Twenty-two partition_run jobs on corpus-v3/wikimedia, the first 22 admissible partitions at
training seeds 1 to 22 and N = 1,850 (jobs 210577 to 210600, pinned at ec8cc05). Every
outcome-neutral check passed and the apparatus held in all 22; the largest leakage rate, own or
sibling, is 0.40% against the registered 2%. Every run scored the same 711 examples (429 changes);
none was dropped.
The ten runs planned above, repeated at PLANT=0.5 (jobs 210234 to 210243, pinned at b05a2d8),
rq1-partition-openstack-*-n1850-plant0.5.json. All ten pass. Every half's lower bound is
above zero (least +0.017); per-half estimates run from +0.053 to +0.142. The pass probability
projected by the logged rule (normal per half, interval half-width over 1.96 as the standard
error, product over halves) is at least 0.9989 for every run and 0.9999 on average. Both conditions
of the rule hold, so check 5 is registered with the plant at a half: neutral.planted_convention,
which reads the plant's strength and window from the run's corpus path and refuses any other, and
partition_pilot.py --planted, which halts the cell when it fails.
Wikimedia was refined under the v3 rules: 5,699 examples kept over 13 months, 1,472 removed as
examples whose target the reviewer wrote (suggested_edit).
A sweep of federated fine-tuning for software engineering, organization-scale adaptation and federated LoRA privacy, each record checked on arXiv. What it changes:
Audit item 7. Current guidance on justifying a smallest effect of interest (Anvari and Lakens, JESP 2021; Lakens, Collabra 2022) prefers a practical or cost-benefit argument, anchor or consensus methods, and advises against taking the value from effects earlier studies reported. The seed-noise origin recorded on 2026-09-22 is withdrawn as a justification.
The ten runs planned above (jobs 210077 to 210086, pinned at b05a2d8), rq1-partition-openstack-*-n1850-plant0.25.json.
Two pass (partitions 2 and 12), seven read mixed, one fail. The planted effect is there in
every run but one half (partition 5, half b, -0.013), from about +0.003 to +0.106, and the
development window's 81 to 125 changes a half resolve the smaller ones no better than the pilot's
183 did. Registered as drafted, check 5 would halt a working apparatus in about four studies of
five. By the rule fixed before these runs, it is not registered at a quarter: the plant is raised
to a half and the same ten partitions and seeds are repeated.
Audit item 5. The Stage 1 draft registers outcome-neutral check 5: on the development window, one run with each half's own comment marker planted in a quarter of its refinements must pass on both halves. Its one pilot (v1, bf16, random halves of 183 changes each, one seed) passed on the weaker half at a lower bound of +0.005, and on a partition the halves hold 125 and 81 development changes. A positive control that halts a working apparatus is not "sufficient" in the sense the registered-report guidelines use, so its pass rate is measured before it is registered.
label-audit-v2-decision.json (Nimble, Ollama 0.35.0, digest 24e550a16a70; raters/C beside the
sheet). The same command that reproduces label-audit-v2.json with rater C added. Nimble's
answers are deterministic: six items re-asked returned the same choice and confidence.
Exploratory, outside the registered pipeline. Ollama 0.35 serves decision models over a System
One API (/v1/systemone): a choice question returns a probability per option and a yes/no
question the probability of yes, in one pass, with no generated text. Nimble (Bespoke Labs, 9B,
Qwen3.5-9B base, Apache 2.0) runs locally, so no corpus text leaves the machine; TypeSafe's hosted
Jev, the API it follows, is not used, since its data terms are unpublished. The only published
evaluation outside its vendors found a System One model's calibration model-specific, to be
audited per model (Rafe and Das, arXiv:2609.24052).
Audit item 8. 0.928 was 0.80 joint over three independent cells, a target of our own making that no longer describes the admitted set. Replaced by the registered-report standard: "For frequentist analysis plans, the a priori power must be 0.95 or higher for all proposed hypothesis tests" (Nature Communications / Nature Human Behaviour Registered Reports author guidelines, read from the guidelines PDF). The hypothesis tests are H1 and H2, not their cells. H1 passes only when every admitted organization's cell is supported (intersection-union), and the organizations are independent, so H1's power is the product of its cells' powers: with k cells each needs 0.95^(1/k), 0.9747 at two and 0.9830 at three. The admitted set is fixed at Stage 1 submission, so the simulation finds the bound for each of one to three cells, and the design reads the one matching the set it registers.
interval-calibration-change-averaged.json, interval_calibration.py --estimator change_averaged,
the same null, window, arm, trials and resamples as the pooled interval-calibration.json. The
claim came from the audit of this log (item 3), which found no artifact behind it.
project-windows-openstack-v3.json, from window-report-openstack-v3.json (both new; the
projection now writes an artifact and reads the corpus it is given). v3's train window holds 1,645
changes over ten months, captured at 0.952, so 172.9 true changes a month; that rate predicts the
dev window at 211 against an actual 214 (-1.3%, inside the 15% tolerance). Over the registered
twelve months, fetched at the earliest month the horizon permits (2027-01, captured 0.973), the
test window projects to 2,018 changes, replacing the ten-month 1,809 scaled to v3 (1,635).
The simulation now reads its test size from this artifact (--projection), not a typed --size.
A reader with no part in writing these entries went through the log oldest to newest, re-deriving figures from the committed artifacts, and checked the current docs and the Stage 1 draft against it. Two findings were re-checked by hand before anything was acted on (the leakage halt and the test size, below). What stands, by consequence:
partition-sensitivity-openstack.json (commit 59f48bf; every input read from its artifact and
recorded under inputs). 1,000 null studies and 300 per power step at each point, the 1,635
planned changes, K = 24, N = 1,850. At each point on the pilot's per-run spread, sigma_run is the
shift at which 24 null runs on the pilot's 206 changes spread that much (100 studies a step):
An independent review of measure/repeated-partitions at 59f48bf reproduced both pilot artifacts,
the sizing (sd 0.014020, bound 0.017657, formula 23.95, K = 24), the admissible list and the
chi-squared bounds against scipy, and found: