Skip to content

LEDGER.md

Latest commit

 

History

History
90 lines (70 loc) · 4.16 KB

LEDGER.md

File metadata and controls

90 lines (70 loc) · 4.16 KB
Raw

THE LEDGER — every cycle this lab has run, and how many of them we lost

Generated by papers/build_ledger.py from committed receipts. Nothing here is typed by hand; re-run it and check.

A research record is only worth what its negatives are worth. This file exists so that the ratio is visible without reading 158 cycle entries.

The count
cycles logged 163
preregistrations frozen 350 (38 carrying a machine-scored gates block)
OATH certificates 173
trust-stack seals 34
cycles ending in a refusal, null, retraction or INVALID 62 of 163
verdicts that were literally INVALID__* — the machinery refusing its own run 16
gates declaring a power basis 7 of 38 gated preregs
— of those frozen since the field existed 6 of 6

Both power-basis rows are reported because either alone misleads. A frozen preregistration is never edited, so the preregs written before power_basis existed can never acquire one and their denominator can never move; quoting only the first row understates the discipline, and quoting only the second hides how much of the corpus predates it.

Certificate verdicts: OATH-HELD × 173

Seal verdicts: SEALED × 33, REFUSED × 1

Every run the machinery refused

These are the cycles where a preregistered gate returned INVALID__* — the experiment was run, the verdict was computed, and the frozen table said the result licenses nothing. Each one cost real compute and produced no claim.

  • cycle 67 (2026-07-24) — INVALID__underpowered
  • cycle 85 (2026-07-28) — INVALID__underpowered
  • cycle 101 (2026-07-30) — INVALID__probe_cells_underpowered__CG1_reasoning_does_not_immunize_the_report
  • cycle 110 (2026-08-01) — TWO
  • cycle 115 (2026-08-03) — TWO
  • cycle 117 (2026-08-04) — INVALID__dose_response_nonmonotone
  • cycle 118 (2026-08-04) — INVALID__whitening_breaks_the_clique
  • cycle 132 (2026-08-06) — INVALID__null_leaks
  • cycle 133 (2026-08-06) — SHIPPED
  • cycle 134 (2026-08-06) — PRODUCT
  • cycle 142 (2026-08-06) — INSTRUMENT_BLIND_TO_ISC__refusals_are_miscalibrated
  • cycle 149 (2026-08-07) — INVALID__breaks_existing_preregs
  • cycle 152 (2026-08-07) — DO
  • cycle 154 (2026-08-07) — REWRITTEN
  • cycle 156 (2026-08-07) — BUILT
  • cycle 157 (2026-08-08) — NO_LEGIBILITY_ISLANDS__the_first_island_does_not_generalize

The rule, measured

On 2026-08-06 this lab adopted a standing rule: no instrument is announced before an adversarial pass. Three modules had shipped before it existed. The record since:

outcome
shipped before the rule 3 modules, 3 broken within hours by red teams, 2 recalled the same day
shipped after the rule released only once an adversary tried and failed
stopped by the rule 2 modules quarantined pre-release, 0 users exposed

The two quarantined modules are retained in the tree with their audit verdicts in their own docstrings, because deleting a failed instrument hides the evidence.

How to check any of this

python -m styxx.corpus_audit papers/      # re-certify every claim against its receipts
python papers/build_ledger.py             # regenerate this file
pytest -q                                 # the suite, including a whole-corpus re-score

The corpus audit re-runs the current verifier over every number this lab has ever published. The test suite re-scores every committed result against its own frozen preregistration and fails on a single changed verdict string.

Why this file is not OATH-certified

Every other claim document here carries an OATH certificate binding its numbers to receipt files. This one cannot: its numbers are counts of the receipts themselves, so there is no receipt beneath them to bind to. It gets a stronger guarantee instead — regeneration. A test in the suite rebuilds this file from the committed corpus and fails if a single character differs, so it cannot drift from the record it describes and cannot be edited by hand without the build going red.

A lab that only publishes what worked is reporting a filtered sample of its own beliefs.