Lab notebook
The record of being wrong
The programme's observations, instrument checks, corrections and research drafts. Each entry states its kind; drafts remain proposals for review. Earlier claims and their corrections remain visible.
Research continuation — September and October 2026
These materials separate a new instrument readout, an audit of existing data, and proposals for later work.…
research overview · source
Bias-corrected divergences: a follow-up plan, committed before its code
2026-10-22 · prospective plan for an offline simulation, not an experiment. No model is called and no repository data is read.…
design draft · source
F* near a vanishing gap: the simulation's results
2026-10-16 · results of the offline simulation registered in the plan, which was committed before the simulation code, itself committed before its registered run.…
methods note · source
The vanishing denominator: a simulation plan, committed before the code
2026-10-16 · prospective plan for an offline simulation, not an experiment. No model is called and no repository data is read.…
design draft · source
Estimand audit: Problem 13's incidents, traced to their arrays
2026-10-07 · offline instrument and report, for human review. No model was called, no published result was recomputed, and no rule is adopted here.…
audit draft · source
Draft sensitivity audit: modal ties in RIVERSIDE-30 self-transfer
Reviewed 2026-10-07. This note's finding that the gpt-oss-composed row mean is 0 under another admissible tie choice is ground (c) of retraction 22.…
audit draft · source
Program-output verifier: offline instrument check
2026-09-28 · instrument prototype, not an E-003 experiment. This is the first implementation step proposed by the domain methods note.…
instrument · source
Qwen register feasibility: instrument readout
2026-09-14 (UTC start date). Collection complete. The complete two-judge corpus result is available.…
instrument · source
Draft audit: RIVERSIDE-30 self-transfer arithmetic and claim scope
Reviewed 2026-10-07. The repository owner adopted this audit's suggested wording, and the claim it reviews is withdrawn as retraction 22.…
audit draft · source
E-003 domain selection: a protocol draft
2026-09-09 · design proposal, not a pre-registration or a result. No new probe text, key, model response or experiment is produced by this document.…
design draft · source
Domain instruments: methods-read supplement
2026-09-14 · literature and engineering note, not a registration or result. This supplements the September 9 domain-selection draft with methods inspected beyond abstracts.…
methods note · source
2026-07-28 — Founding
First entry. The lab notebook is for the things that do not belong in uncertainty.…
journal · source
2026-07-28 — First live run: void, and the reason is worth keeping
E-001 ran against the API for the first time and died. Full record in AMENDMENT-001.md; this entry is for what it means.…
journal · source
E-001 — Findings
Status: the run was interrupted at the CEILING condition and never completed.…
findings · source
E-001b is void — the cost parity gate failed before the sweep ran
Status: void · closed · superseded by E-001c (not yet registered) Voided: 2026-07-29, after composition, before any probe outcome was examined Gate: cost parity across cells, max/min realised cost ≤ 1.30 (PREREGISTRATION §4.1)…
void · source
DEFECT-001 — the analysis path was never executed, and would have crashed
Found: 2026-07-29, ~4 hours into collection Status: repaired; collection restarted Data on disk when found: sender 34/34, PRIOR 34/34, CEILING 34/34,…
defect · source
Can the four registers survive being pinned to a common length?
Instrument check, before registration. No hypothesis is tested here and none may be read off it.…
instrument · source
E-002 is void — the elicitation saturated, and so did the transfer
Status: void · closed · superseded by E-002b (not yet registered) Voided: 2026-07-30, at the first brief's elicitation, by the pre-registered degeneracy gate Cost: ~4.5 hours of the ~13 the full sweep would have taken…
void · source
E-002b — findings
The first run in this repository to reach analysis. Every prior experiment voided. All gates passed; all four rungs of the ladder produced outcome variation.…
findings · source
E-002c — findings
The first pre-registered positive result this programme has produced. All six gates passed, all twenty-four briefs analysed, all four registered hypotheses resolved.…
findings · source
2026-07-30 — Two models disagree exactly where one is wrong
Status: an observation, not a finding. Post-hoc, on data collected for another purpose, with a statistic chosen after seeing it.…
observation · source
2026-07-30 — Two audit holes, found by an agent sent to argue about definitions
Both were found while an adversarial panel was checking arithmetic for an unrelated decision. Both are verified here by independent recomputation.…
audit · source
The public page was still publishing v0.1
2026-07-31. A four-lens consistency audit over the repository returned twenty findings, all claim-bearing, all verified against source. They deduplicate to nine places.…
audit · source
E-004 is void — two of three models were never wrong
Status: void · closed · no successor registered Voided: 2026-08-03, by gate G2, before any hypothesis was read Cost: four days of local GPU, ~$1.50 of API, and the answer to a question this design could not have given…
void · source
Collection stopped at 104 of 126 cells, on one socket timeout
2026-07-31, 19:04 local. Written before the run is resumed and before any analysis, so that what happened is on the record independently of what the remaining data turns out to say.…
instrument · source
DEFECT-001 — the band was calibrated on one instruction, and the run composed under another
Found: 2026-08-04, after the run of 2026-08-03 voided Status: repaired; the void stands as a void, but not for the reason it records Data on disk when found: one results file,…
defect · source
CALIBRATION-001 — the model writes to the ceiling it is told, and slightly past it
Measured 2026-08-04, after the run of 2026-08-03 voided and DEFECT-001 established that the void could not mean what it said.…
instrument · source
E-001c is void — the fluent register has gpt-oss:120b has a floor where the band has a ceiling
Status: void · closed · successor not yet registered Voided: 2026-08-04, during composition, before any probe outcome existed Gate: register filter exhausted in cell A (PREREGISTRATION §4, §6.1)…
void · source
The measure was nine templates and every p-value counted thirty-two
2026-08-18. MERIDIAN-IX32 has thirty-two probes. It does not have thirty-two independent readings, and three published significance levels were computed as though it did.…
audit · source
Seven of the nine probes measured the reader, not the transfer
2026-08-19. MERIDIAN-IX32 has been read by a second architecture, and the route it was built to open is closed.…
observation · source
DEFECT-001 — the noise floor was computed on the wrong pair
Found 2026-08-20, from files already committed, at no measurement cost. Affects the per-brief and per-rung fidelity column only.…
defect · source
E-006 is void before collection
2026-08-20. Zero model calls. Void on its own pre-registered power gate, which was computed before any draw and refused the design.…
void · source
Three laws were blocked by the wrong instrument
2026-08-28. Nine days that began by trying to test L1 and ended by finding that the reason three laws have stood unattacked since founding was never the ideas, and never — as I published mid-way — the programme's instruments in general.…
observation · source
The measurement was already in the repository
2026-09-08. Eleven days that produced one first measurement, one correction worth more than the claim it corrected, one model chosen for a reason that turned out to be false — and, at the end, an audit that read every number in the arc against the file it came from and found three of them wrong.…
audit · source
The ledger, read against the text
2026-10-07. Nothing was measured today and no model was called. The work was reading RETRACTIONS.md against the files it governs, and it moved the count from twenty-one to twenty-four.…
audit · source
Retractions
Every claim this programme has made and then withdrawn, with what killed it.…
audit · source
Prior art
Created 2026-07-29, after an external review found that several of this programme's founding novelty claims were false.…
audit · source