Findings

Each finding below is a standalone measured result. Every number was measured against the file it describes, not estimated. Each carries its DOI. The unflattering results are included: finding 1 is a negative result, and finding 2 reports the system's own error rate rather than only its catches.

Finding 1. Verification does not monotonically improve.

Across 276 checks that were re-run, there were 63 verdict transitions: 31 hardening events (a check was made harder to game and the verdict moved CONFIRMED to REFUTED) and 32 loosening events. The net is −1.

Correction. This finding was published as net zero — 38 hardening, 38 loosening, over 79 transitions. That figure was wrong and is retracted here. The checker grouped ledger rows by the raw check field without guarding rows that have no check name, so 25 unrelated claims were differenced against one another as though they were re-runs of a single instrument. That manufactured 7 hardening events, 6 loosening events and all 3 transitions the checker could not classify. Corrected, the net is −1 — a slightly worse result than the one published. The full correction is at /corrections.

Source: /thesis · DOI 10.5281/zenodo.22752082

Finding 2. The instrument that caught 5,837 false claims had itself been fooled 31 times.

Hardening events are cases where a verdict flipped from CONFIRMED to REFUTED only after the check was strengthened. The system had been ratifying a claim it would later refute. Both directions were published; reporting only the hardening half would have made the instrument look better than it is.

Those events are not independent. 74% of all recorded movement — 56 of 76 events before the correction above — comes from just 12 checks that flip in both directions. The largest single contributor is one claim with one pre-registered threshold, run 35 times, producing 10 transitions: five hardening, five loosening. A near-zero net is therefore as much a property of a handful of oscillating checks as it is of the instrument as a whole, and this finding should be read with that in mind.

Source: /thesis · DOI 10.5281/zenodo.22752082

Finding 3. 99.9% of recorded claims carry the measurement that would have killed them, set before the outcome was known.

16,194 of 16,218 records carry an explicit pre-registered refutation threshold. This removes the largest degree of freedom available to a self-grading system: choosing, after seeing the result, which measurement counted.

Source: /ledger · DOI 10.5281/zenodo.22752075

Finding 4. 5,837 claims were refuted against the system's own ground truth, and none were deleted.

The falsification ledger is append-only and hash-chained at 29,707 entries with the chain verified unbroken. A system that discards its failures has no immune system: it re-proposes the same broken claim next cycle at full confidence.

Source: /ledger · DOI 10.5281/zenodo.22752075

Finding 5. 28,226 divergences were measured in systems that passed every one of their declared checks.

Differential testing across 301 campaigns found 28,226 spec-compliant divergences, 9,830 crashes and 5,404 genuinely compliant runs. The checks passed while real state detached from what the specification meant.

Scope, stated because it limits the finding. The targets are owned synthetic in-process simulators, and their entire model — every threshold and every cap — is written as named constants inside the same module, deliberately, so that specification mining can recover it. The oracle is therefore a hand-written specification by the same author as the system under test, not an independent second implementation. What this measures is how far an implementation drifted from a spec I wrote, inside a system I also wrote. It is not evidence about divergence in anyone else's system, and it should not be cited as though it were.

Source: /isomorphism · DOI 10.5281/zenodo.22752082

Finding 6. The binding constraint on self-improvement is verifier precision, not generator capability.

Written as r* > p_gen: the loop improves only while the verifier's precision exceeds the generator's power to produce convincing falsehoods. A stronger generator makes verification harder, not easier.

Source: /thesis · DOI 10.5281/zenodo.22752082

Finding 7. 135 dated defensive publications, each with its own DOI.

Published 24 June to 13 September 2026. Collection DOI 10.5281/zenodo.22751107; individual records 10.5281/zenodo.22751478 through 10.5281/zenodo.22752083.

Source: /record-of-priority · DOI 10.5281/zenodo.22751107

Finding 8. A 360M model fine-tuned on 69,945 rows beat its base model by 14.1%.

gnarp-m2 was fine-tuned via QLoRA on a 74,395-row corpus, of which 69,945 rows were train-eligible: the trainer discards every row labelled negative, and 4,450 rows carry that label. So the corpus does not include the claims the system refuted against itself — those rows were excluded from the loss outright. Earlier text on this page claimed the opposite and was wrong. The corpus also carries no failure_class field at all, so the fine-grained error grammar is absent from the training signal rather than collapsed within it. The model card's refusal-rate row for m2 is empty because it was never measured.

Source: https://huggingface.co/CatQualia/gnarp-m2 · DOI 10.5281/zenodo.22751569