Run report — A Confession (Исповедь) work-dive

Mode: autonomous (--auto), run deep / multi-pass with a heavy visual fan-out. Date: 2026-06-06 · Slug: 1879-1882-a-confession · Work id: confession Scope approved once upfront (the reader chose Heavy visuals + Go-deep), then detached.

Phase 0 — scope contract (as run)

• Question. Document A Confession as a work: genesis & composition, what the text says, redactions, publication / censorship / translation, reception (Russian society & Church first), and place in the Prophet-period canon — grounded in Tolstoy, the Jubilee apparatus, the composition-window diaries/letters, and prior dives. Propose a workRecord: to fill the near-empty Confession.md.
• Corpus surface. PSS Tom 23: v23_001_059_Ispoved (text, pp. 1–59), v23_488_511_…Plany_i_varianty (variants, pp. 488–511), v23_515_537_Ispoved (commentary, pp. 515–537). Composition-window diaries Tom 48 (1878) + Tom 49 (1881–82); letters Tom 62–63; sibling В чём моя вера? (Tom 23); echoes (Tom 26 Notes of a Madman, Tom 66 letter to Dumas). Period: post-1880 Prophet years; OS dates.
• Pins discovered. Composition ~1879–1882 (roots 1877; complete 1879 treatise reworked into «Исповедь»; finished form ~early July 1881, editor-hedged; type-set 1882). Diary silence 1879–80 (notebooks, not diaries). Title «Исповедь» not Tolstoy’s (first in Elpidin 1884).

Models used (cost note)

• opus: work deep-read; Tom-23 commentary mining; Phase-3 scholarship + triangulation; the Phase-5 verifier (fresh context).
• sonnet: 4 Phase-1 sweeps (diaries, letters Tom 62, letters Tom 63, broad echo); 3 visual channels (Tolstoy / people-places / editions); the PD facsimile render; the visuals dedup/consolidation.
• Rough order ~1.1M subagent output tokens across ~12 subagents + main-context synthesis. Mechanical steps (extract_tei, verify_quotes, serve.py, build_evidence_index, image download) used no model.

Verification (Phase 5)

• Mechanical gate: verify_quotes.py docs/research/1879-1882-a-confession/dossier.yaml → 27/27 quotes verbatim, 0 label warnings — PASS.
• Cross-dive aggregation: build_evidence_index.py --check → exit 0 (0 unresolved refs; entities register cleanly).
• Render: serve.py --build-only → index.html (48 KB) + INDEX.html rebuilt, exit 0.
• Independent verifier (opus, fresh context): verdict CLEAN-WITH-MINORS, no must-fix. Confirmed primary-anchored claims, attributed secondary, valid triangulation refs, accurate vaultStatus, the cousin/daughter A. A. Tolstaya distinction, rights hygiene (visuals git-ignored; only PD in extracts/; nothing in website/src/), honest coverage, evidence-anchored dates. Only minor: workRecord value shapes (flat vs object arrays) to reshape at human ingestion — already in needsReview.

Evaluation self-assessment (the deferred gate, self-scored; not run live)

1. Did the composition-years interlocutor sweep yield people as person entities? — YES. Strakhov, Fet, Alexandrine (cousin), Fedorov, Sutaev, Alekseev, Engelhardt, Yuryev, Elpidin, Pobedonostsev, Turgenev, Urusov surfaced from the letters/diaries and are routed in entities with ingestionPriority. Strong.
2. Is the Russian society & Church reception real and source-confirmed? — PARTIAL (honest). 1880s critics (Gromeka, Ostroumov, Gusev) named via the PSS commentary; the 1882 censorship and the 1901 Synod edict (“names no works”) confirmed (commentary + scholarship). A dedicated reception sweep was not run → coverage: partial, logged in notCovered.
3. Is the workRecord fill accurate and provenanced? — YES, with the shape caveat. Every proposed value is evidence- or commentary-anchored; uncertain ones carry confidence: medium + notes; the two record fields that are currently wrong (publishedDuringLifetime, publishedInRussiaDuringLifetime = false) are flagged to flip to true. Value-shape reshaping deferred to ingestion (needsReview).
4. Is coverage honest (no covered that is really partial)? — YES. Verifier spot-checked; partial/not-covered calls (variants, sibling work, reception, 1901 text, 1879–80 diary) match the prose.
5. Did --choice=reg extract cleanly? — YES. All 18 extractions ran with no dropped-pair warnings; pre-reform orthography resolved to modern.
6. Did the spine stay bare? — YES (with one caution). Contested labels (“crisis”/”conversion”/”Tolstoyanism”) are attributed to biographers, not adopted; the dive leads with «переворот». “crisis” appears in running prose only as the frame being corrected or a neutral descriptor — acceptable, watch on future dives.

Outputs

• docs/research/1879-1882-a-confession/index.md (+ git-ignored index.html)
• docs/research/1879-1882-a-confession/dossier.yaml (evidence ×27, entities ×24, scholarship, workRecord, coverage, visuals, needsReview)
• docs/research/1879-1882-a-confession/extracts/ — 18 byte-faithful .txt extracts + 1 PD facsimile (v13_MyConfession_opening_facsimile-019.jpg, English Wiener 1904)
• docs/research/1879-1882-a-confession/visuals/ — 27 PD images cached (git-ignored; repopulate with fetch_visuals.py 1879-1882-a-confession)
• docs/research/1879-1882-a-confession/session-log.md, run-report.md
• website/src/posts/notes/2026-06-06-a-confession.md (draft: true)

For the reader’s return (deferred, with Johan)

• Human evaluation gate (spec §5): review this self-assessment; decide whether to fold any fixes into SKILL.md/extract_tei.py before running the rest of the canon.
• Pending memory updates (from the handoff): reverse corpus-dive-human-present (unattended --auto runs are wanted again — confirmed this session); reconfirm in-session claude -p --allow-dangerously-skip-permissions is denied by the auto-mode classifier (Johan launches headless, or drives in-session under accept-edits, as here). Left for the with-Johan ritual.
• needsReview highlights: the 1884 Geneva imprint wording (Genève vs Carouge); identity of the “book exposing the deceivers” in two 1882 letters; the 1884–85 Russkaya-Mysl-ban letters likely concern What I Believe, not A Confession; the excommunicationRelated field intent; the cousin Alexandrine needs a NEW wiki page (not the daughter’s).
• Wiki ingestion is a separate human step — the dossier is the pointer, not the writer. Entity work-order (priority 1 first): the «переворот» concept page + Strakhov; then Alexandrine (cousin), Fet, Fedorov, Sutaev, Yuryev, Elpidin, Pobedonostsev, Holy Synod, Russkaya Mysl, Optina Pustyn; then Turgenev, Urusov, Alekseev, Engelhardt, Rumyantsev Museum.