	pj_better_10m ◊	0.7392	1	same	84	59	0.962
FE	heatpump_defrost ●	pj_better_10m ◊	0.7389	1	same	82	61	1.000

Rank of every ranked model on the board without each dataset (main rank in the header)

without	pj_better_10m ◊ (1)	pj_better_3p7m ◊ (2)	pjgood-warp-xview-combo-10M ◊ (3)	Mantis (4)	Chronos-2 (5)	LeSaturn-warppool-factorized-7.4M ◊ (6)	pj_better_1m ◊ (7)	MantisV2 (8)	NuTime (9)	Timestral-209M (10)	Timestral-123M (11)	Toto-Open-Base-1.0 (12)	Timestral-28M (13)	Timestral-5M (14)	TimesFM-2.5-200M (15)
C:MoteStrain	1	3	2	4	5	6	7	8	9	10	11	12	14	13	15
C:SonyAIBORobotSurface2	1	2	3	4	5	7	6	8	9	10	11	12	13	14	15
C:Wafer	1	3	2	4	5	6	7	8	9	10	11	12	13	14	15
C:ACSF1	1	2	3	4	5	6	7	8	9	10	11	12	13	14	15
C:PLAID	1	2	3	4	5	6	7	8	9	10	11	12	14	13	15
C:FreezerRegularTrain	1	3	2	4	5	6	7	8	9	10	11	12	14	13	15
C:FordA	1	2	3	4	5	6	7	8	9	10	11	12	13	14	15
C:ElectricDevices	1	2	3	4	5	6	7	8	9	10	11	12	13	14	15
C:lendb	1	3	2	4	5	6	7	8	9	11	12	10	13	14	15
C:hydraulic	1	2	3	4	5	6	7	8	9	11	12	10	13	14	15
C:gas251	1	2	3	4	5	6	7	8	9	10	11	12	13	14	15
C:stead	1	2	3	4	5	6	7	8	9	10	11	12	14	13	15
C:AsphaltPavementTypeCoordinates	1	2	3	4	5	6	7	8	9	10	11	12	13	14	15
C:aursad095	1	2	3	4	5	6	7	8	9	10	11	12	13	14	15
C:voraus095	1	3	2	4	5	6	7	8	9	10	11	12	14	13	15
C:chillan5	1	2	3	4	5	6	7	8	9	10	11	12	13	14	15
C:haptic_texture8	1	2	3	4	5	6	7	8	9	10	11	12	13	14	15
C:openrainer_m	1	3	2	4	5	6	7	8	9	10	11	14	12	13	15
C:poseit_ft	1	2	3	4	5	6	7	8	9	10	11	12	13	14	15
C:px4_airframe3	1	2	3	4	5	6	7	8	9	10	11	12	14	13	15
C:swiss_load10	1	2	3	4	5	6	7	8	9	10	11	13	12	14	15
C:ved_powertrain3	1	2	3	4	5	6	7	8	9	10	11	12	13	14	15
C:bdg2_meter6	1	2	3	4	5	6	7	8	9	10	11	12	14	13	15
C:robo6_rb6	1	2	3	4	5	6	7	8	9	10	11	12	14	13	15
A:HAI	1	2	3	4	5	6	7	8	9	10	11	12	14	13	15
A:UKWIND	1	3	2	4	5	6	7	8	9	10	12	11	13	14	15
A:HOT	1	3	2	4	5	6	7	8	9	10	11	12	13	14	15
A:PREDIST	1	2	3	4	5	6	7	8	9	10	12	11	14	13	15
A:NOBOOM	1	3	2	4	5	6	7	8	9	10	11	14	12	13	15
A:ESA	1	2	3	4	5	6	7	8	9	10	11	12	14	13	15
A:NASA	1	3	4	2	5	6	7	8	9	10	11	12	13	14	15
A:UAVSEAD	1	2	3	4	5	6	7	8	9	10	11	14	12	13	15
A:UR5E	1	2	3	4	5	6	7	8	9	10	12	11	13	14	15
A:KUKA	1	2	3	4	5	6	7	8	9	10	11	14	12	13	15
FE:stickslip_poincloux	1	2	3	4	5	6	7	8	9	10	11	12	14	13	15
FE:minati_circuits	1	2	3	4	5	6	7	8	9	10	11	13	14	12	15
FE:geyser_strokkur	1	3	2	4	5	6	7	8	9	10	11	12	14	13	15
FE:ndbc_buoy_v2_b	1	3	2	4	5	6	7	8	9	10	11	12	13	14	15
FE:nasa_battery_rw	1	2	3	4	5	6	7	8	9	10	12	11	14	13	15
FE:wind_genbearing_hot	1	3	2	4	5	6	7	8	9	10	11	12	14	13	15
FE:towie_genbearing_hot	1	3	2	4	5	6	7	8	9	10	11	12	13	14	15
FE:memristor_jerk	1	2	3	4	5	6	7	8	9	10	11	12	14	13	15
FE:noaa_coops_surge_b	1	2	3	4	5	6	7	8	9	10	11	13	12	14	15
FE:asos_lows_b	1	2	3	4	5	6	7	8	9	10	11	13	12	14	15
FE:usgs_reach_v2_b	1	2	3	4	5	6	7	8	9	10	11	12	13	14	15
FE:rotary_utrecht	1	2	3	4	5	6	7	8	9	10	11	12	14	13	15
FE:megathrust_sync_v2	1	2	3	4	5	6	7	8	9	10	11	12	14	13	15
FE:cdip_waves_c	1	3	2	4	5	6	7	8	9	10	11	13	12	14	15
FE:heatpump_defrost	1	2	3	4	5	6	7	8	9	10	11	12	13	14	15

Admission of 0.9.7

A. Every candidate was measured without a model before any cell on it (Amendment 13 item 8): Eval series, anomaly segments in them, median share of the scored part that is labelled, and the better of the rising and falling elapsed-time scores. Rules: (a) at least 15 Eval series, (b) at least 20 segments, (c) median share at most 0.3; new datasets also need an elapsed-time best of at most 0.65.

candidate	status		cap	Eval series	segments	median share	elapsed-time best	rules	decided (UTC)	reason
ESA	in	existing	100	95	97	0.032	0.611	a:yes b:yes c:yes gate:yes	–	items 2(a)-(c) hold: 95 Eval, 97 segments, median share 0.032; elapsed-time best 0.611 (gate applies to new datasets only)
HAI	in	existing	100	58	103	0.030	0.533	a:yes b:yes c:yes gate:yes	2026-11-05T19:21Z	items 2(a)-(c) hold: 58 Eval, 103 segments, median share 0.030; elapsed-time best 0.533 (gate applies to new datasets only)
HOT	in	existing	100	50	107	0.014	0.529	a:yes b:yes c:yes gate:yes	–	items 2(a)-(c) hold: 50 Eval, 107 segments, median share 0.014; elapsed-time best 0.529 (gate applies to new datasets only)
KUKA	in	new	300	255	534	0.108	0.519	a:yes b:yes c:yes gate:yes	2026-11-05T19:23Z	items 2(a)-(c) hold: 255 Eval, 534 segments, median share 0.108; elapsed-time best 0.519 (gate 0.65 holds)
NASA	in	existing	100	37	47	0.040	0.516	a:yes b:yes c:yes gate:yes	–	items 2(a)-(c) hold: 37 Eval, 47 segments, median share 0.040; elapsed-time best 0.516 (gate applies to new datasets only)
NOBOOM	in	existing	300	20	37	0.191	0.561	a:yes b:yes c:yes gate:yes	–	items 2(a)-(c) hold: 20 Eval, 37 segments, median share 0.191; elapsed-time best 0.561 (gate applies to new datasets only)
PREDIST	in	existing	100	41	41	0.051	0.512	a:yes b:yes c:yes gate:yes	–	items 2(a)-(c) hold: 41 Eval, 41 segments, median share 0.051; elapsed-time best 0.512 (gate applies to new datasets only)
UAVSEAD	in	new	300	101	147	0.083	0.585	a:yes b:yes c:yes gate:yes	2026-11-05T18:56Z	items 2(a)-(c) hold: 101 Eval, 147 segments, median share 0.083; elapsed-time best 0.585 (gate 0.65 holds)
UKWIND	in	existing	100	17	41	0.057	0.558	a:yes b:yes c:yes gate:yes	–	items 2(a)-(c) hold: 17 Eval, 41 segments, median share 0.057; elapsed-time best 0.558 (gate applies to new datasets only)
UR5E	in	new	300	63	74	0.191	0.616	a:yes b:yes c:yes gate:yes	2026-11-05T18:54Z	items 2(a)-(c) hold: 63 Eval, 74 segments, median share 0.191; elapsed-time best 0.616 (gate 0.65 holds)
BASIC	out	new	300	–	–	–	–	a:– b:– c:– gate:–	2026-11-05T19:12Z	check 1 / item 4 condition: the flights are software-in-the-loop, not real (Ahmad and Akram, Data in Brief 2024, doi:10.1016/j.dib.2024.110069: 'The BASiC dataset is exclusively derived from simulations conducted in a SITL environment using the ArduPilot platform'); not built
CAFUC2	out	new	300	96	286	0.021	0.508	a:yes b:yes c:yes gate:yes	2026-11-05T18:56Z	check 2 (physical label; PREREG09 note of 2026-10-31 ~20:50Z: injected or staged faults pass when the onset is observed in the system): CAFUC2's faults are synthetic signals written into five copies of each recorded flight after the flight; no aircraft experienced them, so the label is not a state of or an event in the physical system. Decided before any cell; reversible by Jonas (needs-you)
CRANFIELD	out	new	–	–	–	–	–	a:– b:– c:– gate:–	2026-11-05T19:23Z	the official download (MATLAB File Exchange #50938) needs a MathWorks sign-in; the other copies (Kaggle, IEEE DataPort) need a login; item 4 / Amendment 13: sets that need a registration are not added
EAWAG	out	new	–	–	–	–	–	a:– b:– c:– gate:–	2026-11-05T18:55Z	only 4 ponds (2D, 3D, 5A, 5B), phase 1, carry the expert 'Anomaly' column, with 5 segments in total: rules 2(a) and 2(b) cannot hold; the licence field is empty (portal: public domain for data outside Swiss copyright)
GREEKWIND	out	new	100	7	8	0.012	0.501	a:no b:no c:yes gate:yes	2026-11-05T19:23Z	2(a): 7 Eval series < 15; 2(b): 8 anomaly segments < 20
MAST	out	new	–	–	–	–	–	a:– b:– c:– gate:–	2026-11-05T18:55Z	no precursor-only cut without an elapsed-time leak: a scored block that ends before the current quench carries the precursor label at its end (a rising elapsed-time score reaches 0.933 on synthetic series of that shape); TokaMark ships no disruption labels or quench times
ONBOARDEU	out	new	100	85	94	0.173	0.708	a:yes b:yes c:yes gate:no	2026-11-05T19:19Z	item 4 elapsed-time gate: 0.708 > 0.65
PMU	out	new	300	–	–	–	–	a:– b:– c:– gate:–	2026-11-05T19:22Z	the continuous 1 s TransnetBW record (Zenodo 15784548) is one channel (CE frequency) and ends 2024-04-25, covering 23 of 88 cases; P4 needs >= 2 channels and 23 cases give at most 19 Eval series and 19 segments (rule 2(b)); not built
RFLYMAD	out	new	300	255	405	0.289	0.731	a:yes b:yes c:yes gate:no	2026-11-05T19:12Z	item 4 elapsed-time gate: 0.731 > 0.65
TCABR	out	new	300	138	138	0.199	1.000	a:yes b:yes c:yes gate:no	2026-11-05T19:22Z	item 4 elapsed-time gate: 1.000 > 0.65

Admission waves, as written before any 0.9.7 cell on their units

wave	written (UTC)	in	out	pending
0	2026-11-05T18:46Z	HOT	none	none
1	2026-11-05T18:56Z	ESA, NASA, UKWIND, HOT, PREDIST, NOBOOM, UAVSEAD and UR5E	EAWAG, CAFUC2 and MAST	HAI, RFLYMAD, PMU, ONBOARDEU, TCABR, BASIC, KUKA, GREEKWIND and CRANFIELD
C1	2026-11-05T19:10Z	nicola15/bdg2_meter6 and nicola15/robo6_rb6	none	nicola15/sns_hvcm_b_end
2	2026-11-05T19:12Z	none	RFLYMAD and BASIC	none
C2	2026-11-05T19:13Z	none	nicola15/sns_hvcm_b_end	none
3	2026-11-05T19:21Z	HAI	ONBOARDEU	none
4	2026-11-05T19:23Z	KUKA	TCABR, PMU, GREEKWIND and CRANFIELD	none
FE1	2026-11-05T19:44Z	heatpump_defrost	none	none
FE2	2026-11-05T21:59Z	ipshocks_fluxes	pyrnet	none

FE. Model-free gates (Amendment 13 item 5; thresholds computed 2026-11-05T18:53:31Z on the 14 kept FE datasets, before any new FE download): at least 300 events, at least 30 independent units or days, time-since-last-event row at most 0.7956, time-of-day row at most 0.5851, best free forecast at most 0.85.

The reference rows on the kept FE datasets (the source of the thresholds)

dataset	time since last event	time of day	position only	event rate	best free forecast
stickslip_poincloux	0.685	no calendar	0.498	0.595	early-warning statistics 0.701
minati_circuits	0.569	no calendar	0.507	0.501	curvature 0.757
geyser_strokkur	0.734	0.504	0.490	0.611	early-warning statistics 0.730
ndbc_buoy_v2_b	0.549	0.476	0.506	0.463	current value 0.688
nasa_battery_rw	0.796	0.512	0.517	0.676	current value 0.873
wind_genbearing_hot	0.651	0.529	0.534	0.594	current value 0.704
towie_genbearing_hot	0.538	0.585	0.566	0.504	level statistics 0.752
memristor_jerk	0.635	no calendar	0.510	0.505	curvature 0.882
noaa_coops_surge_b	0.654	0.503	0.495	0.533	current value 0.797
asos_lows_b	0.508	0.492	0.491	0.436	trend 0.743
usgs_reach_v2_b	0.505	0.496	0.481	0.486	early-warning statistics 0.652
rotary_utrecht	0.582	no calendar	0.503	0.520	level statistics 0.647
megathrust_sync_v2	0.759	no calendar	0.496	0.562	level statistics 0.674
cdip_waves_c	0.566	0.520	0.515	0.494	current value 0.738

candidate	status	events	units or days	time since last event	time of day	best free forecast	reason
heatpump_defrost	in	53,978	113	0.659	0.566	0.729	item 5 gates hold: 53,978 events >= 300; 113 systems >= 30; time since last event 0.659 <= 0.796; time of day 0.566 <= 0.585; best free forecast 0.729 <= 0.85
ngafid_hastall	out	–	–	–	–	–	Amendment 13 item 5
substorm_omni	out	–	–	–	–	–	Amendment 13 item 5
epa_aqs_pm25	out	–	–	–	–	–	Amendment 13 item 5
ipshocks_fluxes	admitted by item 5; off the board by check 3 (FE at least 1,000 events)	914	629	0.490	0.513	0.695	item 5 gates hold: 914 events; 629 distinct shock days (both spacecraft at L1, counted by day); time since last event 0.490; time of day 0.513; best free forecast 0.695
pyrnet	out	–	–	–	–	–	not built and gated by the 21:30Z cutoff (the 4.8 GB download finished 21:58Z); builder written, never run

ipshocks_fluxes passed the item 5 gates (at least 300 events) but has fewer than the 1,000 events of the standing admission check 3, which Amendment 13 item 7 keeps; the board leaves it out, so 0.9.7 has 15 FE datasets. Neither text said whether the 300-event gate replaces check 3 for new FE datasets; Jonas Petersen decided that ipshocks_fluxes stays out of 0.9.7, with the board that includes it known (PREREG09 note of 2026-11-06 11:05Z; see the sensitivities of the test above). Its cells exist and are kept.

C. New C datasets were rebuilt from their builders and accepted only if the sha256 of X equals the logged one (Amendment 13 item 6), after the exp-nicola-15 screening that scored a ◊ row on them (see What changed).

dataset	status	kind	X sha256 equal	classes	windows	groups	licence	reason
nicola15/bdg2_meter6	in	new (item 6)	yes	6	75,409	1,615	CC BY-SA 4.0 (github.com/buds-lab/building-data-genome-project-2 LICENSE)	rebuilt from the exp-nicola-15 builder; X sha256 equals LOG.md (item 6)
nicola15/robo6_rb6	in	new (item 6)	yes	6	6,869	5,352	MIT (Hugging Face dataset card kyselica/RoBo6)	rebuilt from the exp-nicola-15 builder; X sha256 equals LOG.md (item 6)
nicola15/sns_hvcm_b_end	out	new (item 6)	yes	2	3,323	15	CC BY 4.0 (Mendeley Data kbbrw99vh8 v5)	item 6 gate: combo-10M minus random twin positive in 3 of 5 repeats (needs 4) and mean +0.0169 (needs +0.02); results/cgate097.json
mv06/3w	out	dropped (item 6)	–	–	–	–	–	Amendment 13 item 6
mv06/bosch	out	dropped (item 6)	–	–	–	–	–	Amendment 13 item 6
nicola15/soredd_abc	out	dropped (item 6)	–	–	–	–	–	Amendment 13 item 6

Combo minus twin on every C dataset (Amendment 13 item 6): pjgood-warp-xview-combo-10M minus the random-init LeSaturn row (3.7 M parameters, another architecture; not a matched twin of the 10.6 M combo), per repeat, under the C head; the criterion is positive in at least 4 of 5 repeats and a mean of at least +0.02. It decided only the new SNS HVCM unit; a kept dataset that fails it is reported here and not dropped.

C dataset	positive repeats	of	mean difference	per repeat	criterion
nicola15/sns_hvcm_b_end (candidate)	3	5	+0.0169	-0.019 +0.016 +0.047 -0.015 +0.055	fails
ucr/MoteStrain	1	1	+0.0279	+0.028	n/a: one fixed split
ucr/SonyAIBORobotSurface2	1	1	+0.2475	+0.247	n/a: one fixed split
ucr/Wafer	1	1	+0.0209	+0.021	n/a: one fixed split
ucr/ACSF1	0	1	-0.0000	-0.000	n/a: one fixed split
ucr/PLAID	1	1	+0.1064	+0.106	n/a: one fixed split
ucr/FreezerRegularTrain	1	1	+0.1232	+0.123	n/a: one fixed split
ucr/FordA	1	1	+0.0229	+0.023	n/a: one fixed split
ucr/ElectricDevices	1	1	+0.0263	+0.026	n/a: one fixed split
phys/lendb	5	5	+0.1144	+0.103 +0.131 +0.122 +0.102 +0.114	passes
mv06/hydraulic	5	5	+0.0706	+0.077 +0.074 +0.096 +0.039 +0.066	passes
mv06/gas251	5	5	+0.0979	+0.107 +0.152 +0.099 +0.069 +0.062	passes
mv06/stead	5	5	+0.0614	+0.077 +0.072 +0.056 +0.053 +0.049	passes
uea/AsphaltPavementTypeCoordinates	1	1	+0.0797	+0.080	n/a: one fixed split
mv06/aursad095	5	5	+0.0850	+0.078 +0.098 +0.093 +0.070 +0.086	passes
mv06/voraus095	5	5	+0.0839	+0.069 +0.080 +0.110 +0.072 +0.089	passes
nicola15/chillan5	5	5	+0.0605	+0.056 +0.051 +0.076 +0.051 +0.069	passes
nicola15/haptic_texture8	5	5	+0.0545	+0.054 +0.044 +0.069 +0.048 +0.057	passes
nicola15/openrainer_m	5	5	+0.0424	+0.058 +0.031 +0.025 +0.039 +0.059	passes
nicola15/poseit_ft	5	5	+0.0869	+0.164 +0.081 +0.044 +0.051 +0.094	passes
nicola15/px4_airframe3	5	5	+0.0761	+0.087 +0.067 +0.087 +0.050 +0.090	passes
nicola15/swiss_load10	5	5	+0.0476	+0.069 +0.047 +0.045 +0.043 +0.034	passes
nicola15/ved_powertrain3	5	5	+0.0736	+0.077 +0.072 +0.074 +0.055 +0.089	passes
nicola15/bdg2_meter6	5	5	+0.0329	+0.050 +0.018 +0.040 +0.029 +0.027	passes
nicola15/robo6_rb6	5	5	+0.0850	+0.079 +0.084 +0.085 +0.083 +0.094	passes

What took how long

From the run's event log (timing097.json, UTC): the first event is at 2026-11-05T18:33:41Z and the last at 2026-11-06T10:13:14Z, 15.66 hours apart. Builds and measurements ran in parallel, so the spans below overlap and do not add up to the run.

Phases: 4, 835.3 minutes in total, the longest phase2 (615.0 min).

what	start	end	minutes	note
phase0	18:33:41	18:42	8.3	setup, probes, dashboard, runner; ended when builders were launched (~18:42Z); logged late
phase1	18:42	21:59	197.0	final FE list (PREREG09 note 21:59Z); logged after the outage
phase2	18:49	05:04	615.0	last cell job 3868626 ended (unattended, VM offline); tabular references queued separately 2026-11-06
phase3	09:17	09:32	15.0

Downloads: 27, 407.9 minutes in total, the longest fe_pyrnet (182.5 min).

what	start	end	minutes	note
UAVSEAD	18:38:58	18:40:31	1.6
UR5E	18:39:00	18:46:24	7.4
BASIC	18:39:08	18:40:03	0.9
TCABR	18:39:33	18:55:30	15.9
HAI	18:40:13	18:42:48	2.6
KUKA	18:40:13	18:59:19	19.1
RFLYMAD	18:40:25	18:57:42	17.3
EAWAG	18:40:26	18:49:38	9.2
ONBOARDEU	18:41:04	18:54:25	13.3
PMU	18:41:14	18:45:05	3.9
CAFUC2	18:41:29	18:42:12	0.7
CRANFIELD	18:41:48	18:46:10	4.4
GREEKWIND	18:41:48	18:46:10	4.4
PREDIST	18:42:31	18:44:26	1.9
C_robo6	18:42:32	19:04:32	22.0	fetch_inside_build_robo6.py_end_is_card_mtime
C_bdg2	18:43:14	18:43:45	0.5
C_sns-hvcm	18:43:14	18:44:56	1.7
NOBOOM	18:44:47	18:46:28	1.7
MAST	18:51:56	18:54:51	2.9	out_no_build
fe_ipshocks	18:54:51	19:45:34	50.7
fe_heatpump	18:55:01	19:10:46	15.8
fe_pyrnet	18:55:42	21:58:12	182.5
UTAQ (0.9.7.1 work, Amendment 14)	22:36:20	22:40:44	4.4
PVLAZ (0.9.7.1 work, Amendment 14)	22:36:42	22:37:37	0.9
UKWIND (0.9.7.1 work, Amendment 14)	22:36:55	22:54:36	17.7
RCFPOWER (0.9.7.1 work, Amendment 14)	22:38:11	22:40:39	2.5
NOBOOM_M202210 (0.9.7.1 work, Amendment 14)	22:42:18	22:44:27	2.1

Builds: 22, 97.5 minutes in total, the longest C_robo6 (22.0 min).

what	start	end	minutes	note
HOT_reselect	18:40:18	18:40:42	0.4
C_robo6	18:42:32	19:04:32	22.0	end_is_card_mtime
C_bdg2	18:43:45	18:45:09	1.4	end_logged_late_actual_end_is_card_mtime
UAVSEAD	18:43:47	18:53:11	9.4
HAI	18:43:54	18:56:25	12.5
PREDIST	18:44:31	18:48:37	4.1
C_sns-hvcm	18:44:56	18:45:14	0.3	end_logged_late_actual_end_is_card_mtime
CAFUC2	18:45:31	18:46:38	1.1
NOBOOM	18:46:32	18:46:37	0.1
UR5E	18:50:24	18:52:11	1.8
select_UR5E	18:52:23	18:53:08	0.8
ONBOARDEU	18:54:30	19:02:46	8.3
GREEKWIND	18:59:23	18:59:40	0.3
KUKA	18:59:40	19:09:07	9.4
RFLYMAD	19:00:47	19:08:21	7.6
fe_heatpump	19:10:46	19:12:32	1.8
TCABR	19:11:50	19:20:09	8.3
fe_ipshocks	19:45:34	19:47:15	1.7
PVLAZ (0.9.7.1 work, Amendment 14)	22:37:37	22:37:45	0.1
RCFPOWER (0.9.7.1 work, Amendment 14)	22:40:39	22:40:47	0.1
UTAQ (0.9.7.1 work, Amendment 14)	22:40:53	22:46:48	5.9
NOBOOM_M202210 (0.9.7.1 work, Amendment 14)	22:44:27	22:44:35	0.1

Model-free measurements: 21, 50.2 minutes in total, the longest ESA (10.8 min).

what	start	end	minutes	note
ESA	18:41:10	18:51:55	10.8	ESA_NASA_UKWIND
fe_F3	18:45:57	18:54:32	8.6
CAFUC2	18:46:45	18:49:26	2.7
PREDIST	18:50:34	18:51:01	0.5
NOBOOM	18:51:01	18:51:19	0.3
UR5E	18:53:20	18:53:51	0.5
UAVSEAD	18:54:57	18:55:43	0.8
HAI	18:57:41	18:58:19	0.6
GREEKWIND	18:59:49	19:00:08	0.3
KUKA	19:09:57	19:19:45	9.8
RFLYMAD	19:10:03	19:10:59	0.9
fe_heatpump	19:12:32	19:13:02	0.5
fe_heatpump_r2	19:13:15	19:20:05	6.8
ONBOARDEU	19:16:24	19:17:07	0.7
HAI	19:20:05	19:20:32	0.5
TCABR	19:20:22	19:21:31	1.1
HAI_withEnd	19:20:32	19:21:14	0.7
PVLAZ (0.9.7.1 work, Amendment 14)	22:37:50	22:37:56	0.1
RCFPOWER (0.9.7.1 work, Amendment 14)	22:40:54	22:41:28	0.6
NOBOOM_M202210 (0.9.7.1 work, Amendment 14)	22:45:09	22:45:51	0.7
UTAQ (0.9.7.1 work, Amendment 14)	22:46:52	22:49:37	2.8

Started and not logged as ended when this card was built: crefs_C_new (since 19:11:13), build_UKWIND (0.9.7.1 work, Amendment 14) (since 22:38:07), download_REASSEMBLE (0.9.7.1 work, Amendment 14) (since 22:46:07) and phase4 (since 10:13:14).

Cluster jobs of the run (28, from sacct; node-hours = elapsed x nodes):

job	name	partition	state	start (UTC)	end (UTC)	task list	node-hours
3867625	s097	debug	COMPLETED	11-05 18:49Z	11-05 18:53Z	jobs097_smoke.txt	0.07
3867627	p097	preemptable	COMPLETED	11-05 18:50Z	11-05 18:55Z	jobs097_w0.txt	0.08
3867656	p097	preemptable	PREEMPTED	11-05 18:58Z	11-05 19:17Z	jobs097_w1.txt	0.31
3867709	g097	preemptable	COMPLETED	11-05 19:08Z	11-05 19:08Z	–	0.01
3867710	c097	preemptable	CANCELLED	–	11-05 19:08Z	–	0.00
3867712	c097	preemptable	COMPLETED	11-05 19:09Z	11-05 19:10Z	–	0.02
3867715	g097	preemptable	COMPLETED	11-05 19:09Z	11-05 19:10Z	–	0.01
3867716	c097	preemptable	COMPLETED	11-05 19:10Z	11-05 19:11Z	–	0.01
3867721	e097	preemptable	COMPLETED	11-05 19:10Z	11-05 19:11Z	–	0.02
3867722	c097	preemptable	COMPLETED	11-05 19:15Z	11-05 19:16Z	–	0.01
3867750	gate097	preemptable	COMPLETED	11-05 19:15Z	11-05 19:16Z	–	0.02
3867751	c097	preemptable	COMPLETED	11-05 19:17Z	11-05 19:17Z	–	0.01
3867791	p097	preemptable	COMPLETED	11-05 19:21Z	11-05 19:29Z	–	0.12
3867812	p097	preemptable	PREEMPTED	11-05 19:26Z	11-05 19:34Z	–	0.15
3867837	p097	preemptable	COMPLETED	11-05 19:26Z	11-05 19:41Z	–	0.26
3867848	c097	preemptable	COMPLETED	11-05 19:26Z	11-05 19:26Z	–	0.01
3867960	p097	preemptable	COMPLETED	11-06 04:04Z	11-06 04:04Z	–	0.01
3867977	f097	preemptable	CANCELLED	–	11-05 22:35Z	–	0.00
3867978	fe097	preemptable	CANCELLED	–	11-05 22:35Z	–	0.00
3868446	r097	preemptable	CANCELLED	–	11-05 22:35Z	–	0.00
3868447	fe097	preemptable	CANCELLED	–	11-05 22:35Z	–	0.00
3868625	a097	preemptable	COMPLETED	11-06 04:06Z	11-06 05:01Z	–	0.90
3868626	a097	preemptable	COMPLETED	11-06 04:06Z	11-06 05:04Z	–	0.97
3868685	a097	preemptable	COMPLETED	11-06 00:38Z	11-06 01:16Z	–	0.63
3871836	t097tab	debug	COMPLETED	11-06 09:36Z	11-06 09:48Z	jobs097_tabsmoke.txt	0.21
3872079	t097tab	preemptable	CANCELLED	–	11-06 10:00Z	–	0.00
3872371	t097tab	preemptable	PENDING	–	–	jobs097_tab.txt	0.00
3872372	t097tab	preemptable	PENDING	–	–	jobs097_tab.txt	0.00

Other logged events

time	event	note
18:33:41	run_start	session launched ~18:33Z
18:38:13	measure097_validated	measure097 reproduces nulls095 HOT 29/29 max diff 1.1e-16
18:45:00	download_C_sns-hvcm	bytes_correction_previous_end_entry
18:47:09	build_UAVSEAD_restart	added_asset_field
18:50:41	gate_A_repro	A6g through CAFE-097 tree on Clariden: ZERO NOBOOM 20/20, pj_better_3p7m PREDIST 29/29 equal to 0.9.6 cells, max diff 0; smoke job 3867625
19:29:20	phase1_A_C_done	final A list and C admission written by 19:23Z; FE pending
10:13:14	outage	VM froze 2026-11-05 ~22:56Z (memory), network lost 23:11Z, reset 2026-11-06 09:01Z; agents stopped 00:16Z-02:55Z (DECISIONS097.md)

Inference and benchmarking time

FE feature-extraction time on the new FE datasets is not measured for 28 model-dataset tasks: their tasks saved the features and then ended with an error in a scoring step the board does not use (DECISIONS097.md, 2026-11-06 09:35Z), so they have no clean timing record. Their FE head time is measured (the FT tasks).

Per model, from the task-queue records of the cell jobs (timing097_infer.json, timing097.py). task wall = end - start of one queue task (one row on one dataset: embedding + head + scoring, with up to NW tasks sharing one GH200, so it is an upper bound of the GPU time of that task); A throughput = Eval series per second and scored windows per second of task wall; GPU-hours = sum of task walls / 3600 (an upper bound: tasks share GPUs); node-hours = sacct elapsed x nodes. Seconds are the sum over the model's tasks of each kind. Time is reported, never scored.

model	parameters	A cells, s	A gate, s	A reference, s	C embedding (Toto m1), s	C embedding (menu row), s	C embedding (stored row), s	FE cells, s	FE cells (head on stored features), s	FE embedding, s	FE gate, s	FE reference, s	A windows / s	A series / s	GPU-hours, upper bound
Chronos-2	119.5 M	2,600	81	–	–	63	–	–	34	–	–	–	85	0.218	0.772
Timestral-209M	209.3 M	1,876	–	–	–	32	–	–	31	–	–	–	117	0.303	0.539
TimesFM-2.5-200M	231.3 M	1,793	–	–	–	–	20	–	39	–	–	–	123	0.317	0.515
Toto-Open-Base-1.0	151.3 M	1,545	–	–	28	–	–	–	35	–	–	–	142	0.368	0.447
Timestral-123M	122.6 M	1,436	–	–	–	32	–	–	28	–	–	–	153	0.396	0.415
MantisV2	–	1,123	–	–	–	26	–	–	37	–	–	–	196	0.506	0.330
rawknn	–	–	–	1,188	–	–	–	–	–	–	–	–	–	–	0.330
Mantis	8.0 M	889	–	–	–	23	–	–	38	–	–	–	247	0.639	0.264
Timestral-28M	27.8 M	795	–	–	–	26	–	–	26	–	–	–	277	0.715	0.235
pjgood-warp-xview-combo-10M ◊	10.6 M	595	–	–	–	33	–	–	26	3	–	–	370	0.954	0.182
pj_better_10m ◊	10.6 M	571	–	–	–	19	–	–	26	3	–	–	385	0.995	0.172
LeSaturn-warppool-factorized-7.4M ◊	7.4 M	376	–	–	–	22	–	–	30	3	–	–	585	1.510	0.120
pj_better_3p7m ◊	3.8 M	322	13	–	–	25	–	–	27	3	27	–	683	1.763	0.116
Timestral-5M	5.2 M	340	–	–	–	24	–	–	33	–	–	–	646	1.669	0.110
NuTime	2.4 M	341	–	–	–	–	17	–	29	–	–	–	645	1.667	0.108
Random-init LeSaturn (control)	3.7 M	274	–	–	–	–	19	6	28	–	–	–	803	2.073	0.091
ZERO	0	181	12	–	–	–	9	–	32	–	35	–	1,215	3.138	0.075
pj_better_1m ◊	1.0 M	177	–	–	–	13	–	–	23	3	–	–	1,244	3.212	0.060
FEREF	–	–	–	–	–	–	–	–	–	–	–	87	–	–	0.024
levelonly	–	–	–	46	–	–	–	–	–	–	–	–	–	–	0.013
maxz	–	–	–	21	–	–	–	–	–	–	–	–	–	–	0.006
maxdz	–	–	–	17	–	–	–	–	–	–	–	–	–	–	0.005

The cluster jobs of the run used 3.82 node-hours (15.29 GPU-hours at 4 GH200 per node), from sacct.

A material per dataset (the denominators of the A throughput)

A dataset	Eval series	scored windows	rows
ESA	95	37,929	1,187,429
NASA	37	14,049	234,147
UKWIND	17	6,633	177,245
HOT	50	19,818	524,844
PREDIST	41	15,980	220,652
NOBOOM	20	7,750	130,077
HAI	58	22,887	688,644
UAVSEAD	101	37,668	404,044
UR5E	63	23,617	327,889
KUKA	255	99,952	2,933,620

Earlier cost table, for comparison (costs09.json, CAFE 0.9.1 datasets and settings, One A100 40 GB, 6 cores. 0.9.1: A measured on a shared GPU; part of the board estimated (see the text). The hardware and the datasets differ from 0.9.7, so compare orders of magnitude only):

model	C, min	A, min	FE, min	total, min	FE measured
TimesFM-2.5-200M	18.7	112.4	74.8	206	yes
TiRex-2	4.5	116.2	63.5	184	scaled
Chronos-2	6.1	99.4	66.1	172	scaled
Timestral-209M	9.7	42.6	69.2	121	scaled
Timestral-123M	7.1	43.1	59.8	110	scaled
Toto-Open-Base-1.0	12.4	16.4	37.8	67	yes
Timestral-28M	2.7	12.6	45.9	61	scaled
Mantis	2.2	35.7	18.4	56	scaled
LeSaturn-scm-10M-s43	6.4	12.0	25.2	44	scaled
pjgood-warp-xview-combo-10M ◊	3.5	23.5	14.4	41	scaled
LeSaturn-scm-3.8M-s43	3.1	9.6	28.0	41	scaled
LeSaturn-scm-10M-s42	5.0	11.3	23.4	40	scaled
LeSaturn-scm-3.8M-s44	2.8	9.5	26.7	39	scaled
pj_better_10m ◊	3.2	20.6	13.3	37	yes
LeSaturn-scm-10M-s44	5.3	8.7	22.9	37	scaled
Timestral-5M	4.9	4.9	20.8	31	scaled
LeSaturn-scm-3.8M	6.8	10.4	11.6	29	scaled
LeSaturn-warppool-factorized-7.4M ◊	3.5	11.9	12.6	28	scaled
pj_better_3p7m ◊	2.6	13.0	10.5	26	scaled
NuTime	0.0	12.9	11.1	24	yes
Twin-LeSaturn-scm-3.8M	3.4	4.6	14.9	23	scaled
LeSaturn-warppool-factorized-3.8M	3.3	5.4	12.1	21	scaled
LeSaturn-warppool-factorized-3.8M-s43	3.2	5.2	11.7	20	scaled
ZERO	4.9	2.7	9.1	17	yes

Whole-board hours, runs one after another (C / A / FE): 0.8.6: 1.5 / 15.9 / 38.4 = 55.8; 0.8.6 protocol, same 24 runs: 2.63 / 25.91 / 54.25 = 82.79; 0.9: 2.09 / 3.52 / 4.15 = 9.76.

The 0.9.1 speed note (speed09.json)

Scoring one model on CAFE 0.9.1 (15 C, 10 A and 16 FE datasets) takes about 37 min for pj_better_10m and 17 min for ZERO on one A100, against 2.4 h and 2.0 h in 0.8.6. The slowest model, TimesFM-2.5-200M, takes about 3.4 h. The whole board, 24 runs (17 models, 5 extra pretraining seeds, 2 baselines) one after another, costs about 25 hours, against 83 h for the same runs under the 0.8.6 protocol (NuTime, not on the 0.8.6 board, is missing from that total, and its C time from the 0.9.1 total).

Where the time goes now. A and FE now cost about the same (10.7 h and 11.7 h for the board). On A the kNN head is cheap, but every window of every A series is embedded, and large encoders are slow at that (Chronos-2 1.7 h, TimesFM 1.9 h on A). The 9 FE datasets that 0.9.1 restores cost 520 s for pj_better_10m and 391 s for ZERO: on the fast path they are cheap, because their 0.8.6 cost was the head (10 seeds on the CPU), not the encoder.

The fast path changes no number. With one exception, every fresh cell equals the stored cell bit for bit: five models rerun from scratch on the 0.9.0 datasets, and pj_better_10m and ZERO on the 9 restored FE datasets (18 of 18 cells equal seeds 0–2 of their 0.8.6 cells). The new models' cells were each repeated on one dataset with identical results. The exception is one Mantis FE value, 8.9e-7 away from its 0.7-era cell, because of that run's forward pass.

How it was measured: one A100 40 GB with 6 physical cores. C and the 7 FE datasets of 0.9.0: 5 models timed one at a time on an otherwise idle machine, 19 estimated from their 0.8.6 times; the C times are for the 13 C datasets of that study, which lack hydraulic, 3w and FreezerRegularTrain of the 15 in 0.9.1 (and include FreezerSmallTrain). Restored FE: measured for pj_better_10m, ZERO, NuTime, Toto and TimesFM; the other rows scaled by the measured ratio. A: measured for every model except TiRex-2 (its cells were copied), on 12 A sources (the board uses 10) and on a GPU shared with other jobs, so the A times lean high. Basis per run in cafe09/results/costs09.json.

model	0.8.6 total	0.9.1 C	0.9.1 A	0.9.1 FE (16 datasets)	0.9.1 total
pj_better_3p7m	1.1 h	3 min	13 min	11 min	26 min
ZERO	2.0 h	5 min	3 min	9 min	17 min
pj_better_10m	2.4 h	3 min	21 min	13 min	37 min
Mantis	2.9 h	2 min	36 min	18 min	56 min
Chronos-2	4.6 h	6 min	99 min	66 min	172 min

C and the 7 FE datasets of 0.9.0 measured for all five; the 9 restored FE datasets measured for pj_better_10m and ZERO, scaled for the other three; A measured on a shared GPU. 0.8.6 totals of Mantis and Chronos-2 are modelled (0.8.6 study).

Cost per model

Minutes of GPU time per model (sum of its task wall times, an upper bound because up to several tasks share one GH200), on this version's cells only; reused cells of earlier versions cost nothing here. So the basis differs by row: rows differ in which of their tasks ran and left a clean record in this run (pj_better_3p7m and Chronos-2 also ran gate tasks; FE feature tasks on the new FE datasets count only where they ended cleanly, see above), and the front below compares this run's cost, not the cost of scoring a model on the whole of CAFE (see the cost table of the earlier versions for that). Cost is reported, never scored. The whole board of ranked rows took 4.38 GPU-hours by this measure; the cluster jobs used 3.82 node-hours.

model	seeds	CAFE	GPU minutes	on the front
pj_better_1m ◊	1	0.7241	3.6	yes
NuTime	1	0.7050	6.5
Timestral-5M	1	0.6943	6.6
pj_better_3p7m ◊	1	0.7341	7.0	yes
LeSaturn-warppool-factorized-7.4M ◊	1	0.7275	7.2
pj_better_10m ◊	1	0.7384	10.3	yes
pjgood-warp-xview-combo-10M ◊	1	0.7340	10.9
Timestral-28M	1	0.6943	14.1
Mantis	1	0.7326	15.8
MantisV2	1	0.7181	19.8
Timestral-123M	1	0.6969	24.9
Toto-Open-Base-1.0	1	0.6954	26.8
TimesFM-2.5-200M	1	0.6662	30.9
Timestral-209M	1	0.6989	32.3
Chronos-2	1	0.7301	46.3

On the front: no other ranked row has a higher CAFE score with the same or less GPU time.

Jobs and cells that did not finish

Logged as started and not as ended in the event log: 4 items (crefs_C_new, build_UKWIND, download_REASSEMBLE and phase4). Rows with a missing cell on a task board (unranked on that task): 23. The list is in data/unfinished097.json.

CAFE score against GPU minutes, one point per ranked model; a diamond is a ◊ row; the orange line joins the front, whose rows are labelled. Hover a point for its numbers. Cost is reported, never scored.

Change log

One section per version since 0.9.4, generated from the boards' dataset lists and the amendments; earlier releases follow in CHANGELOG.md.

v0.9.7 (2026-11-06)

24 C, 10 A and 15 FE datasets; 15 ranked models; CAFE = mean of C, A, FE; first: pj_better_10m.

C: 2 added (nicola15/bdg2_meter6, nicola15/robo6_rb6).
C: 3 removed or pooled (mv06/3w, mv06/bosch, nicola15/soredd_abc).
A: 5 added (ESA, NASA, UAVSEAD, UR5E, KUKA).
A: 9 removed or pooled (ESA_M1, ESA_M2, CARE_A, SMAP, MSL, CARE_C, RIGS, W3, SKAB).
FE: 1 added (heatpump_defrost).
FE: 3 removed or pooled (epa_aqs_pm25, substorm_omni, ngafid_hastall).
Pairs resolved on the CAFE score: 85 on the main board (10 within seed noise), 61 of 105 with the datasets resampled.
Amendment 13 (CAFE 0.9.7, written 2026-11-05 ~18:32Z (first headed ~18:40Z, a typo; commit adbffba1d at 18:32Z), before any 0.9.7 download, build, cell, or model read on a new unit; no 0.9.7 score exists). The heading's "before any 0.9.7 ... model read on a new unit" does not hold for the two C additions bdg2_meter6 and robo6_rb6: the exp-nicola-15 screening scored pjgood-warp-xview-combo-10M ◊ and its random-init twin on these exact builds at 09:01Z to 09:22Z on 2026-11-05, before the amendment was written (see New datasets and the conflict of interest on the card). Item 1 calls the 3W drop structural; it was made on combo minus twin (item 6 and the same screening), so six drops were score-informed.
The test of Amendment 13 item 9: not passed; the CAFE 1.0 candidate is 0.9.5.

v0.9.6

25 C, 14 A and 17 FE datasets; 15 ranked models; CAFE = mean of C, A, FE; first: pj_better_10m.

C: 8 added (nicola15/chillan5, nicola15/haptic_texture8, nicola15/openrainer_m, nicola15/poseit_ft, nicola15/px4_airframe3, nicola15/soredd_abc, nicola15/swiss_load10, nicola15/ved_powertrain3).
A: 2 added (PREDIST, NOBOOM).
FE: 1 added (ngafid_hastall).
Pairs resolved on the CAFE score: 73 on the main board (1 within seed noise), 51 of 105 with the datasets resampled.
Amendment 11 (CAFE 0.9.6, written 2026-11-04 ~21:45Z, before any 0.9.6 build of the A and FE datasets below, before any 0.9.6 cell, and before any model is run on the new A and FE data).
Amendment 12 (CAFE 0.9.6, written 2026-11-04 ~22:00Z, before any 0.9.6 cell and before any new C head is fitted; replaces the C list of Amendment 11 item 2 and adds a C head rule).

v0.9.5

17 C, 12 A and 16 FE datasets; 15 ranked models; CAFE = mean of C, A, FE; first: pj_better_10m.

C: 2 added (mv06/aursad095, mv06/voraus095).
C: 1 removed or pooled (mv06/voraus).
A: 5 added (ESA_M1, CARE_A, MSL, W3, SKAB).
Ranked rows no longer ranked: TiRex-2.
Pairs resolved on the CAFE score: 73 on the main board (1 within seed noise), 49 of 105 with the datasets resampled.
Amendment 6 (CAFE 0.9.5, written 2026-11-03 ~11:15Z, before any 0.9.5 cell, any 0.9.5 dataset build, or any model read on the new builds).
Amendment 7 (CAFE 0.9.5, written 2026-11-03 ~13:54Z, before any 0.9.5 A cell; replaces item 2 of Amendment 6).
Amendment 8 (CAFE 0.9.5, written 2026-11-03 ~14:32Z, before any 0.9.5 A cell; replaces item 2 of Amendment 7 and rule 4 of Amendment 5 as an admission rule).
Amendment 9 (CAFE 0.9.5, written 2026-11-03 ~16:40Z, before any 0.9.5 readout cell; replaces the readout rules of Amendment 5 rule 7 and the C long-sample rule of Amendment 5 rule 2).
Amendment 10 (CAFE 0.9.5, written 2026-11-03 ~17:09Z, before any 0.9.5 cell on voraus).

v0.9.4

16 C, 7 A and 16 FE datasets; 16 ranked models; CAFE = mean of C, A, FE; first: MantisV2.

Pairs resolved on the CAFE score: 39 on the main board (1 within seed noise), 16 of 120 with the datasets resampled.
Amendment 5 (CAFE 0.9.4, written 2026-11-02, first committed 21:35Z as 3be8b9d4; its heading first said ~22:20Z, a typo; before any cell under these rules exists).

Provenance of the rules

Every amendment since 0.9.4 and every dated note of 0.9.7, with the commit that first added it to PREREG09 (git log), the version it belongs to, the last released board known when it was written, what of 0.9.7 existed at that commit (first 0.9.7 cell: 2026-11-05 18:49Z; 0.9.7 board first committed: 2026-11-06 09:38Z; a provisional board, without the new FE datasets, was run from about 22:10Z on 2026-11-05 as a code check), and who requested it where the text says so. The headings in PREREG09 give the stated writing times.

rule	version	board known	0.9.7 at commit	first committed (UTC)	commit	requested by
Amendment 5	0.9.4	0.9.3.1	–	2026-11-02 21:35Z	`3be8b9d49`	Jonas Petersen
Amendment 6	0.9.5	0.9.4.1	–	2026-11-03 11:14Z	`799bbc68b`	Jonas Petersen
Amendment 7	0.9.5	0.9.4.1	–	2026-11-03 13:54Z	`ee6e213a0`	–
Amendment 8	0.9.5	0.9.4.1	–	2026-11-03 14:32Z	`4e04a33d8`	–
Amendment 9	0.9.5	0.9.4.1	–	2026-11-03 16:40Z	`06a56d1c1`	–
Amendment 10	0.9.5	0.9.4.1	–	2026-11-03 17:09Z	`4449cea4c`	–
Amendment 11	0.9.6	0.9.5	–	2026-11-04 21:47Z	`6024a5a7d`	Jonas Petersen
Amendment 12	0.9.6	0.9.5	–	2026-11-04 22:01Z	`20333d8f5`	Jonas Petersen
Amendment 13	0.9.7	0.9.6	no 0.9.7 cell	2026-11-05 18:32Z	`adbffba1d`	Jonas Petersen
Note 2026-11-05 18:46Z	0.9.7	0.9.6	no 0.9.7 cell	2026-11-05 18:46Z	`a008e2cae`	–
Note 2026-11-05 18:57Z	0.9.7	0.9.6	0.9.7 cells existed	2026-11-05 18:57Z	`198f252d5`	–
Note 2026-11-05 19:02Z	0.9.7	0.9.6	0.9.7 cells existed	2026-11-05 19:02Z	`24d87c8ff`	–
Note 2026-11-05 19:05Z	0.9.7	0.9.6	0.9.7 cells existed	2026-11-05 19:05Z	`007d876a8`	–
Note 2026-11-05 19:10Z	0.9.7	0.9.6	0.9.7 cells existed	2026-11-05 19:10Z	`597b2c8e1`	–
Note 2026-11-05 19:12Z	0.9.7	0.9.6	0.9.7 cells existed	2026-11-05 19:12Z	`110f9d1aa`	–
Note 2026-11-05 19:13Z	0.9.7	0.9.6	0.9.7 cells existed	2026-11-05 19:13Z	`5fc6706a9`	–
Note 2026-11-05 19:15Z	0.9.7	0.9.6	0.9.7 cells existed	2026-11-05 19:15Z	`aba266e99`	–
Note 2026-11-05 19:21Z	0.9.7	0.9.6	0.9.7 cells existed	2026-11-05 19:21Z	`0fe475fab`	–
Note 2026-11-05 19:23Z	0.9.7	0.9.6	0.9.7 cells existed	2026-11-05 19:23Z	`b5a9676b8`	–
Note 2026-11-05 19:44Z	0.9.7	0.9.6	0.9.7 cells existed	2026-11-05 19:44Z	`c9d83054c`	–
Note 2026-11-05 21:59Z	0.9.7	0.9.6	0.9.7 cells existed	2026-11-05 21:59Z	`cf7f5facb`	–
Note 2026-11-05 22:32Z	0.9.7	0.9.6	0.9.7 provisional board run (code check)	2026-11-05 22:32Z	`4755a995b`	–
Note 2026-11-06 10:02Z	0.9.7	0.9.6	0.9.7 board built	2026-11-06 10:03Z	`84f5cc9ab`	–
Note 2026-11-06 10:10Z	0.9.7	0.9.6	0.9.7 board built	2026-11-06 10:22Z	`ba8370e94`	–
Note 2026-11-06 11:05Z	0.9.7	0.9.6	0.9.7 board built	2026-11-06 11:05Z	`eb73278bb`	Jonas Petersen
Amendment 14	0.9.7.1	0.9.7	–	2026-11-05 22:35Z	`b70ca1a90`	Jonas Petersen

Baselines

Per dataset, the best baseline (rows with no learned encoder: ZERO, and current value on FE) against the best and the median ranked model. A free forecast beating the best model was part of the reason to drop epa_aqs_pm25 and substorm_omni, while kept FE datasets where a free forecast also wins stay in (see the FE table below); the 0.9.7 admission is not free of model scores: the SNS HVCM gate read model cells and six of the drops were informed by scores: C 3W, bosch and soredd_abc, FE ngafid_hastall, substorm_omni and epa_aqs_pm25 (Amendment 13 items 1 and 6; item 1 calls the 3W drop structural, but item 6 and the screening log drop it on combo minus twin; see the conflict-of-interest section).

All 49 datasets

task	dataset	best baseline	its score	best model	its score	median model	models below the baseline	flag
C	ucr/MoteStrain	ZERO	0.859	NuTime	0.949	0.927	1 of 15
C	ucr/SonyAIBORobotSurface2	ZERO	0.798	LeSaturn-warppool-factorized-7.4M ◊	0.911	0.836	3 of 15
C	ucr/Wafer	ZERO	0.761	Timestral-28M	0.998	0.990	0 of 15
C	ucr/ACSF1	ZERO	0.622	Chronos-2	0.906	0.850	0 of 15
C	ucr/PLAID	ZERO	0.598	Mantis	0.942	0.886	0 of 15
C	ucr/FreezerRegularTrain	ZERO	0.960	pj_better_3p7m ◊	0.994	0.980	1 of 15
C	ucr/FordA	ZERO	0.489	Timestral-209M	0.948	0.919	0 of 15
C	ucr/ElectricDevices	ZERO	0.628	Timestral-123M	0.805	0.788	0 of 15
C	phys/lendb	ZERO	0.574	Timestral-209M	0.939	0.898	1 of 15
C	mv06/hydraulic	ZERO	0.987	Timestral-123M	0.983	0.958	15 of 15	baseline above every model
C	mv06/gas251	ZERO	0.866	MantisV2	0.902	0.882	2 of 15
C	mv06/stead	ZERO	0.599	Timestral-123M	0.977	0.957	0 of 15
C	uea/AsphaltPavementTypeCoordinates	ZERO	0.586	pjgood-warp-xview-combo-10M ◊	0.954	0.919	0 of 15
C	mv06/aursad095	ZERO	0.653	Timestral-28M	0.726	0.692	0 of 15
C	mv06/voraus095	ZERO	0.639	Chronos-2	0.757	0.731	0 of 15
C	nicola15/chillan5	ZERO	0.565	Chronos-2	0.894	0.885	0 of 15
C	nicola15/haptic_texture8	ZERO	0.524	MantisV2	0.606	0.571	0 of 15
C	nicola15/openrainer_m	ZERO	0.805	Chronos-2	0.890	0.874	0 of 15
C	nicola15/poseit_ft	ZERO	0.595	pjgood-warp-xview-combo-10M ◊	0.634	0.614	3 of 15
C	nicola15/px4_airframe3	ZERO	0.607	Timestral-28M	0.827	0.776	0 of 15
C	nicola15/swiss_load10	ZERO	0.614	pj_better_1m ◊	0.647	0.631	3 of 15
C	nicola15/ved_powertrain3	ZERO	0.823	pjgood-warp-xview-combo-10M ◊	0.921	0.905	0 of 15
C	nicola15/bdg2_meter6	ZERO	0.580	NuTime	0.792	0.764	0 of 15
C	nicola15/robo6_rb6	ZERO	0.594	pjgood-warp-xview-combo-10M ◊	0.776	0.741	0 of 15
A	HAI	ZERO	0.660	Mantis	0.664	0.599	14 of 15	baseline at or above the median model
A	UKWIND	ZERO	0.667	Timestral-209M	0.664	0.643	15 of 15	baseline above every model
A	HOT	ZERO	0.645	NuTime	0.690	0.617	13 of 15	baseline at or above the median model
A	PREDIST	ZERO	0.701	NuTime	0.693	0.650	15 of 15	baseline above every model
A	NOBOOM	ZERO	0.789	pj_better_1m ◊	0.813	0.771	10 of 15	baseline at or above the median model
A	ESA	ZERO	0.654	pjgood-warp-xview-combo-10M ◊	0.685	0.624	10 of 15	baseline at or above the median model
A	NASA	ZERO	0.585	MantisV2	0.671	0.638	1 of 15
A	UAVSEAD	ZERO	0.616	Toto-Open-Base-1.0	0.629	0.592	14 of 15	baseline at or above the median model
A	UR5E	ZERO	0.491	Chronos-2	0.612	0.587	0 of 15
A	KUKA	ZERO	0.571	pj_better_10m ◊	0.571	0.536	14 of 15	baseline at or above the median model
FE	stickslip_poincloux	ZERO	0.617	Toto-Open-Base-1.0	0.765	0.682	0 of 15
FE	minati_circuits	current value	0.647	Toto-Open-Base-1.0	0.821	0.775	0 of 15
FE	geyser_strokkur	current value	0.623	Mantis	0.740	0.687	0 of 15
FE	ndbc_buoy_v2_b	current value	0.689	Toto-Open-Base-1.0	0.683	0.632	15 of 15	baseline above every model
FE	nasa_battery_rw	current value	0.872	Mantis	0.879	0.860	14 of 15	baseline at or above the median model
FE	wind_genbearing_hot	current value	0.703	Toto-Open-Base-1.0	0.652	0.623	15 of 15	baseline above every model
FE	towie_genbearing_hot	current value	0.739	pj_better_3p7m ◊	0.716	0.674	15 of 15	baseline above every model
FE	memristor_jerk	ZERO	0.844	Toto-Open-Base-1.0	0.933	0.900	3 of 15
FE	noaa_coops_surge_b	current value	0.795	Toto-Open-Base-1.0	0.737	0.690	15 of 15	baseline above every model
FE	asos_lows_b	current value	0.730	Toto-Open-Base-1.0	0.727	0.686	15 of 15	baseline above every model
FE	usgs_reach_v2_b	ZERO	0.580	pj_better_3p7m ◊	0.708	0.686	0 of 15
FE	rotary_utrecht	ZERO	0.612	pj_better_10m ◊	0.660	0.637	6 of 15
FE	megathrust_sync_v2	ZERO	0.542	pj_better_10m ◊	0.740	0.712	0 of 15
FE	cdip_waves_c	current value	0.738	Toto-Open-Base-1.0	0.701	0.635	15 of 15	baseline above every model
FE	heatpump_defrost	current value	0.701	pj_better_10m ◊	0.708	0.664	13 of 15	baseline at or above the median model

On FE, the best forecast with no encoder (current value, trend, curvature, early-warning statistics, spectral shift, level statistics) per dataset against the best ranked model. On 7 of 15 FE datasets a no-encoder forecast scores above every ranked model (ndbc_buoy_v2_b, wind_genbearing_hot, towie_genbearing_hot, noaa_coops_surge_b, asos_lows_b, cdip_waves_c and heatpump_defrost). For the new FE datasets the last column is pjgood-warp-xview-combo-10M ◊ minus the random-init LeSaturn row (3.7 M; not a matched twin), the second number Amendment 13 item 5 asks the card to show.

All 15 FE datasets

FE dataset	best no-encoder forecast	its score	best ranked model	its score	ranked rows below it	flag	combo-10M minus random-init LeSaturn (new datasets)
stickslip_poincloux	early-warning statistics	0.702	Toto-Open-Base-1.0	0.765	10 of 15
minati_circuits	curvature	0.761	Toto-Open-Base-1.0	0.821	6 of 15
geyser_strokkur	early-warning statistics	0.731	Mantis	0.740	12 of 15
ndbc_buoy_v2_b	current value	0.689	Toto-Open-Base-1.0	0.683	15 of 15	free forecast above every ranked model
nasa_battery_rw	current value	0.872	Mantis	0.879	14 of 15
wind_genbearing_hot	current value	0.703	Toto-Open-Base-1.0	0.652	15 of 15	free forecast above every ranked model
towie_genbearing_hot	level statistics	0.754	pj_better_3p7m ◊	0.716	15 of 15	free forecast above every ranked model
memristor_jerk	curvature	0.883	Toto-Open-Base-1.0	0.933	6 of 15
noaa_coops_surge_b	current value	0.795	Toto-Open-Base-1.0	0.737	15 of 15	free forecast above every ranked model
asos_lows_b	trend	0.744	Toto-Open-Base-1.0	0.727	15 of 15	free forecast above every ranked model
usgs_reach_v2_b	early-warning statistics	0.653	pj_better_3p7m ◊	0.708	6 of 15
rotary_utrecht	level statistics	0.647	pj_better_10m ◊	0.660	9 of 15
megathrust_sync_v2	level statistics	0.674	pj_better_10m ◊	0.740	6 of 15
cdip_waves_c	current value	0.738	Toto-Open-Base-1.0	0.701	15 of 15	free forecast above every ranked model
heatpump_defrost ●	curvature	0.729	pj_better_10m ◊	0.708	15 of 15	free forecast above every ranked model	+0.027

Per task, the three strongest unranked rows (controls, no-encoder references and the specialist row) and how many ranked rows score below them on the task score; a row with no ranked row below it is not listed. The per-dataset baseline above is ZERO or current value only; on FE several no-encoder forecasts score higher than current value.

task	row	kind	task score	ranked rows below it
C	MultiROCKET-Hydra	specialist	0.864	15 of 15
A	ZERO	control	0.638	11 of 15
A	Random-init LeSaturn (control)	control	0.600	2 of 15
A	LevelOnly	admission	0.594	2 of 15
FE	level statistics	admission	0.719	9 of 15
FE	curvature	admission	0.707	7 of 15
FE	Random-init LeSaturn (control)	control	0.703	7 of 15

On A, the best of the three data-free rows (no data read: a score that rises or falls with elapsed time, and a random score per window) against the ranked models. The elapsed-time gate of Amendment 13 item 4 (at most 0.65) admits a dataset; it does not promise that every model beats these rows.

A dataset	best data-free row	its score	best model	its score	median model	models below it	flag
HAI	null: elapsed time rising	0.533	Mantis	0.664	0.599	0 of 15
UKWIND	null: elapsed time rising	0.558	Timestral-209M	0.664	0.643	0 of 15
HOT	null: elapsed time rising	0.529	NuTime	0.690	0.617	0 of 15
PREDIST	null: elapsed time rising	0.512	NuTime	0.693	0.650	0 of 15
NOBOOM	null: elapsed time rising	0.561	pj_better_1m ◊	0.813	0.771	0 of 15
ESA	null: elapsed time rising	0.611	pjgood-warp-xview-combo-10M ◊	0.685	0.624	1 of 15
NASA	null: elapsed time rising	0.516	MantisV2	0.671	0.638	0 of 15
UAVSEAD	null: elapsed time rising	0.585	Toto-Open-Base-1.0	0.629	0.592	6 of 15
UR5E	null: elapsed time rising	0.616	Chronos-2	0.612	0.587	15 of 15	data-free row above every ranked model
KUKA	null: elapsed time falling	0.519	pj_better_10m ◊	0.571	0.536	2 of 15

Flagged: UR5E (null: elapsed time rising 0.616 against the best model Chronos-2 0.612). On such a dataset the ranking orders models that do no better than knowing the position in the series.

Per model, the datasets where it scores more than 0.01 (the practical-equality margin) below the best baseline:

model	datasets	which
pj_better_10m ◊	12	C:hydraulic, A:HAI, A:UKWIND, A:HOT, A:PREDIST, A:UAVSEAD, FE:ndbc_buoy_v2_b, FE:wind_genbearing_hot, FE:towie_genbearing_hot, FE:noaa_coops_surge_b, FE:asos_lows_b, FE:cdip_waves_c
pj_better_3p7m ◊	11	C:hydraulic, A:HAI, A:HOT, A:PREDIST, A:UAVSEAD, FE:ndbc_buoy_v2_b, FE:wind_genbearing_hot, FE:towie_genbearing_hot, FE:noaa_coops_surge_b, FE:asos_lows_b, FE:cdip_waves_c
pjgood-warp-xview-combo-10M ◊	12	C:hydraulic, A:HAI, A:UKWIND, A:HOT, A:PREDIST, A:UAVSEAD, FE:ndbc_buoy_v2_b, FE:wind_genbearing_hot, FE:towie_genbearing_hot, FE:noaa_coops_surge_b, FE:asos_lows_b, FE:cdip_waves_c
Mantis	11	C:hydraulic, A:UKWIND, A:PREDIST, A:ESA, A:UAVSEAD, A:KUKA, FE:ndbc_buoy_v2_b, FE:wind_genbearing_hot, FE:towie_genbearing_hot, FE:noaa_coops_surge_b, FE:cdip_waves_c
Chronos-2	16	C:hydraulic, A:HAI, A:UKWIND, A:HOT, A:PREDIST, A:NOBOOM, A:ESA, A:UAVSEAD, A:KUKA, FE:ndbc_buoy_v2_b, FE:nasa_battery_rw, FE:wind_genbearing_hot, FE:towie_genbearing_hot, FE:noaa_coops_surge_b, FE:cdip_waves_c, FE:heatpump_defrost
LeSaturn-warppool-factorized-7.4M ◊	15	C:hydraulic, A:HAI, A:UKWIND, A:HOT, A:PREDIST, A:UAVSEAD, A:KUKA, FE:ndbc_buoy_v2_b, FE:nasa_battery_rw, FE:wind_genbearing_hot, FE:towie_genbearing_hot, FE:noaa_coops_surge_b, FE:asos_lows_b, FE:cdip_waves_c, FE:heatpump_defrost
pj_better_1m ◊	13	C:SonyAIBORobotSurface2, C:hydraulic, A:HAI, A:UKWIND, A:HOT, A:PREDIST, FE:ndbc_buoy_v2_b, FE:wind_genbearing_hot, FE:towie_genbearing_hot, FE:noaa_coops_surge_b, FE:asos_lows_b, FE:cdip_waves_c, FE:heatpump_defrost
MantisV2	15	C:hydraulic, A:HAI, A:UKWIND, A:PREDIST, A:NOBOOM, A:ESA, A:UAVSEAD, A:KUKA, FE:ndbc_buoy_v2_b, FE:wind_genbearing_hot, FE:towie_genbearing_hot, FE:noaa_coops_surge_b, FE:asos_lows_b, FE:cdip_waves_c, FE:heatpump_defrost
NuTime	15	C:hydraulic, A:HAI, A:UKWIND, A:NOBOOM, A:ESA, A:UAVSEAD, A:KUKA, FE:ndbc_buoy_v2_b, FE:wind_genbearing_hot, FE:towie_genbearing_hot, FE:noaa_coops_surge_b, FE:asos_lows_b, FE:rotary_utrecht, FE:cdip_waves_c, FE:heatpump_defrost
Timestral-209M	17	C:hydraulic, A:HAI, A:HOT, A:PREDIST, A:NOBOOM, A:ESA, A:UAVSEAD, A:KUKA, FE:ndbc_buoy_v2_b, FE:nasa_battery_rw, FE:wind_genbearing_hot, FE:towie_genbearing_hot, FE:noaa_coops_surge_b, FE:asos_lows_b, FE:rotary_utrecht, FE:cdip_waves_c, FE:heatpump_defrost
Timestral-123M	17	A:HAI, A:UKWIND, A:HOT, A:PREDIST, A:NOBOOM, A:ESA, A:UAVSEAD, A:KUKA, FE:ndbc_buoy_v2_b, FE:nasa_battery_rw, FE:wind_genbearing_hot, FE:towie_genbearing_hot, FE:noaa_coops_surge_b, FE:asos_lows_b, FE:rotary_utrecht, FE:cdip_waves_c, FE:heatpump_defrost
Toto-Open-Base-1.0	19	C:SonyAIBORobotSurface2, C:FreezerRegularTrain, C:lendb, C:hydraulic, C:gas251, C:poseit_ft, A:HAI, A:UKWIND, A:HOT, A:PREDIST, A:NOBOOM, A:ESA, A:KUKA, FE:nasa_battery_rw, FE:wind_genbearing_hot, FE:towie_genbearing_hot, FE:noaa_coops_surge_b, FE:cdip_waves_c, FE:heatpump_defrost
Timestral-28M	19	C:hydraulic, A:HAI, A:UKWIND, A:HOT, A:PREDIST, A:NOBOOM, A:ESA, A:UAVSEAD, A:KUKA, FE:ndbc_buoy_v2_b, FE:nasa_battery_rw, FE:wind_genbearing_hot, FE:towie_genbearing_hot, FE:memristor_jerk, FE:noaa_coops_surge_b, FE:asos_lows_b, FE:rotary_utrecht, FE:cdip_waves_c, FE:heatpump_defrost
Timestral-5M	17	A:HAI, A:HOT, A:PREDIST, A:NOBOOM, A:ESA, A:UAVSEAD, A:KUKA, FE:ndbc_buoy_v2_b, FE:nasa_battery_rw, FE:wind_genbearing_hot, FE:towie_genbearing_hot, FE:memristor_jerk, FE:noaa_coops_surge_b, FE:asos_lows_b, FE:rotary_utrecht, FE:cdip_waves_c, FE:heatpump_defrost
TimesFM-2.5-200M	23	C:MoteStrain, C:hydraulic, C:gas251, C:swiss_load10, A:HAI, A:UKWIND, A:HOT, A:PREDIST, A:NOBOOM, A:ESA, A:NASA, A:UAVSEAD, A:KUKA, FE:ndbc_buoy_v2_b, FE:nasa_battery_rw, FE:wind_genbearing_hot, FE:towie_genbearing_hot, FE:memristor_jerk, FE:noaa_coops_surge_b, FE:asos_lows_b, FE:rotary_utrecht, FE:cdip_waves_c, FE:heatpump_defrost

Noise budget

Signal against test-material noise, over the ranked rows: the SD of their scores (signal) and the bootstrap SD of the paired difference of two rows (test material, median over pairs). A single-seed pair closer than 0.009 is marked within seed noise; that margin comes from the 0.9.2 noise budget, which also counted pretraining seed, head training set and readout choice and was not rerun.

board	signal SD	test material, SD of a pair	signal over test material
C	0.0184	0.0018	10.15
A	0.0196	0.0056	3.54
FE	0.0438	0.0030	14.39
CAFE	0.0212	0.0022	9.55

The datasets themselves are a further source: with them resampled, 61 of 105 pairs stay resolved, against 85 on the main board (10 of them within seed noise).

Elo

Elo from paired per-dataset wins of the ranked rows (ZERO = 1000; 2,000 draws; elo095.py's method). It is a figure and gives no verdict: the CAFE score, its Holm verdicts and tiers are the board. Inner interval: test material resampled; outer: datasets resampled too. Guests are rated against the frozen ratings of the ranked rows; a guest without a cell on a dataset takes ZERO's value there, and the share is in the last column. Third-party pool: the Elo of a fit without the ◊ rows in the pool.

‡: some of the row's dataset cells are missing and were filled with ZERO's value for the Elo matches (column imputed). The tabular rows are among them because their cells on the new and regrown datasets are pending (v0.9.7-r1).

row	kind	Elo	test material	and datasets	Elo C	Elo A	Elo FE	third-party pool	imputed
pj_better_10m ◊	ranked, one readout	1271	1247 to 1293	1222 to 1319	1506	1045	1483	1262
pj_better_3p7m ◊	ranked, one readout	1233	1213 to 1251	1187 to 1279	1478	1022	1408	1232
pjgood-warp-xview-combo-10M ◊	ranked, one readout	1233	1213 to 1253	1190 to 1278	1544	1025	1315	1240
Mantis	ranked, one readout	1198	1178 to 1219	1147 to 1252	1416	926	1498	1195
Chronos-2	ranked, one readout	1177	1159 to 1197	1128 to 1225	1540	830	1374	1189
pj_better_1m ◊	ranked, one readout	1150	1133 to 1169	1098 to 1199	1377	964	1304	1165
LeSaturn-warppool-factorized-7.4M ◊	ranked, one readout	1149	1131 to 1167	1101 to 1196	1395	864	1418	1156
MantisV2	ranked, one readout	1135	1114 to 1156	1076 to 1198	1455	909	1207	1141
NuTime	ranked, one readout	1054	1035 to 1076	994 to 1110	1396	876	989	1072
Timestral-209M	ranked, one readout	1037	1019 to 1053	992 to 1084	1513	787	856	1042
Timestral-123M	ranked, one readout	1024	1008 to 1041	977 to 1071	1493	777	847	1031
Toto-Open-Base-1.0	ranked, one readout	1014	998 to 1032	953 to 1074	1143	651	1453	1000
Timestral-28M	ranked, one readout	1001	984 to 1020	955 to 1050	1457	775	796	1001
ZERO	no encoder	1000	974 to 1022	916 to 1069	–	–	–	–
Timestral-5M	ranked, one readout	975	957 to 990	919 to 1029	1367	750	879	984
Random-init LeSaturn (control)	control	948	926 to 971	880 to 1019	–	–	–	–
Raw reference	no encoder	919	897 to 940	807 to 1015	–	–	–	–
TimesFM-2.5-200M	ranked, one readout	814	789 to 841	744 to 879	1242	595	631	814
CatBoost ‡	tabular, defaults	1139	1117 to 1164	1052 to 1220	–	–	–	–	35%
LightGBM ‡	tabular, defaults	1114	1087 to 1144	1029 to 1193	–	–	–	–	18%
XGBoost ‡	tabular, defaults	1114	1091 to 1143	1033 to 1188	–	–	–	–	20%
RealMLP ‡	tabular, defaults	1025	999 to 1056	939 to 1102	–	–	–	–	49%

Open dataset questions

Questions about datasets that this version leaves open; none changes a score.

asos_lows_b: its event onsets are not independent of those of asos_lows_c (a review on CAFE #16 found most within a day); whether to pool them is a review question carried over from earlier releases (carryover097.md); no decision is recorded.
Hill of Towie (HOT): its label rule (stopping alarms of at least 6 hours, weather codes removed) is a review question carried over from earlier releases (carryover097.md); no decision is recorded.
cdip_waves_c: a review question carried over from earlier releases (carryover097.md); no decision is recorded.
SMAP and MSL (pooled into NASA): their near-duplicate clusters are a review question carried over from earlier releases (carryover097.md); no decision is recorded.
KUKA: threshold-easy. Its admission record says: Disclosed: a single threshold of ~1.75 Nm on the external-torque channels finds the collisions with F1 0.92.
HAI: HAIEnd 23.05 left out (same two 23.05 recordings and labels through other tags); whether to add it is open.
CAFUC2: out by check 2 (synthetic faults written into recorded flights); only Jonas Petersen can reverse it.

Heat maps

One row per model, one column per dataset. Lighter is lower and darker is higher, so the scale also reads in grey and for colour-blind readers. A thick edge marks the best ranked model on a dataset. The rows follow the switches above the board.

Hover or tap a cell for its value.

Breakdowns

Per dataset

Best and median are over ranked models. A flag says a baseline beats every model on that dataset, or reaches at least the median model. Those datasets stay in on purpose.

Dataset groups

Domain groups are for browsing and carry no weight in any score. Numbers are the main board's. To see the board without a group, use the switches at the top.

Pair verdicts

Row against column. Hover or tap a cell for the difference, its interval and the Holm p. Unresolved does not mean equal.

Hover or tap a cell.

Tiers

CAFE score

Tier 1 pj_better_10m ◊

Tier 2 pj_better_3p7m ◊, pjgood-warp-xview-combo-10M ◊, Mantis, Chronos-2

Tier 3 LeSaturn-warppool-factorized-7.4M ◊, pj_better_1m ◊

Tier 4 MantisV2

Tier 5 NuTime, Timestral-209M

Tier 6 Timestral-123M, Toto-Open-Base-1.0, Timestral-28M, Timestral-5M

Tier 7 TimesFM-2.5-200M

C, classification

Tier 1 Chronos-2, pjgood-warp-xview-combo-10M ◊, Timestral-209M

Tier 2 Timestral-123M, pj_better_10m ◊, pj_better_3p7m ◊

Tier 3 Timestral-28M

Tier 4 Mantis, LeSaturn-warppool-factorized-7.4M ◊, NuTime, Timestral-5M, MantisV2

Tier 5 pj_better_1m ◊

Tier 6 TimesFM-2.5-200M

Tier 7 Toto-Open-Base-1.0

A, anomaly detection

Tier 1 pj_better_10m ◊, pj_better_3p7m ◊, pjgood-warp-xview-combo-10M ◊, pj_better_1m ◊, Mantis, MantisV2, NuTime

Tier 2 LeSaturn-warppool-factorized-7.4M ◊, Chronos-2, Timestral-209M

Tier 3 Timestral-28M, Timestral-123M, Timestral-5M

Tier 4 TimesFM-2.5-200M, Toto-Open-Base-1.0

FE, forecasting events

Tier 1 Mantis, pj_better_10m ◊, Toto-Open-Base-1.0, LeSaturn-warppool-factorized-7.4M ◊

Tier 2 pj_better_3p7m ◊, Chronos-2

Tier 3 pjgood-warp-xview-combo-10M ◊, pj_better_1m ◊

Tier 4 MantisV2

Tier 5 NuTime

Tier 6 Timestral-5M, Timestral-209M, Timestral-123M

Tier 7 Timestral-28M

Tier 8 TimesFM-2.5-200M

Sensitivities

With and without A

Computed from the per-dataset values on this page. Ranks count the ranked rows only.

Other rule sets and dataset choices

Each row is a whole board under one changed rule, from sens097.json.

Datasets resampled

The datasets are drawn again with replacement within each task. Computed by sens097.json (sens095.py), 10,000 draws, each also resampling the test material. This asks how much the order depends on which datasets are in the benchmark, which the board's own interval does not cover.

Noise

Spread over pair noise, this version (the 1.0 test's A signal-to-noise is in the 1.0 test section)

Recomputed from this version's draws: the spread of the ranked rows' scores divided by the test-material noise of a pair. The other sources below are not in it.

A null floors per dataset

Each null is normalised as the dataset's score is, so 0.5 is chance. They are shown as unranked rows; the elapsed-time nulls also gate new A datasets (Amendment 13 item 4). In 0.9.7 an A dataset is admitted only with at least 15 Eval series, at least 20 anomaly segments and a median anomaly share of at most 0.3, and a new one also needs the better elapsed-time null at most 0.65 (Amendment 13 items 2 and 4; the card's Admission section lists every candidate). These scores read no model.

10 A datasets are admitted; the rule asks for at least 5. The mean of their best data-free scores is 0.554, and 15 of 15 ranked rows score above it on those datasets; the rule asks for at least half. So A counts in the CAFE score.

Source: the data-free rows of board097.json (centre assignment, Eval series).

Rules in plain words

Inputs

Every encoder gets real samples with no padding and no mask. What an encoder does inside, such as padding to its own context or normalising its input, counts as part of the model under test. The models table lists the input rules each row used.

On C, a recording of 512 samples or fewer is stretched to 512 by linear interpolation for every model, which changes the time scale in the same way for every encoder. A longer recording goes to the model whole: a model that accepts variable length gets it up to its maximum context, most recent samples kept, and a model with a fixed input length of 512 gets it block-averaged to 512. On A, the window is 256 real samples ending at the scored moment, never stretched; its score goes to the window's centre step, and a series needs a prefix of at least 512 samples or it leaves the list. On FE, every model gets the last 512 samples before the forecast time. Rows that embed no window, the tabular rows and hand-made references, keep their registered inputs.

Heads

The head is the only thing fitted, and it is the same for every model. C uses L2 logistic regression on one block per channel: a model's native width when it is at most 256 per channel, and wider embeddings projected to 256 per channel by a fixed random projection (at most 8,192 in total), with its penalty chosen by cross-validation inside the training split. A uses no labels: it scores a window by the mean of its 5 smallest Mahalanobis distances to the windows of the series' own prefix. FE fits one ridge regression per lead time on the embedding at the end of the context, with 3 head seeds.

Scores

Each dataset gives one number on the same scale, 0.5 for chance and 1 for perfect. C uses balanced accuracy rescaled for the number of classes, A uses VUS-PR after the prefix rescaled against random scores, and FE uses stratified AUROC averaged over lead times and head seeds. A task score is the plain mean over its datasets. The CAFE score is the mean of C, A and FE.

Readouts

Each model has one readout, the same on C, A and FE. The candidates are m0 (the released embedding), m1, m2, m3 and m6, as far as the model's code exposes them; the pick has the highest validation CAFE score, the mean over the tasks of cross-validation on the training split for C, the Tuning series for A and the inner cross-validation of the ridge head for FE. A candidate within 0.005 of the best loses to an earlier one. No test value is read. The picks were made in 0.9.5 and not rerun; the A part of the validation score used the Tuning series of the 0.9.5 A datasets, which were the Forgis rows' development set (Amendment 9 item 2).

Intervals

An interval is a paired 95% percentile bootstrap of the test material inside each dataset: test samples or splits on C, Eval series on A with near-duplicates resampled together, time blocks on FE. There are 10,000 draws, the same for every model. The datasets are held fixed, so an interval says nothing about how a model would do on other datasets; the resampled table under Sensitivities speaks to that.

Verdicts and tiers

Every pair of ranked models gets one of three verdicts. It is better if the Holm-corrected two-sided bootstrap p is below 0.05. It is practically equal if not resolved and the 95% interval of the difference lies inside plus or minus 0.01. Otherwise it is unresolved, which means the data do not say. A pair of single-seed models less than 0.009 apart is also marked within seed noise. Tier 1 holds the models above the first one resolved below the top model, and the next tier starts there, so a model in a lower tier can still be unresolved against the top one. Rank inside a tier is not a claim.

Admission checks

A dataset must come from an inanimate physical system, its label must be a state, type or identity of that system or an event observed in the record, it needs at least 100 test samples on C or 1,000 events on FE, and only the largest dataset of a recording campaign stays. In 0.9.7 an A dataset needs at least 15 Eval series, at least 20 anomaly segments and a median anomaly share of at most 0.3, with at most 300 series per source of independent runs and 100 per source with several series per asset; a new A dataset also needs the better of the rising and falling elapsed-time nulls at most 0.65, a data-free check that removed RflyMAD, OnboardEU and TCABR; a new FE dataset needs at least 300 events, 30 independent units or days and data-free reference rows below fixed limits (Amendment 13 items 2, 4 and 5). A series needs the prefix above. Not every check is free of model scores: the C candidate SNS HVCM had to pass a combo-10M minus random-init LeSaturn gate on its own cells, and six of the 0.9.7 drops (C 3W, bosch, soredd_abc; FE ngafid_hastall, substorm_omni, epa_aqs_pm25) were informed by scores (Amendment 13 items 1 and 6). A free forecast beating the best model was part of the reason to drop epa_aqs_pm25 and substorm_omni, while kept FE datasets where a free forecast also wins stay in; the boards show where that happens.

Conflict of interest

Forgis maintains CAFE and developed 5 of the 15 ranked models, all with CAFE Eval scores visible. Jonas Petersen of Forgis requested Amendment 13 with the 0.9.6 board known. Forgis staff chose the drops, adds, caps and thresholds of 0.9.7 with that board, its separability table and a per-dataset comparison of the Forgis rows known; some drops were informed by scores (Amendment 13 item 1), the two C additions were accepted after a screening that scored pjgood-warp-xview-combo-10M ◊ against its random-init twin on them, and the A thresholds were placed in gaps of the 0.9.6 data. Applied to the 0.9.6 scores, the drops alone raise the mean ◊ lead (see What changed in 0.9.7). The pair count of the test that decides between 0.9.7 and 0.9.5 as the CAFE 1.0 candidate was fixed before any 0.9.7 cell existed; its reading of the A signal-to-noise was fixed at 19:02Z on 2026-11-05, after the first 0.9.7 A cells existed and before any was read. The boards under older rules stay published at their tags.

What the board does not show

The board does not show performance after fine-tuning, because nothing is fine-tuned. It does not cover datasets outside the list, and its intervals leave out pretraining seed, head training set and readout choice. It says nothing about value forecasting, because FE scores events only. Run time is reported and never scored. CAFE publishes scores and no data.

Known limitations, as written on the dataset card

Each model keeps the readout chosen for it in 0.9.5 (Amendment 13 item 7), chosen without the 0.9.7 datasets.
The drops of 0.9.7 were chosen with the 0.9.6 board known; some of them were informed by scores (Amendment 13 item 1), and the A thresholds sit in gaps seen in the 0.9.6 data, so they are post-hoc cut-offs.
The SNS HVCM gate cells were computed after the A cells of the first admission waves, not first as Amendment 13 item 6 said (PREREG09 note of 2026-11-05 19:05Z); no score had been read.
UR5e is under a non-commercial licence (CC BY-NC 4.0).
On HAI (7 recordings), UKWIND (1 recording), HOT (9 recordings), ESA (2 recordings) and KUKA (33 recordings), some A Tuning series are other blocks of the same recordings as Eval series, though no series is in both. The readouts were chosen in 0.9.5 on the Tuning series of CARE_A, CARE_C, ESA_M1, ESA_M2, HAI, HOT, MSL, SKAB, SMAP, UKWIND and W3, so this matters only where those datasets overlap the 0.9.7 ones; for datasets new in 0.9.7 (KUKA among them) no readout was chosen on their Tuning series. The recording-clustered sensitivity of the test is under the 1.0 test.
On 7 of 15 FE datasets (ndbc_buoy_v2_b, wind_genbearing_hot, towie_genbearing_hot, noaa_coops_surge_b, asos_lows_b, cdip_waves_c and heatpump_defrost) a forecast with no encoder scores above every ranked model; FE is also where the ◊ lead is largest. A free forecast beating the best model was part of the reason to drop epa_aqs_pm25 and substorm_omni, and the same holds on these kept or added datasets (see Baselines).
On UR5E a data-free elapsed-time row (0.616) scores above every ranked model (best 0.612); see Baselines. Each UR5e series puts the last rows of a separate normal run, recorded earlier, before the anomalous run (build_ur5e.py), so prefix and scored part come from different runs.
PreDist: the 7 days appended after each fault are the one build step that reads an event time; on PreDist and NoBOOM a score that follows elapsed time does well, and the data-free rows show by how much.
Recording length carries label information in two C datasets, in the source data itself: a classifier that sees only the length scores uea/AsphaltPavementTypeCoordinates 0.720 and ucr/PLAID 0.596 (see Datasets admitted and removed); every other C dataset has one length per recording. The variable-length models see the length of recordings above 512 samples and the block-averaged models do not.
FactoryNet's README warns that recording context partly predicts fault labels, so real fault evaluation should stay within one recording session. mv06/voraus and the other pooled C datasets mix sessions, and a voraus split tests settings not seen in training.
voraus-AD comes from FactoryNet's copy, not the publisher's files, under CC BY-NC-SA 4.0. Forgis is a company, so CAFE publishes scores only, and a legal reading is still open.
The weights of the Forgis rows are not public. Each carries a sha256 under Models, so a later release can be checked against it, and until then outsiders cannot rerun those rows.
At least 66 distinct Saturn-family checkpoints, including the five ranked Forgis rows and two HEPA-lineage precursors, have CAFE or PrecursorBench 1.4 Eval scores on datasets of the 0.9.3.1 board, with records dated from 2026-08-04 to 2026-10-28. Before its submission, each ranked row had at least this many other variants scored: LeSaturn-warppool-factorized-7.4M 43, pjgood-warp-xview-combo-10M 59, pj_better_3p7m and pj_better_10m 64, and pj_better_1m 65. These are lower bounds counted from repository records, because some run trees and lane outputs are unavailable, and they exclude readout variants, controls, evaluation seeds, PrecursorBench v1 runs and development-harness runs.
MantisV2 is read through the sweep's adapter, and its m0 is the sweep's released-readout rule, not independently confirmed.
Hill of Towie labels are inferred stops, with rules made without any score. cdip_waves_c was rebuilt because series.npy differs from the pinned file; every ranked row was recomputed on the rebuilt file.
The near-duplicate rule clusters SMAP series with long constant stretches although their correlation is near zero, and four UKWIND turbines have identical series that still need checking at the source.
Mantis and NuTime were pretrained on data that includes UCR and UEA training splits. Whether a Forgis checkpoint was pretrained on FactoryNet data was not established.

How to submit a model

CAFE has no upload form. To propose a model, open a discussion on the dataset page with a public checkpoint pinned to a commit, its licence, code that turns an array of shape (samples, channels, T) into an embedding, with the largest T it accepts, the list of readouts it exposes, and the number of variants you scored on CAFE datasets before submitting, which the protocol publishes with the row. Forgis then runs its cells under the rules above, and the row appears with its marks in a dated version.

Datasets and models

Datasets and licences

CAFE redistributes no data.

Models

Every model is frozen.

Version history

Protocol · Changelog · leaderboard.json

Figures and tables for reports

Every figure is a file in this Space. Use the links under each one. Tables are below in a form you can copy, and each also exists as a CSV file. Version 0.9.7.

The main board with intervals, one panel, drawn for this page.SVG

Heat map of C on the main board, drawn for this page. Colour runs from 0.5 or lower to the highest score on the task.SVG

Heat map of A on the main board, drawn for this page. Colour runs from 0.5 or lower to the highest score on the task.SVG

Heat map of FE on the main board, drawn for this page. Colour runs from 0.5 or lower to the highest score on the task.SVG

Not offered here because they predate this board: baselines.png, cost.png, elo_cost.png, gaps.png, leaderboard.png, method_board.png, money_plot.pdf, money_plot.png, variance_breakdown.png, variance_pie.png.

Tables

Glossary

A: Anomaly detection: find the abnormal stretches in a series. Scored by VUS-PR after the prefix.
anomaly segments: Labelled runs of consecutive anomalous steps in the Eval series of an A dataset.
AUROC: The chance that a randomly picked positive case gets a higher score than a randomly picked negative one. FE uses a stratified version, compared within fold, unit and month.
balanced accuracy: Accuracy averaged over classes, so a large class does not dominate. C rescales it so that chance is 0.5.
baseline: A row with no learned encoder: ZERO on every task, and current value on FE. Shown on every board, never ranked. The random-init control and the other references are listed beside them.
best baseline: On one dataset, the highest-scoring of the baselines the board compares with.
board effect: Pairs resolved on the CAFE score without a dataset minus with it; negative means the dataset helps.
bootstrap: Drawing the test material again with replacement to see how far a score moves. CAFE uses 10,000 draws.
C: Classification: say what a recording shows. Scored by balanced accuracy.
CAFE score: The mean of the task scores that count on this release: C, A and FE.
cap: The most series a source may give an A dataset: 300 for independent runs, 100 for several series per asset (Amendment 13 item 2).
cell: One model's score on one dataset, stored with its bootstrap draws.
chance-normalised: Rescaled so that random output gives 0.5 and a perfect result gives 1.
check 2: The admission check that the label is a state of, or an event in, the physical system.
combo minus twin: C score of pjgood-warp-xview-combo-10M minus that of the random-init LeSaturn row (3.7 M parameters, another architecture; not a matched twin of the 10.6 M combo), per repeat: a model-specific check that pretraining helps on a dataset.
conflict of interest: Forgis runs CAFE and also builds some of the models it ranks. Those rows carry the mark ◊.
datasets resampled: The board recomputed 10,000 times with the datasets of each task drawn with replacement as well as the test material; the measure of the 1.0 test.
drops only: The 0.9.6 board with the datasets removed in 0.9.7 taken out and nothing added.
elapsed-time best: The better of the two data-free scores that rise or fall with time since the prefix.
elapsed-time gate: A new A dataset is admitted only if the better of the rising and falling elapsed-time scores is at most 0.65.
embedding: The vector of numbers a frozen model returns for a window. It is all a head ever sees.
Eval series: The series of an A dataset that are scored. Tuning series are kept apart.
F3 thresholds: The highest time-since-last-event and time-of-day scores on the kept FE datasets; a new FE dataset may not exceed them.
FE: Forecasting events: say whether an event will happen within the next lead time. Scored by AUROC.
foundation model: A network pretrained on many time series and then used for jobs it did not see.
free forecast: An FE reference row that reads the series without a model, such as current value or trend.
frozen: The weights are not changed for CAFE. No fine-tuning on any dataset.
GPU-hours, upper bound: Sum of task wall times over 3600; tasks share GPUs, so the GPU time used is at most this.
H0, H1, H2: The C head candidates of Amendment 12: the 0.9.5 head, the same with an 11-value penalty grid, and that after a 256-component projection fitted on the training split.
head: The small model fitted on the embedding for one task. It is the same for every model.
Holm: A step-down correction that keeps the error rate over a whole family of pair tests at the stated level.
input rules: What each encoder gets: on C the whole recording (block-averaged to 512 for fixed-length models; stretched to 512 when it is 512 samples or shorter), on A a window of 256 real samples, on FE the last 512 samples. No padding, no mask.
Kendall tau: Agreement between two orderings: 1 means the same order, 0 means no relation.
lead time: How far ahead an FE head has to warn.
Mahalanobis distance: Distance that takes account of the spread and correlation of the normal windows.
median share: Median over Eval series of the labelled share of the scored part; at most 0.3 to be admitted.
near-duplicate cluster: Eval series joined by a rule on the data alone and resampled as one block, so that copies do not narrow the interval.
node-hours: Elapsed time of a cluster job times its nodes, from sacct.
non-commercial: A licence that allows use for non-commercial purposes only, such as CC BY-NC 4.0 (UR5e).
not admitted: A candidate measured without a model and left out by a rule, with the reason recorded before any cell.
null: A score that reads no model: elapsed time rising, elapsed time falling, or a random number per window.
null floor: A dataset's highest null. A model at or below it has shown nothing the clock or chance could not.
One readout: The one readout a model uses on C, A and FE, picked by the highest validation CAFE score without reading any test value. This is the main board's rule.
pair noise: Median over all pairs of ranked models of the bootstrap SD of their paired score difference.
paired: Every model is scored on the same random draws, so a difference between two models does not depend on which material a draw picked.
pooled: Datasets of one release or paper counted as one dataset, its score the mean over all their Eval series (Amendment 13 item 2).
practically equal: A verdict: not resolved, and the 95% interval of the difference lies inside plus or minus 0.01.
prefix: The normal stretch at the start of an A series, used to fit the head.
readout: Which internal tensor of a model is taken as its embedding.
reference: A hand-made row that embeds nothing, such as max-|z| or current value. Shown, never ranked.
resolved alone: Pairs of ranked models resolved on a board made of one dataset alone.
signal-to-noise (separability): On one dataset or task, the mean absolute score gap between two ranked models divided by the median bootstrap SD of that gap.
spread SD: SD of the ranked models' scores on one dataset or task: how far apart the models are.
tabular row: A gradient-boosted or MLP learner on hand-written statistics of the window. Shown, never ranked.
tau with the task: Kendall tau between the models' order on one dataset and their mean over the task's other datasets.
tier: Models in score order. Tier 1 holds the models above the first one resolved below the top model; the next tier starts there and is built the same way, so a model in a lower tier can still be unresolved against the top one.
Tuning series: Series of an A dataset kept out of scoring. Only readout selection reads them.
unresolved: A verdict: the data do not separate the two models, and the interval is too wide to call them equal.
VUS-PR: Volume under the precision-recall surface, the metric for A.
windows per second: Scored A windows of a model's A cells divided by their task wall time.
within seed noise: A pair of single-seed models closer than 0.009 on the CAFE score or on a task board, 1.96 times the total SD of a single-seed difference in the 0.9.2 noise budget (0.0044, of which pretraining seed is 0.0018), a normal approximation.
ZERO: A fixed random projection of the raw window with no learning. The no-encoder reference.
†: A mean over fewer datasets than the view has, because some cells do not exist in this readout mode.
◊: A model developed by Forgis with access to CAFE Eval scores.
◊ lead: Mean CAFE score of the ◊ rows minus that of the other ranked rows; reported, part of no rule.

Main board

Is 0.9.7 more separable than 0.9.6?

What changed in 0.9.7

Comparison across versions since 0.9.4

0.9.6 and 0.9.7 side by side

Separability by task, dataset and model

Every dataset's effect on the board

Admission of 0.9.7

What took how long

Inference and benchmarking time

Cost per model

Jobs and cells that did not finish

Change log

v0.9.7 (2026-11-06)

v0.9.6

v0.9.5

v0.9.4

Provenance of the rules

Baselines

Noise budget

Elo

Open dataset questions

Heat maps

Breakdowns

Per dataset

Dataset groups

Pair verdicts

Tiers

CAFE score

C, classification

A, anomaly detection

FE, forecasting events

Sensitivities

Noise

A null floors per dataset

Rules in plain words

Inputs

Heads

Scores

Readouts

Intervals

Verdicts and tiers

Admission checks

Conflict of interest

What the board does not show

How to submit a model

Datasets and models

Version history

Figures and tables for reports

Tables

Glossary

Scores with 95% intervals

Datasets: spread and uncertainty