PCAP reporting revision: verification record

Implemented and running locally

The new PCAP report-writing stage is available after packet analysis and approved enrichment. It produces a concise, cited analyst-review report and a focused /pcap/:analysisId/report view. The packet evidence and original reports remain intact. All case data is from public training captures; no malware was executed.

Implementation is in the isolated adversarygraph-pcap-quality worktree on branch feat/pcap-readable-summary-20260922, based on 817f74b13d9c4155db6c161e260747df0b0d1abe. Current-turn changes are not committed or pushed, and no website publication was performed. Unrelated changes in the canonical checkout were preserved.

Changed components

Component	Change
`backend/app/services/pcap_story.py`	Bounded short-report evidence selection, prompt, validation, conservative optional-claim filtering, rendering and provenance
`backend/app/api/routes/pcap.py`	Read-only preflight, saved summary access and explicit report-writing endpoint; permissions, TLP, consent and concurrent-source checks
`backend/app/services/investigation_story.py`	Reusable citation-bound generation, bounded repair diagnostics and actual model/usage metadata
`backend/app/services/ai/openai.py`	Strict server-issued citation/indicator enums and returned model metadata
`backend/app/services/ai/claude.py`	Returned model, stop reason and usage accounting including cache input tokens
`backend/app/services/pcap_assessment.py`	Hash-qualified PE-under-image-path file-review candidates; no automatic host-IP promotion
`frontend/src/components/PcapSummaryPanel.tsx`	Explicit final reporting stage, readable report card, exports and separate evidence details
`frontend/src/pages/PcapReport.tsx`, `frontend/src/App.tsx`	Focused live report route
`frontend/src/api/client.ts`	Typed PCAP summary API contract
`frontend/src/pages/Analyze.tsx`	Integrate the final stage and bound oversized investigation handoffs
Backend unit/integration and frontend PCAP browser tests	Evidence limits, provenance, authorization, source changes, readability and handoff regressions
`docs/PCAP-SHORT-REPORTS.md`	Report contract, governance and review checklist

Validation

Final backend regression output: 1,624 passed, 180 skipped, 46 warnings in 64.44 seconds. Skips are not counted as passed integration coverage. Full backend test output.
Frontend lint passed; TypeScript/Vite production build passed. Vite retains an existing large-code-chunk warning; this is not reported as a clean warning-free build.
All four targeted PCAP browser tests passed: report writing/consent/readability, reputation selection, report quality, and large-investigation handoff. Browser test output.
The two previously rejected oversized handoffs returned HTTP 200 through the real local UI. Original texts were 297,320 and 502,983 characters; bounded handoff descriptions were 537 and 541 characters. Full reports remain stored. Okay-Boomer handoff proof, Eggnog Soup handoff proof.
Selected deployed backend source hashes are checked against the worktree, and exact service image IDs are retained in runtime provenance.
Original packet results and internal-context snapshots are checked against the 490-file baseline freeze. Final native reports have a separate pre-answer freeze.
Every published case screenshot must have a matching validation JSON recording live IDs, hash, model, DOM checks, image hash and completed visual inspection. Image presentation success does not mean investigation correctness.

All 20 final report-card images now have matching checksums and individual visual inspection records: 16 show saved reports and four show an explicit withheld state. No screenshots were fabricated or substituted with answer-assisted report images.

The final optional-citation reliability fixes were deployed in backend image summary-20260922-v11; the frontend is summary-20260922-v6. Their unit and browser regressions passed. The four intended live Claude retries were blocked by the execution approval mechanism, pending explicit consent to send the exact bounded public-training metadata to Anthropic. They are not reported as successfully retested. Subsequent reference-answer review did not alter any frozen native report or submit any answer text to a model.

Known boundaries

Strict JSON, exact quotations and literal-identifier checks do not establish semantic correctness. Current reputation is not capture-time reputation. A download is not execution, and missing telemetry is not a clean verdict. Global negative claims, weak POST-based leads and incomplete attack chains are assessed explicitly in the per-case answer comparison instead of being hidden by a passing software test result.