[R3.2] Correct cleanliness evidence and freeze admission protocol

[lemone112]

Purpose

Evidence/admission owner for the human AppearanceConvention branch of local-roadmap R3. It cannot directly add a runtime scalar and does not block TechnicalQuality.

Correct the source state

Update the durable ledger for Shinomori & Takata:

• primary paper: doi:10.1364/JOSAA.546984;
• public Figshare supplement v3: doi:10.6084/m9.figshare.28555439.v3;
• record exact retrieved bytes, filenames, SHA-256, license and extraction output;
• verify any previously reported digest against actual bytes; never copy a prose checksum blindly;
• distinguish participant-level raw availability from aggregate numerical supplements.

This source is development/negative evidence, not a future untouched holdout.

Construct-validity protocol

Before human lint/action admission freeze: construct/task/translation equivalence; culture/population/viewing/stimulus/output applicability; measurement invariance; preregistered features and multiplicity; study-level validation; uncertainty/practical margin; conflicts across populations; no pooling of incompatible adjective tasks; exact action class and protected outcomes.

Two sources or one source plus independent published holdout is necessary but not sufficient.

Explicit rejected shortcuts

Absolute chroma, C-star/Cmax, Berns vividness, saturation-is-cleaner, brightness-is-cleaner, one universal clean/dirty threshold, and silent cross-cultural transfer cannot be sole/default cleanliness laws. Sources may provide scoped candidate features or falsification evidence, not direct action authority.

Output algebra

LintAdmitted(profile, scope, uncertainty)
AutoActionAdmitted(profile, action class, non-inferiority proof)
Rejected(claim, reason)
NotExamined(missing source/access/contract)
Experimental(explicit opt-in)

Lint admission never implies action admission.

RED / reproducibility

• immutable source manifest/checksums/license/access;
• executable extraction from source bytes;
• negative/null/contradictory results retained;
• study-level split guard;
• no participant bootstrap without participant rows;
• mutation from ordinal/distributional output to universal scalar fails;
• mutation from lint to auto authority fails;
• protected readability/identity/sentiment/distinction cannot regress;
• no human profile still leaves TechnicalQuality total.

Acceptance

• corrected evidence ledger and captured analysis;
• construct-validity/admission preregistration;
• no invented holdout/universal formula;
• every human release binds exact source/applicability/uncertainty/license;
• separate lint/action verdicts;
• feeds R3 implementation, Implement: bind every public product claim to a versioned evidence and applicability manifest #431, Audit: production-wide coverage matrix and zero-gap disposition ledger #408 and release join.

Refs #408 #431.