[jrh-mann]

Findings handoff

One Inspect agent investigates, submits a portable immutable findings bundle, then continues publishing in the same conversation. inspect_audit/publish can start a fresh publishing agent from that bundle. Publisher plugins remain replaceable.

• Bash/Python edits /workspace/findings.json; collected auditor records and subject remain host-authoritative.
• Findings, reviews and ratings include explanations and evidence. Submission checks sample references, original auditor records and the selected publisher's requirements before freezing.
• Publishing skills/tools are withheld until submission. Job tools disappear afterwards, and host launch code refuses new experiments.
• The bundle preserves selected evidence and a checksum manifest. Native logs remain external references.
• Checkpoints preserve the submitted claims, manifest and phase. Progress and publication share one interpretation of missing/unresolved sample/check results.

Default publisher: Laurence's GL LaTeX audit report

latex is now the default; publisher=gl retains the optional web publisher. The publisher downloads the actual template from Generality-Labs/audit_reports at ad7de89658b6794112c2f65847991c7ed41d9e30, verifying every file against publishers/template.json. The class, style, logo, scoring criteria, build configuration and Example structure are fetched into a host cache. No TeX template copy is vendored. Investigation uses generated analytical registry data from that revision, including the correct A.0 contribution numbering.

The output follows the upstream report: cover sheet, contents, A-D eval scorecard with nine dimensions and key contributions, datasheet, audit setup, dimension findings and scoring criteria.

• Grade is frozen in investigator.ratings.overall; assessments use framework criterion IDs.
• Findings carry their criterion, e.g. C or C.1. Contribution findings render as A.n boxes, and broader dimension findings as sequential F-n boxes.
• Code generates rows, assessments, finding boxes, counts and evidence references from the bundle. The publishing agent supplies only escaped prose in prose.json.
• latexmk runs on the publishing host or in the separate publishing image with networking disabled, never in the investigation sandbox.
• Output preserves the PDF, generated source, prose, template provenance and original findings bundle.

The GL web plugin still supports predefined bars, stacked bars and histograms and emits HTML/PDF/SVG with data provenance.

Cleanup

The superseded LaTeX helpers and their separate findings schema are removed, along with review/finding/rate recording tools, assessments.json, register.json mirroring, old report sync, old publishing skills and templates. This is a publisher over the shared frozen bundle, not restoration of the old reporting pipeline. TeX is confined to publishing dependencies.

Validation and CI

Before the request to stop local heavy work:

• 251 non-Docker tests passed, with the TeX integration case skipped in that run.
• The real fixture compiled into an 11-page PDF in the separate publishing image. All pages were visually reviewed. Its scorecard, nine rows, assessments, both finding-box formats, sample counts and evidence were checked against the frozen fixture.
• The PDF test's expected F-number was corrected to F-1 to match template dimension order; all assertions were checked against the existing PDF without recompiling.
• Type checking and formatting hooks passed.

No more local heavy tests/builds. Heavy verification is delegated to CI. A dedicated LaTeX fixture job uses the audit-template environment and its read-only AUDIT_REPORTS_TOKEN secret. The dedicated job now fails immediately if that read-only secret is missing; it must not pass with its private-template tests skipped. No paid model calls were made.

README is untouched. Hawk 4 deployment remains unvalidated and tracked in docs/hawk-4-migration.md. See docs/findings.md for contracts, cache/auth setup and publisher commands.

Python line budget

Compared with main: +673 package Python lines and +78 test Python lines (+751 total) for the shared handoff and publisher functionality after the removals above. #67 reduces production Python by 5 lines and adds 33 lines of authorizer tests; #68 changes no Python.

Related PRs and follow-up

• Restore benchmark task-argument authorization and remove gpt-5-mini #67: main-based task-argument authorization fix and removal of gpt-5-mini. Merge separately.
• Split safe dev-tool bumps from #64; retain Hawk runtime pins #68: dev-tool bumps extracted from Bump the python group across 1 directory with 7 updates #64, preserving semver<3.1 and Inspect 0.3.266. Bump the python group across 1 directory with 7 updates #64 must not merge as-is.
• After this PR merges: a separate behavior-preserving role rename, with auditor meaning the top-level agent and investigator meaning the per-sample agent. That rename is deliberately not mixed into this PR.

Registry and private-cache follow-ups (fa0ea20)

• test_registry_is_regenerated_from_the_pinned_template compares the complete generated JSON with regeneration from verified upstream sources, including all 56 dimension/contribution definitions, all four grades, all six assessment descriptions and provenance. A stdlib-only check against the existing verified cache passed locally; no local heavy tests were run.
• python -m inspect_audit.publishers._template --update-registry is the reproducible update command when the pin changes.
• A named-context Dockerfile can layer the verified private template cache into the Python publishing runner image without putting a GitHub token in the image or pod. The compiler image alone does not provide the Python host's cache. The derived image must remain private.
• CI includes an offline cache-image test with networking disabled and no runtime GitHub credential, alongside registry regeneration and PDF compilation. These heavy checks are CI-only; default pytest excludes both Docker and LaTeX markers.
• The audit-template environment currently has no AUDIT_REPORTS_TOKEN. A fine-grained read-only token for audit_reports must be supplied by the operator; the personal gh-login token was not copied into CI. Missing access is now an explicit failing check.
• Coordination: this branch does not move Claude's dispatch, log-resolution, pricing or launcher modules. The optional GL web publisher stays as previously requested. Role renaming remains a separate follow-up after Freeze portable findings before exposing pluggable publishing #66 merges, before real findings bundles are saved.