Every flagship claim in this repo ships with receipts that re-run on an 8GB consumer GPU. This file converts that from a statement into a ledger. The first independent re-run of each target earns a named credit line in the next Zenodo version of the corresponding paper and a permanent row here. Zero replications after six months is also an answer, and we will publish that too.
A moat nobody has tried to cross is indistinguishable from a field. The claims below are only as strong as the number of people who have independently re-run them. We are asking you to try to break them — a replication that fails to match is MORE valuable than one that matches, and earns the same credit.
| target | claim | frozen code | canonical receipt | replication bar |
|---|---|---|---|---|
| B2-adaptive | the chasing eraser never converges; private-13 read survives (SURVIVES__vs_adaptive_erasure_1p5B) |
papers/calib-poison-general/b2_adaptive_erasure.py |
papers/calib-poison-general/b2_adaptive_erasure_result.json |
verdict string matches; per-cell private13_auroc within tolerance |
| B2 static | subspace verifiably zeroed, signal relocates (SURVIVES__vs_subspace_erasure_1p5B) |
papers/calib-poison-general/b2_subspace_erasure.py |
papers/calib-poison-general/b2_subspace_erasure_result.json |
same |
| parity control | the flagship private>naive recovery is probe-capacity-dominated (PRIVACY_PARTIAL__between_bars) |
papers/calib-poison-general/honesty_parity_control.py |
papers/calib-poison-general/honesty_parity_control_result.json |
verdict matches; parity_gap sign per cell matches |
| E1 | stage-2 "evasion" was calibration poisoning; private-calibration recovers the read | papers/read-neq-write/e1_clean_auditor.py |
papers/read-neq-write/e1_result.json |
verdict matches; recovery reads within tolerance |
| legibility matrix (CPU-only, easiest) | three models mutually legible label-free cross-family; qwen an island from every direction (LEGIBILITY_LAW_CANDIDATE, matrix licensed / law not) |
papers/disjoint-worlds/run_b37.py |
papers/disjoint-worlds/b37_result.json |
anchor discoveries (0.5918, 0.0536) match to the digit; clique/island topology holds. One command from a clone, no GPU — see REPLICATE_legibility.md |
| island frame geometry (CPU-only, ~4 s — the single easiest check in this repo) | the clique's concept-frames co-align far above random; the island sits below the clique in every seed (SHARED_FRAME_CONFIRMED_GEOMETRICALLY) |
papers/disjoint-worlds/run_b45.py |
papers/disjoint-worlds/b45_result.json |
verdict matches; clique median 0.848 vs null p95 0.0566; qwen below in 5/5 seeds |
| the bridge + dose + cliff (CPU-only) | the island's barrier is causal (0.0612→0.9745 vs random 0.0), rank-2 at core (k*=2), and switch-like (knee t=0.8) | run_b41.py · run_b42.py · run_b46.py (same dir) |
b41_result.json · b42_result.json · b46_result.json |
verdict strings match; medians within tolerance — see REPLICATE_legibility.md |
| OATH corpus audit (CPU-only) | every published claim doc certifies against its receipts, with five disclosed exceptions | python -m styxx.corpus_audit papers/ |
committed *.certificate.json files |
your output matches the expected line below, character for character (CPU-deterministic) |
What the corpus audit prints today, so a divergence means something. As of 2026-09-01 the audit does not come back clean, and the honest bar is that you reproduce the same nine exceptions rather than zero:
corpus papers: 208 certificates | HELD 200 FAILED 8 unresolved 0 verdict-drift 1 receipt-drift 0 incomplete 1 receipt-changed 1
[OATH-HELD] INCOMPLETE-RECEIPTS(changed) CAPSTONE_universal_mind_2026_06_10.md
[OATH-FAILED] ANALYSIS_base_rate_ceiling_2026_09_01.md
[OATH-FAILED] verdict-CHANGED FINDING_behavioral_sycophancy_blackbox_2026_06_09.md
[OATH-FAILED] PREREG_oath_v12_formula_constant_2026_08_26.md
[OATH-FAILED] RECON_oath_external_reach_2026_08_26.md
[OATH-FAILED] RESULT_oath_verified_channel_internal_2026_08_27.md
[OATH-FAILED] RESULT_obligate1_does_not_ship_2026_08_31.md
[OATH-FAILED] RESULT_struct1_beats_the_null_2026_08_31.md
[OATH-FAILED] SYNTHESIS_mention_and_use_2026_08_26.md
This block had itself gone stale, which is worth recording rather than quietly fixing. It
listed six exceptions while the audit printed nine: RESULT_obligate1_does_not_ship and
RESULT_struct1_beats_the_null were published OATH-FAILED on 2026-08-31 and never added here,
and ANALYSIS_base_rate_ceiling joined them on 2026-09-01. The pinned counts were tested;
the pinned list was not, so only the numbers were kept honest.
receipt-drift reads 0 again after a repair on 2026-09-01. It had silently been 1:
external1_summary.json was regenerated by the CORRECTION commit of 2026-08-31 under a
certificate that had already sworn to the older bytes. Re-certifying against the corrected
receipt returns the identical verdict and identical counts — the correction moved the bytes but
no claim the paper swears to — so the certificate was re-issued rather than excused. This is
the third document hit by the same defect, after CORPUS_STATE earlier the same day: a
receipt is history too, and regenerating one in place silently invalidates every certificate
citing it.
Four of those OATH-FAILED lines are deliberate: each is a document accused on the example it
quotes, published failing rather than reworded until it passes.
One is a partial-evidence warning, and it is new. CAPSTONE_universal_mind cites twelve
receipts; one of them, mind_v0_validation.json, is present in the tree with content that is not
what was certified. The resolver refuses to accept a changed file as the certified one — correctly,
because this repository is full of files called *_result.json, and accepting a mismatch would
certify a document against another experiment's data. So the document is certified against the
eleven that resolved. Until 2026-08-27 that printed exactly like a document certified against all
twelve; incomplete and receipt-changed are the counters that stop it. The verdict still holds
on the evidence that resolved, and the absent twelfth is now visible instead of implied.
One is a real drift. FINDING_behavioral_sycophancy_blackbox is recorded in
KNOWN_VERDICT_DRIFT in tests/test_certificate_reproduces.py with its diagnosis: line 13 of
that FINDING is truncated mid-sentence, so a dangling 4 lost the vocabulary that bound it to
n_nogate. It was invisible for months because the guard resolved receipts only next to the
certificate, and that document cites two from another folder — one of 36 certificates the guard
was skipping. Repairing a published document is its own cycle; the entry exists so it cannot be
forgotten.
Three are committed OATH-FAILED on purpose, and each is accused on the example it cites.
The RECON reports that the verifier does not transfer outside this lab and is accused on the
tokens it quotes as false accusations. The SYNTHESIS reports that four instruments cannot tell a
mention from a use, and is accused on the digits of the LaTeX formula it quotes as the specimen.
The v0.12 PREREG proposes to close that formula case and is a member of the class it proposes to
close — with a pre-committed transition saying that when the clause lands, its own verdict and
the SYNTHESIS's must both flip to OATH-HELD, or the cycle under-reached. All three were
published failing rather than reworded until they passed.
A note for anyone replicating on Linux. Receipt hashes in this corpus were recorded from a
Windows working tree, so they are CRLF hashes, while git stores and Linux checks out LF. Until
2026-08-26 that meant cross-directory receipts silently failed to resolve on Linux and their
documents were dropped from the drift guard — CI reported a document "repaired" when it had
merely become invisible. _resolve_receipts now compares content modulo line endings, so the
audit gives the same answer on either platform. If you see a different count from the block
above, that is a divergence worth filing.
The binding: line, and what it is for. Since SPEC_oath_receipt_binding_2026_09_04.md
the audit prints a third line beneath the verdict line and the uncovered line. On a full clone
(--history auto, the default, or on) it reads
binding: over N/M certificates — citations same … at-issue … elsewhere … unbacked … unrecoverable … | documents same … at-issue … moved … | certificates: … — for every receipt a
certificate cites, where the bytes it swore to are: in the working tree (anywhere in it; a
receipt outside the audit root is same with a note), at the certificate's issuing commit,
elsewhere in history, nowhere in history, or unanswerable; the same for the document; and for
every certificate whether the current verifier reproduces the recorded verdict class over the
document and receipt bytes at the issuing commit (stands-over-sworn-bytes, null with a reason
when those bytes cannot all be fetched). On a shallow clone, or with --history off, it reads
binding: history unavailable (…) and no cell is filled; --history on additionally exits 2.
CI never runs this audit, so nothing in CI prints either line; the shallow reason is pinned by
tests/test_certify_by_digest.py on a clone the test builds. The first line is byte-identical
in every mode; it is the only line this file pins. The corpus census over all 213 tracked
certificates is papers/closed-model-frontier/receipt_binding_census_result_2026_09_05.json (the third run; the
second, receipt_binding_census_result.json, stays as committed); the census
refuses to overwrite a tracked result, so a new census is a new dated file; its reading is
RESULT_oath_receipt_binding_2026_09_05.md, sworn to its leaves. One thing it settles for the
block above: CAPSTONE_universal_mind's changed twelfth receipt has its sworn bytes at the
issuing commit, and the certificate stands over the full set — the receipt moved seventeen
minutes after issue on 2026-06-10; the certificate stands over the bytes it swore to. Whether
those bytes were true is not a question this audit answers.
All of it is stated here because a replicator who runs the advertised command and gets an unexplained failure has been handed a divergence that is really our undisclosed known state — which would waste the first thing an outside checker ever does for us.
Tolerance for GPU targets: bf16 CUDA training is non-deterministic across hardware; the honest replication bar is the frozen VERDICT STRING plus per-cell decisive reads within ±0.05 AUROC, with your per-cell deltas disclosed in the PR. The CPU target (corpus audit) must match exactly.
- Clone at the commit recorded in the target's RESULT doc (each RESULT names its frozen commit).
- Run the frozen script verbatim — no flags beyond those in the RESULT's Reproducibility section. GPU targets need ~8GB VRAM and a few hours; the corpus audit needs only CPU.
- Open a PR titled
[replication] <target> — <your name/handle>adding:- your result JSON at
replications/<target>__<handle>.json(verbatim, unedited); - one row to the ledger below (date, hardware, verdict match yes/no, max per-cell delta).
- your result JSON at
- CI runs
python scripts/verify_replication.py <target> replications/<your file>— it checks the verdict string and tolerances against the canonical receipt and posts the comparison. Honest mismatches are merged too, labeledreplication-divergent— divergence is data.
- First matching replication per target: named acknowledgment line in the next Zenodo version of the paper that claim belongs to, plus your row here, permanently.
- First DIVERGENT replication per target (verdict flip that survives our re-run of your exact setup): co-credit on the correction note itself. Breaking our claim earns more than confirming it.
| date | target | replicator | hardware | verdict match | max per-cell delta | PR |
|---|---|---|---|---|---|---|
| — | — | none yet — be first | — | — | — | — |
Published 2026-07-13. If this table is still empty in 2027-01, that emptiness will be reported as-is in the next paper's limitations section: unreplicated is a property of a claim, and we do not hide properties of claims.