CAFE

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 ((only Jonas Petersen can reverse this, in a dated note); 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.

Amendment 15 (CAFE 0.9.7.2, written 2026-11-06 ~23:50Z (first headed 2026-11-07 ~00:10Z, a clock error; commit 199d87582 at 23:51:45Z), before any 0.9.7.2 build, measurement or cell on a dataset not yet measured, and before any model cell on a 0.9.7.1 or 0.9.7.2 candidate; the released 0.9.7 board, card and tags are not changed by it). Requested by Jonas Petersen on 2026-11-06 at ~23:45Z: a 0.9.7.2 with more audits, more robustness checks and more separability on A, adding further A datasets that pass every rule (the TSB-AD collection named as a source), aiming at an A signal-to-noise above 9; published with the earlier versions kept on the leaderboard, and checked so that the 1.0 conversion can run on his go. Disclosure: the released 0.9.7 board and its per-dataset A separability (signal-to-noise per dataset from 1.0 on UKWIND to 6.1 on KUKA; the task value 4.04 is limited mainly by the noise of the datasets with few Eval series) were known. So were the 0.9.7.1 measurements of 2026-11-05 22:37Z to 22:49Z (UTAQ, RCFPOWER, PVLAZ, NOBOOM_M202210; no model was run on them). Conflict of interest: as in Amendment 13 item 1. 1. Versions. 0.9.7.1 was never built past the measurements or published. Amendment 14's candidates are carried into 0.9.7.2, and the name 0.9.7.1 is not used for a release. The tabular reference update planned as card revision v0.9.7-r1 is also folded into 0.9.7.2. 2. Rules. Amendment 13 items 2 and 4 and Amendment 14 items 1 and 2 are unchanged and apply to every candidate, measured with measure097.py before any cell. Checks 1 to 4 of Amendment 8 apply as written (physical system, physical label, test material, one dataset per campaign). No existing A dataset is dropped or changed, except by a pooling that Amendment 14 already names (NOBOOM_M202210 into NOBOOM; the later UKWIND years into UKWIND). 3. TSB-AD. Each source collection in TSB-AD-U and TSB-AD-M is judged on its own against item 2, from the publisher's original files where TSB-AD has cut, resampled or normalized them. Sources already decided keep their decision (SMAP and MSL are in NASA; SWaT needs an access form; SKAB was removed by Amendment 13). Sources whose anomalies were injected into recorded data by software, simulated plants, IT or web metrics, and recordings of people or animals fail check 1 or check 2 and are listed with that reason. 4. The 0.9.7.2 board is the 0.9.7 board plus the admitted A datasets and the tabular reference rows. C and FE do not change. Item 9 of Amendment 13 applies unchanged: at least 56 of the 105 pairs resolved with the datasets resampled, and an A signal-to-noise of at least 6 (goal 9). 0.9.7 is shown beside 0.9.7.2. If 0.9.7.2 passes, it is the 1.0 candidate; the conversion still runs only on Jonas's explicit go. 5. Order of work and cut-off. As Amendment 13 item 8: build and measure, then a dated admission note, then cells. A candidate whose units are not on Clariden by 2026-11-07 03:30Z is not in 0.9.7.2. Its builder and measurement are handed over in a ranked list. A candidate on which any ranked row has no cell after two attempts leaves 0.9.7.2 as a whole, with the reason on the card. No admission, pooling or cut-off decision reads a model score on a candidate. 6. Exit. At least three fresh audits, the last two in a row with no blocker, before the upload (Amendment 5 rule 13 for 1.0 is unchanged).

Note 2026-11-06 23:54Z (first headed 2026-11-07 00:35Z, a clock error; commit dcf9de80a at 23:54:39Z) (0.9.7.2 admission, wave 1; Amendment 15 item 5; before any model cell on the dataset). In: UTAQ, the Utah aquatic sensor networks (Logan River Observatory with the iUTAH GAMUT Red Butte and Provo networks; HydroShare, CC BY 4.0, open). Measured on 2026-11-05 at 22:49Z with measure097.py, no model run: 85 Eval and 15 Tuning series (site-year P2 blocks, cap 100), 404 anomaly segments in Eval, median scored anomaly share 0.027, all 85 Eval series with a prefix of at least 512, elapsed-time best 0.560 (centre; gate 0.65). Rules 2(a), 2(b), 2(c) and the elapsed-time gate hold. Amendment 14's condition holds as built: the label is the union of the technicians' qualifier codes for sensor malfunction, power failure, maintenance, sediment, low water and calibration, fixed in the builder before any series was built (~/stage097/logs/build_utaq_labelrule_frozen_2240Z.py, 22:40Z), and the inputs are the raw logger files, not the quality-controlled files. Checks 1 to 4 of Amendment 8: in-stream sondes are a physical system; the qualifier codes mark instrument and site events entered by technicians, not thresholds; Eval and Tuning are disjoint site-year blocks; the three networks are one dataset (same authors, item 2 pooling rule). Not a flight set. The A gate threshold is recomputed over the final 0.9.7.2 A list in the note that closes the list, before any 0.9.7.2 A cell is read.

Note 2026-11-07 00:02Z (first headed 00:15Z, a clock error; commit a1f71c5f6) (Amendment 15 item 3 result; two further candidates named before any download or build of them). TSB-AD: no source collection passes (work/r097/tsbad_screen0972.md). GHL, CATSv2 and Exathlon are simulated or carry injected events; SMD, PSM, WSD, IOPS, NEK, YAHOO, Stock and CreditCard are IT, web or finance metrics; MITDB, SVDB, LTDB, Daphnet and OPPORTUNITY record people; GECCO's events were imputed by the organisers; TAO has no documented label and no series with a 512-sample prefix; Power labels holidays; Genesis needs a login; NAB has two physical non-human series; the UCR families are synthetic or hand-edited anomalies with no 15 independent series from one campaign. SMAP and MSL stay in NASA; SWaT needs an access form. Nothing was built. Named as candidates under the unchanged rules of item 2 (each out unless it passes every rule and checks 1 to 4 when built): LEAD1.0 (building electricity meters with human-annotated anomalies; admitted only if its labels are human annotations of recorded events, the licence is usable, and it is not the same campaign as the C unit nicola15/bdg2_meter6, which is drawn from BDG2; if it is, it is out under check 4); USGS NWIS gauges with equipment or ice qualifier codes (admitted only if the qualified stretches keep the recorded signal in the published record). No score of any kind exists on either.

Note 2026-11-07 00:10Z (committed 00:07:31Z as 4683f1fec; the heading time was written ahead of the clock) (the two candidates of the 00:02Z note; no download of model data, no build, no score). Out: LEAD1.0, check 4: its meters are the BDG2 electricity meters of the ASHRAE GEPIII subset (LEAD paper, arXiv 2203.17256; BDG2 README), the campaign the C unit nicola15/bdg2_meter6 is built from; it would also fail the licence rule (no licence stated; the full data sit behind a Kaggle competition login). Its labels would have passed (people annotated every point with a web tool). USGS NWIS: on ice-affected stretches the published discharge is an estimate written in place of the record and the recorded gage height carries no qualifier, so the label marks an estimate or a gap, not a recorded event (check 2); no equipment-coded stretch was found in the probes.

Note 2026-11-07 01:45Z (committed 01:42:32Z as bced91633; the heading time was written ahead of the clock) (0.9.7.2 admission, wave 2, the final 0.9.7.2 A list and the A gate threshold; Amendment 15 item 5; before any model cell on the wave-2 datasets and before any 0.9.7.2 A cell is read). Measured with measure097.py, no model run (builders in data097/src/, record data097/admit0972.json from admit0972.py). In: RCFPOWER (Zenodo, CC BY 4.0): 84 Eval series, 103 segments, median share 0.010, elapsed best 0.513, all with a prefix of 512. Amendment 14's condition holds: only the physically induced categories are labelled (load, voltage, frequency, power quality, outage; sensor, context and compound categories excluded), 136 such events (109 without the gradual-drift and flicker types, whose induction protocol the datasheet does not state). UKWIND pooled with the later years of the same publisher (Amendment 14): Kelmarsh and Penmanshiel 2016 to 2022, 85 Eval series under the cap of 100 in hash order, 128 segments, median share 0.027, elapsed best 0.541; the 2016 rebuild reproduces all 21 stored series byte for byte. Kelmarsh 2023 and 2024 are left out because each turbine file holds 81 overlapping exports (2.17 million rows for 52,560 timestamps), which the unchanged 0.8 builder cannot read without a new de-duplication rule; this was decided from the file structure before any UKWIND measurement. Of the 17 Eval series of 0.9.7, 12 stay in Eval and 5 fall outside the cap. The inherited flag of 0.8 stays: availability and lost-production channels computed by the SCADA are inputs. REASSEMBLE (TU Wien, CC BY 4.0): 18 Eval series, 44 segments, median share 0.115, elapsed best 0.556. Its condition holds: the files hold only measured robot state (no command or haptic-master stream) and the failed actions are physical outcomes (196 of 1,957 action segments). Disclosed: the fetch was stopped at 00:57Z because the VM was overloaded, so REASSEMBLE is built from the recordings at zip positions 0 to 78 (one failed transfer), a label-blind subset; 74 recordings after the README known-issue drops give 18 Eval series. Fetching the rest (about 35 GB) would roughly double them; it is handed over. Out: PVLAZ (rule 2(a), 5 Eval series; elapsed-time gate 0.683). NOBOOM_M202210: 85 of its 89 runs are NOBOOM runs with the same inputs and a second label column (21 of its kept series are byte-identical to NOBOOM series, and the pooled list would put identical inputs in Eval twice and in Eval and Tuning), so it adds no new test material (checks 3 and 4); NOBOOM stays as in 0.9.7. Final 0.9.7.2 A list (13 datasets, 992 Eval series): ESA 95, NASA 37, UKWIND 85, HOT 50, PREDIST 41, NOBOOM 20, HAI 58, UAV-SEAD 101, UR5e 63, KUKA 255, UTAQ 85, RCFPOWER 84, REASSEMBLE 18. One flight set (UAV-SEAD) of 13. A gate threshold (Amendment 5 rule 5(b), recomputed over the final list): 0.5502 = mean over the 13 datasets of the highest data-free null (centre): ESA 0.6109, NASA 0.5162, UKWIND 0.5408, HOT 0.5293, PREDIST 0.5124, NOBOOM 0.5605, HAI 0.5328, UAV-SEAD 0.5851, UR5e 0.6156, KUKA 0.5194, UTAQ 0.5604, RCFPOWER 0.5130, REASSEMBLE 0.5560. No 0.9.7.2 A cell has been read (UTAQ's cells exist on Clariden and were only checked for completeness).

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.

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).