The Indus script, tested claim by claim.

Every research program accumulates results it likes too much to re-examine. This week we did the opposite on purpose: we adopted a stricter study framework — claims with stable IDs and named falsifiers, evidence counted by origin rather than by report, controls that must be able to fail, and verdicts that include CONTRADICTED and INVALID alongside SUPPORTED — and then we applied it retroactively to everything we have published in this program. 116 prior items (Specs 001–025, Phases 52–141) were classified under criteria that were frozen before anyone looked at which items the criteria would hit. Eight items were rerun under pre-declared corrections. The outcome: 75 of 116 items carry a changed evidentiary status. The originals were not edited. Each change is an appended historical assessment, so the record shows what we believed, when, and what replaced it.

What fell

• The keyed-transcription pilot is now CONTRADICTED, not merely stopped. Rescoring from the raw records gives two-pass agreement 0.20 with a confidence interval of [0.10, 0.32] against a pre-registered floor of 0.80. That is not a near miss, and the dataset stays unpublished. A retry needs human expert transcription — no third attempt with AI coders is a reopening path.
• The “site dialects” result on the Holdat layer is CONTRADICTED at a declared margin. The bounded claim's interval, [0.059, 0.078], sits entirely below the pre-declared practical margin of 0.10, converging with the original test: on this layer, site differences are a preservation artifact.
• Three test batteries and the blind affiliation rounds are INVALID — worse than null. Their instruments could not discriminate the hypothesis from grammar-free controls (in one case a random generator won the blind test outright). We have stopped describing these as “rejected at calibration.” They never measured anything.
• The motif-coding pilots soften to INCONCLUSIVE. The honest surprise ran the other way: Phase-134's agreement interval [0.77, 0.91] reaches its gate, so the story we told at the time — “missed by one object” — was more confident than the data supports. The motif arm stays closed, but for the recorded reason (the AI instrument is unstable on worn plates), not a borrowed certainty. The audit also established structurally that all AI pilot coding was a single evidence origin group: agreement between two passes of one model family is consistency, not independent corroboration.

What stood

• Terminal-class × object type (seals vs tablets, Mantel-Haenszel OR 2.384, CI 1.847–3.077) was untouched by the reruns and remains SUPPORTED, stable under leave-one-site-out.
• Site repertoire differences on the ICIT-lineage layer stand (interval [0.182, 0.210]), and the preservation-controlled follow-up (G1) stands at its declared margin (interval [0.184, 0.210]) — with its named weakness intact: chronology is not controlled, and we say so wherever the result appears.
• Four foundational results replayed exactly in a disposable environment (Phases 116, 125, 127, 131), including the finding that 93.9% of the disagreement between published corpora remains unexplained.

The part we are proudest of

One of the audit's own rerun methods failed its independence control — its confidence interval excluded zero even at a true effect of zero — and was declared INVALID under the framework's rules, then replaced with a corrected method whose results are the ones reported above. A framework that cannot catch its own side cheating is decoration. Ours caught one, in its first week.

What this does not change

The script is not deciphered, and nothing here moves that line. The strict core remains 94 readings; the 287-anchor set is byte-identical before and after the audit; PRED-2026 remains pending, waiting for an independent corpus. What changed is the quality of the ground the next studies stand on — and the program's newest study, on population movement, culture, and language (Spec 027, frozen today), is the first to be built entirely under the new rules.

Links: Spec 026 record · Movement study · Zenodo archive