Version 2.0 (2026-10): the trial report Before this version the record has three published versions: 1.9 (2026-08-17), the pre-registration of a randomized trial of memory dosing in a production agent fleet; 1.11 (2026-08-17), a correction of it (1.10 was written and not deposited); and 1.12 (2026-08-26), an amendment. Version 2.0 is the report of that trial: the manuscript A registered horizon that outlived the intervention the trial ran: a pre-registered randomized trial of memory dosing in a production agent fleet (registered-horizon-outlived-the-intervention-the-trial-ran-v2.0.md, and the same text as PDF), with the artifacts its numbers come from. This version is also the grouped amendment of the registration in this Zenodo record: it declares the deviations from the deposited registration and reports the result in one record. Appendix A of the manuscript lists the twelve deviations a reader cannot reconstruct from the estimates; the append-only log DEVIATIONS-FOR-PAPER.md, in the package, holds the others. The OSF registration yf7d2 (2026-08-18) is not amended and has not changed since it was registered. The registration files stay where they were deposited, OSF yf7d2 and Zenodo v1.12, 10.5281/zenodo.22110203 (2026-08-26), and are not copied into this version. The surface measurements of the same system are reported separately (Paper A, 10.5281/zenodo.23163119). Result. Under the registered analysis no hypothesis rejects. The primary outcome reported, H1c (the share of opportunities that yield a repeated failure), is null: the registered re-randomization test gives p = 0.4006, and the estimate is −0.0140 (95% BCa interval −0.0464 to +0.0110, 19 epochs, 11 treatment and 8 control). H1, the primary named by v1.12, gives p = 0.3294 tested alone, and every Holm-adjusted p-value in the H1a–H1c and H2 family is 1.0. In the sensitivity analysis kept from earlier drafts, which splits sessions across the epochs they touch, H1 rejects under the deposited reading (p = 0.0127); the registered analysis counts each session in the epoch of its start, as the registration requires, and does not reject. No interval of the H1 family excludes zero. Two H2 intervals (cost) exclude zero in the direction of a costlier treatment, under the BCa construction only: the percentile intervals contain zero, and on time the raw difference has the opposite sign (−1.18 s). The registered H2 test does not reject. The null does not show that the mechanism has no effect. Under the planning assumptions, no reduction from the planning baseline, including total elimination, reaches 80% power at 19 epochs; that calculation uses an intra-cluster correlation estimated for a different quantity and is not a calibrated measure of this study's uncertainty (§4.1.1). Why the trial was short. The registration set a horizon of 234 daily epochs. The 19 items designated for dosing all left the candidate pool together at 2026-09-20 22:51:23Z, during the 20th epoch; past that instant no dose reaches them, and treatment and control are the same intervention. The designation was chosen on 2026-08-26, after v1.12 was deposited, and was not deposited. The trial was closed before its registered horizon. Of the 20 epochs realized, one (2026-09-02, lost to an outage) served no brief, and the other 19 form the analysis set. Six of the deviations reported in the manuscript (§1, Appendix A; the full list is there): Primary outcome: v1.12 names H1 (repeated-failure density per session-hour); the paper reports H1c as primary, a switch decided on 2026-08-30, before the assignment seed was drawn, in documents that were not deposited. H1 is reported beside it under both readings. Horizon: 234 epochs registered, 20 realized, because of the designation above. Stopping rule: a rule written before the seed (close the arm with a null if Epoch-1 coverage fell below 36.7%) was found only after the trial; no contemporaneous record shows it was evaluated, and its condition was met under one reading and not under another. Assignment rule: the seed declaration named a rule other than the registered assign_arms.py and retracted it in writing 24 min 38 s after the drand round, before Epoch 1. The arms served came from assign_arms.py; run offline, the two rules differ in 11 of the 20 realized epochs. Adjudication: the registration locked a census of the live study; the analysis adjudicates the error stratum in full and a hash-ordered sample of 800 of the other 5,556 episodes (weight 6.945). An attempt to restore the census after unblinding stopped at a test-retest gate declared before it started. Analysis specification: written on 2026-09-10, before any outcome was computed, and not deposited. The analysis reported combines it with the deposited provisions and with a panel-substitution rule declared after the trial; the analysis of earlier drafts is kept as a sensitivity analysis. Sham replays, an instrument check. The dosing code was replayed over the served briefs with the real designation and with 20 sham designations matched to it. On the 2,646 brief states of the w = 4 epochs, where production served that dose and the replay reproduces it, the real designation changed more states than every sham: 132 against 81–122, p = 1/21, the floor for 20 shams. Against these 20 shams, that is evidence that the dose acted specifically on what those epochs served. A second replay applied w = 4 to all 11,812 reconstructible states of the hash-verified trial window, control epochs included, so outside the w = 4 epochs it is a counterfactual, not a record of what was served; there too the real designation exceeded every sham (790 against 449–583, p = 1/21). The two replays share 2,016 states (the w = 4 epochs except 2026-09-01), so the second is not a replication. Neither says anything about outcomes, and the rank is not a calibrated randomization p-value. Files. artifacts-v2.0.zip holds the registered-analysis outputs (out/ITT-REGISTRADO-*.json, v1 to v4), the results notes and frozen copies of the estimator (B-registered/), the adjudication census record without its raw episode files (B-censo/), the sham replays (summaries, determinism checks and the two runs archives), the figures, the review receipts and verdicts, the per-round records of how each review finding was checked and applied, and the trial's own outputs kept outside the repository (_ballast/, with the hash manifest that covers them). scripts-v2.0.zip holds the measurement and check scripts and the PDF build. _derived/panel-verdicts/, also in artifacts-v2.0.zip, holds the panel's verdict on each adjudicated episode, from the two verdict files the manuscript cites: episode id, panelist, model requested and model served, verdict and severity level, call status. The panelists' free-text reasons are removed, because they describe the episodes. MANIFEST-v2.0.json lists the SHA-256 of every packaged file, which copies were redacted (local paths and host names only), every path the manuscript cites and where it is found, and what was left out and why. Episode text, the panelists' free-text reasons and the production databases are not included.
No takes yet. Share an insight, caveat, or question.
Luiz Antonio Busnello (2026) studied this question.