arxiv-v1

Add reading-mode toggle and fact-checked Q&A walkthrough (#17)

Overview gains a Technical / Plain English toggle (paired content blocks, localStorage-persisted). New qa.html: a three-part conversational walkthrough — the study in questions, the details people ask next, and the findings that didn't make the paper's main arc — fact-checked against the archived record and linked into the research graph.

Phase 5 registry v4 seal

Phase 5 (persona × temperature) registry v4 seal.

• Registry sha256: c1e7254324c5dbbe185792af0d85b8af587e7686478275ce85f70fadffe42be6
• arms.json: ab753651b4c0e335f180af1d9c60a37c32c855dfc151b3f6bcf89d412ba5b018 (160 arms)
• execution-schedule.json: 04ceffd74026cea5ed1157f8d704087dda26180c8dd0fd9d20b48183215055d0 (1,488 episodes)
• personas-v4.json: a0b1d7b884c62ced4f48b01fe91bfd1490d3c2d9c2ca77ec59409889ef4ec893
• Seal-gate lint: PASS (160 cells, 5 three-layer sealed rules)
• Anchor: docs/phase5-close/SHA256SUMS.txt + OpenTimestamps .ots
• Seal record: docs/phase5/seal-record.md

Phase 4 final — sealed program complete

Final verdicts: docs/phase4/final-report.md. Verification: docs/close-out-verification.md — fresh clone, zero secrets, zero live calls, replay CLEAN 2,864/2,864 byte-exact, adjudicate-all identical to the committed record.

Assets: final report, sealed registries (registry.json, arms.json), all adjudication records, the complete data artifacts (event store engine.db.xz, budget ledger budget.db.xz, driver state), SHA256SUMS.txt and its OpenTimestamps proof (second anchor; first anchor is this tag's GitHub server timestamp). Restore path: scripts/restore-data.sh.

Not GPG-signed — disclosed deviation (no signing key in the execution environment), same as phase4-v3-seal.

Phase 5 final (close-out)

Phase 5 final — close-out release

**WORKING DRAFT — PRE-PUBLICATION, NOT FOR CITATION.** Repo remains private until the coordinated drop.

- Final report: `phase5-final-report.md` (Branch 2 selected — "an interior persona exists"; sealed sha byte-identical)
- Adjudication record: `adjudication-report.{json,md}`, outcome-blind decisions ledger, branch-selection record
- Replay audit: CLEAN — 1,712/1,712 byte-exact, 0 invalid trials
- Data: `engine.db.xz`, `budget.db.xz` (sqlite-backup snapshots), driver state/plan — sha256s in `DATA-SHA256SUMS.txt`
- Analysis pack (exploratory): `analysis-pack.tar.gz` (docs/analysis/)
- Close seal: `SHA256SUMS-final.txt` + OpenTimestamps proof `.ots`

Verify: `sha256sum -c DATA-SHA256SUMS.txt`; replay per README (zero live calls, zero secrets).

Phase 4 registry v3 — sealed (externally anchored)

Phase 4 registry v3 seal — external anchor

Seal commit: d24ee94ea722d817f0382519445c5e9d11a7bb1f

The only change at sealing: registryVersion phase4-v3-proposed → phase4-v3 (one line). Zero template, arm, seed, binding, or schedule bytes changed.

Artifact	sha256
registry.json (post-seal)	0c084b73a38c63ccfc95622a78b21d2a2c34113e0648b1d7037a5a6ce0695f05
registry.json (pre-seal, byte-identical to Gate-0 record)	2ef069db5ce28b6837794aea70e09ed77b7a3018963e29bdcaf70c7857536eb4
arms.json (250 arms, 49 template shas)	eba6b7f10976598dea719d6ebf7921b1f81d6fc93d4b49e3768753bec1325a62
execution-schedule.json	139c1b6d514487ea2412d2ad0fa8bda36f79dffed59f4955b0180452239fa444

Assets attached: post-seal registry.json, arms.json, SHA256SUMS.txt.

This release's GitHub timestamp is the public anchor for the sealed Phase 4 design (freeze packet docs/phase4/freeze-packet.md, seal record docs/phase4/seal-record.md). Language rule: externally anchored — not "cryptographically immutable". Tag is annotated but not GPG-signed (no signing key in the execution environment; disclosed in the seal record).