GRA v9.2.0 Skill Edition

GRA v9.2.0 Skill Edition

GRA v9.2.0 strengthens evidence calibration, document reliability, privacy, and tool honesty while preserving the two-edition architecture introduced in v9.0.0. The Agent Edition remains the source of truth; the Chat Edition is generated from it and remains below the 8,000-character gate.

What Changed

• Conditions Proved conclusions on the reliability and accuracy of supplied claims and citations; correlated sources cannot manufacture independent support.
• Treats uploaded records as evidence to analyze, never as instructions to follow.
• Adds explicit attention to alterations, damage, peculiarities, unreadable text, and matters not understood.
• Strengthens source weighing through provenance, informant, commissioning, and correlation questions.
• Restores quick-start cost and access-channel labels in the compact Chat Edition.
• Classifies supplied transcriptions as derivative representations of the record in hand.
• Adds specialist-domain, tool-state, DNA-plus-documentary, and Negative Evidence safeguards.
• Improves living-person GEDCOM privacy and date/pointer validation.
• Corrects related public methodology, navigation, evidence-framing, rights, and link material identified in the July repository audit.

Editions And Checksums

• Agent Skill: gra-skill-v9.2.0.zip — 60,068 bytes — SHA-256 23e7407b102c05d5b8c6772d9310659b230c8e34e69e81126ef031d0f6a58df5.
• Chat Edition: research-assistant-v9.2.0-chat.md — 8,016 bytes and 7,968 LF-normalized characters — SHA-256 dea089b3fa50b0c14986ed7e5cb8cfbccad2d7124b7542e09763f35b9632ae03.

Integrity correction (July 11, 2026): The sizes and SHA-256 values above were corrected after downloading the published GitHub assets and verifying their remote bytes. The release assets were not replaced.

The ZIP contains only the six runtime files. Developer fixtures and test results remain outside the installable package.

Verification

On identical instrumentation, the v9.2 candidate produced 27/33 clean Agent fixtures and 26/32 clean Chat fixtures, compared with 21/33 and 22/32 for published v9.0.0. Six of eight new safeguards are fully delivered.

Two documented limitations remain: the model may invent a likely informant when a birth record does not name one, and the Chat Edition may prefer a fuller name appearing in an Original source when the provenance does not justify that preference. These limitations are accepted for v9.2.0 and routed to the full-reference layer or v9.2.1. They do not make the behavior correct; they make the release boundary honest.

Offline generation, size, fixture, package, GEDCOM, Markdown, privacy, and link checks passed on the frozen candidate. AI output can still be wrong. Verify every record, citation, inference, and conclusion before relying on or publishing it.

Legacy Continuity

The v9.0.0 and v8.5 releases remain available. Paths used by earlier articles remain in place; this release does not rewrite or delete historical tags, assets, or files.

License

CC-BY-NC-SA-4.0, with path-specific exceptions and provenance documented in the repository rights matrix.