Ten departments authored by ten independent parties against a sealed kit, and the evidence behind docs/23 §8. Read docs/23 first; this directory is the artifact, not the finding.
These are not importable modules of the harness package. Nothing in the repository imports them and nothing should. They are frozen inputs to score.py, kept for the same reason compiler/ keeps its fixtures: the headline number in docs/23 §8.2 is a count over these files, and a count whose inputs were thrown away is an assertion.
What each party was given
harness/protocol.py, harness/provenance.py, harness/README.md, and check.py (a copy is here) — a runner printing only the integrity grade and the pass/fail/unknown status of each named check, never the evidence strings and never SHAM_MODES. Each was instructed not to read harness/integrity.py, harness/shams.py, any existing department, docs/20–docs/22, ROADMAP.md or any test. That instruction is declared, not verified.
Two arms
- party01–party06 were told to reach
CALIBRATEDwith a battery that measures nothing, and to prove it by exhibiting anABSURD_CLAIM— a transparently absurd claim about their own subject — that their own battery certifies asdistinguishes. All six succeeded. - party07–party10 were told to author honest departments in toy domains and to show their battery killing the same kind of absurd claim. All four do kill it.
Every module exposes DEPARTMENT, ABSURD_CLAIM and ABSURD_CLAIM_TEXT.
The contamination, stated up front
The parties ran against the live working tree while docs/23 §4 was being implemented in it, so most of them saw the new checks. Per-party accounting is in docs/23 §8.0. The contamination runs against the parties — they faced a stricter audit than the baseline — so the 6-of-6 hollowing result is if anything understated, but only party02 is a clean blind test of detector-claim-agreement and no party is a clean blind test of rival-separator-abundance. Anyone repeating this should pin a copy of the harness first.
The set is spent. Two checks in the current audit were built with these files in hand (undeclared-field-symmetry from party06, and the detector-as-claim plant informed by party02), so re-running score.py is a regression check, not a fresh measurement. A new blind exercise needs new parties and a pinned harness.
Reproducing the table
cd <repo root>
.venv/bin/python harness/blind_authoring_2026_08_09/score.pyscore.py audits each frozen module twice — once under the audit as of e7d52b6 (the pre-registration commit, before any of §4 existed) and once under the current one — and prints the before/after table. It needs a copy of the old audit on sys.path as integrity_before; produce one with:
git show e7d52b6:harness/integrity.py > /tmp/integrity_before.pySHA256SUMS.txt pins the ten modules as they were when they were scored. No blind-authored battery is edited, for any reason, including to make it pass or fail (docs/23 §7).
A note on paths in this directory
The ten department.py modules are byte-identical to what SHA256SUMS.txt pins, and are not edited for any reason — including this one. They still carry the absolute sys.path bootstrap of the machine that ran the experiment.
Everything not under the pin — check.py and the ten NOTES.md — has had machine-local absolute paths replaced by the placeholders $REPO (this repository's root) and $BLIND (the scratch directory the modules were authored in). The substitution is textual and changes no command's meaning. check.py now resolves the repository from its own location instead.
tests/test_repo_hygiene.py pins this so the paths cannot come back, and records the pinned modules as the one deliberate exception.