# DECISIONS — attributed log

Tags: `[regis]` sponsor · `[F]` Fable · `[O]` Opus · `[S]` Sonnet · `[FL]` Flash. One line per decision; append-only.

- **D-001** 2026-07-30 [regis] Formal three-round council with separate Fable/Opus/Sonnet seats is mandated. [F] Noted: my default conductor doctrine prefers a narrower panel; sponsor mandate overrides — protocol written as requested, reservation recorded here, not re-litigated later.
- **D-002** 2026-07-30 [F] Control plane lives in `00_ADMIN/`; every session resumes via `NOW.md` → `TASKS.md`. Disk state is the report — no out-of-band status.
- **D-003** 2026-07-30 [F] Research split into 6 non-overlapping bundles (B1–B6) sized for bounded Flash passes; all interpretation/weighing reserved for judgment models (see briefs §"Judgment work"). Flash never concludes, only collects.
- **D-004** 2026-07-30 [F] Self-signed TLS certificate observed at `baby.nurt.cloud` recorded as a hard deployment blocker (gate G2). No public "live" claim until proxy/certificate ownership is established and a trusted cert is served.
- **D-005** 2026-07-30 [F] Niche/beachhead selection deferred to P3, after evidence (P1) and council (P2). Anti-goal: premature convergence on childcare merely because it is the seed hypothesis.
- **D-006** 2026-07-30 [F] Concurrency: sponsor cap of ≤3 sessions adopted as the hard ceiling; intersected with box discipline (≤2 concurrent subagents, serialize unless boxguard GREEN). Council seats may therefore run sequentially — the protocol's independence rules make ordering irrelevant to fairness.
- **D-007** 2026-07-30 [F] Evidence contract: every load-bearing claim carries URL + access date (+ verbatim quote or datum where feasible) or the label `UNVERIFIED`. Invented or unlabelled claims are protocol violations and get struck, not repaired silently.
- **D-008** 2026-07-30 [F] B2-A foreground Opus attempt failed with `Error: Reached max turns (16)` (exit code 1); local verification found no `01_RESEARCH/evidence/B2-A-global-marketplaces.md` and no B2-A source-register rows. B2-A remains the next slice; failure receipt: `reports/evidence/b2-a-opus-failure-2026-07-30.md`.
- **D-009** 2026-07-30 [O] B2-B completed as a bounded, corpus-only artifact at `01_RESEARCH/evidence/B2-B-global-marketplaces.md`; nine B2-B source/receipt rows were appended to `01_RESEARCH/evidence/SOURCE_REGISTER.md`. The corpus establishes fetch outcomes only (one Yoopies title, two Yoopies 404s, three Rover 403s, three UrbanSitter 403s); 0/12 requested model/revenue/breadth/scale fields are established, so all three comparators remain unclassified and no business-model conclusion is carried forward.
- **D-010** 2026-07-30 [F] The first B2-B write attempt was refused by the harness permission layer and an intermediate retry was rejected by CLI stdin parsing; recovery used one fresh foreground Opus call with the required model/permission/session flags and allowed local write tools. Failure and recovery details are preserved at `reports/evidence/b2-b-opus-failure-2026-07-30.md`. The self-signed TLS/proxy blocker remains unchanged at `05_PROTOTYPE/DEPLOYMENT_GATES.md` G2.
- **D-011** 2026-07-30 [O] B2-C completed as a bounded local-corpus artifact at `01_RESEARCH/evidence/B2-C-managed-and-specialists.md`; twelve B2-C source/receipt rows (`B2-C-01`…`B2-C-12`) were appended to `01_RESEARCH/evidence/SOURCE_REGISTER.md`. The corpus establishes a self-service multi-care marketplace (Sittercity), a benefit-funded/direct-pay managed companion-care flow (Papa), and a franchise/network elder-care model (Honor/Home Instead), while recording no monetary amount and no niche recommendation.
- **D-012** 2026-07-30 [F] The fresh foreground B2-C Opus invocation used `--model opus --permission-mode default --no-session-persistence --max-turns 28` but the terminal returned `Execution error / [Command timed out after 600s]` (exit `124`). Post-call local verification found the readable 612-line artifact, all twelve parseable JSON receipts, twelve empty `.err` files, and exactly twelve matching register rows; the execution receipt is `reports/evidence/b2-c-opus-failure-2026-07-30.md`. No deployment action occurred; the self-signed TLS/proxy blocker remains unchanged at `05_PROTOTYPE/DEPLOYMENT_GATES.md` G2.
- **D-013** 2026-07-30 [F] B2-S capture-first corpus completed with 27 parseable `webget --json` receipts and 27 empty `.err` sidecars under `reports/evidence/B2-S-raw/` (three URLs each for Care.com, Babysits, Yoopies, Rover, UrbanSitter, Sittercity, Papa, Honor, and Home Instead). The required fresh foreground Opus synthesis with `--model opus --permission-mode default --no-session-persistence --max-turns 28` timed out at the 600-second terminal limit (exit `124`); local verification found no `01_RESEARCH/evidence/B2-mature-market-comparators.md` and no `B2-S-` source-register rows. Failure receipt: `reports/evidence/b2-s-opus-failure-2026-07-30.md`. B2-S remains open; no niche or deployment action occurred, and the self-signed TLS/proxy blocker remains unchanged at `05_PROTOTYPE/DEPLOYMENT_GATES.md` G2.
- **D-014** 2026-07-30 [F] A fresh B2-S capture-first corpus was rebuilt and locally verified as 27 parseable `webget --json` receipts plus 27 empty `.err` sidecars under `reports/evidence/B2-S-raw/`. The required fresh foreground Opus retry again used `--model opus --permission-mode default --no-session-persistence --max-turns 28` and timed out at the 600-second terminal limit (exit `124`); local verification found no `01_RESEARCH/evidence/B2-mature-market-comparators.md` and zero `B2-S-` source-register rows. Retry receipt: `reports/evidence/b2-s-opus-retry-failure-2026-07-30.md`. B2-S remains open; no niche or deployment action occurred, and the self-signed TLS/proxy blocker remains unchanged at `05_PROTOTYPE/DEPLOYMENT_GATES.md` G2.
- **D-015** 2026-07-30 [F] A third fresh B2-S capture-first run completed its designated 27-receipt corpus: 27 parseable `webget --json` JSON receipts plus 27 empty `.err` sidecars under `reports/evidence/B2-S-raw/`, with status tally 11 HTTP 200, 6 HTTP 403, and 10 HTTP 404. The required fresh foreground Opus call used `--model opus --permission-mode default --no-session-persistence --max-turns 28` and again timed out at the 600-second terminal limit (exit `124`); local verification found no `01_RESEARCH/evidence/B2-mature-market-comparators.md` and zero `B2-S-` source-register rows. Failure receipt: `reports/evidence/b2-s-opus-retry2-failure-2026-07-30.md`. An older unreferenced `babysits-safety.json` / `.err` duplicate remains in the raw directory and was excluded from the current 27-receipt tally. B2-S remains open; no niche or deployment action occurred, and the self-signed TLS/proxy blocker remains unchanged at `05_PROTOTYPE/DEPLOYMENT_GATES.md` G2.
- **D-016** 2026-07-30 [F] B3 capture-first collection completed as 25 parseable `webget --json` receipts plus 25 empty `.err` sidecars under `reports/evidence/B3-raw-2026-07-30/` (eight public platforms, within the five-source-per-platform cap). The single fresh foreground Opus pass used `--model opus --permission-mode default --no-session-persistence --max-turns 28` over the local corpus and returned `Error: Reached max turns (28)` (exit `1`); local verification found no `01_RESEARCH/evidence/B3-trust-safety-mechanics.md` and no B3 source-register rows. B3 is not done and remains blocked/open; receipt `reports/evidence/b3-opus-failure-2026-07-30.md`. No niche, deployment, B2-S retry, or external action occurred.
- **D-017** 2026-07-30 [GO] OpenCode Go credentials were rotated and verified with `opencode-go/glm-5.1` (`GO_CLI_SMOKE_OK`). The direct Go REST path remains runtime-WAF-blocked (`403/1010`), but the CLI path succeeded. [GO] completed B3-A at `01_RESEARCH/evidence/B3-A-trust-mechanics-marketplaces.md` from the deterministic local extract over nine Niania/Sitly/Babysits captures; controller validation confirmed ≤250 lines, all source files, and required citations. B3-B remains a separate ≤6-URL policy/help gap slice.
- **D-018** 2026-07-30 [GO] B3-B capture-first collection completed with exactly six first-party public Niania/Sitly URL probes under `reports/evidence/B3-B-raw/` (six parseable JSON receipts and six empty `.err` sidecars). The single fresh foreground `opencode run --model opencode-go.glm-5.1` pass wrote `01_RESEARCH/evidence/B3-B-policy-help-mechanics.md`; independent verification confirmed 72 lines, six matching receipt citations, first-party URLs only, honest 200/404 limits, and no email/personal name in the artifact. Source-register rows B3-B-01…06 were appended. B2-S stayed on technical hold; no niche, deployment, DNS/Traefik/Coolify, cron, outreach, or product action occurred.
- **D-019** 2026-07-30 [O] B4 completed as a bounded primary-source collection: six official Sejm ELI source documents covering the captured child-care contract form, employment-agency registration wording, minor-care register duty/exception, elder/social-care boundaries, animal-welfare duty, and a superseded 1974 employment baseline. The 218-line artifact is `01_RESEARCH/evidence/B4-polish-legal-regulatory-sources.md`; receipts are under `reports/evidence/B4-raw/`; rows `B4-01`…`B4-06` were appended only after independent local URL/citation verification. The S2 HTML zero-byte response and the Opus collection-pass timeout are preserved as receipts. The artifact states current-law and product-flow questions not established; it gives no legal advice, niche selection, deployment, or product recommendation.
- **D-020** 2026-07-30 [O] A fresh foreground Claude Code Opus pass over the existing local B4 corpus overwrote `01_RESEARCH/evidence/B4-polish-legal-regulatory-sources.md` at 219 lines (≤220). The artifact remains bounded to contract/employment classification, employment-agency registration, personal/verification data, and child/elder/pet category boundaries; it marks current-law and product-flow questions not established, preserves the S2 HTML zero-byte alternate-representation failure, and gives no legal advice or recommendation. Independent GET, hash, URL/final-URL, receipt-field, citation, and register checks passed in `reports/evidence/b4-independent-verification-2026-07-30-v2.json`; the pre-existing six matching B4 rows required no duplicate append. No B2-S retry, niche, deployment, DNS/Traefik/Coolify, outreach, data collection, or product action occurred.
- **D-021** 2026-07-30 [O] B5 completed as a bounded official-source tranche. Six Polish public documents (five GUS pages and one Sejm ELI regulation) were already captured and preserved under `reports/evidence/B5-raw/`; no seventh source or linked GUS detail file was fetched. A fresh foreground Claude Code Opus pass wrote `01_RESEARCH/evidence/B5-polish-demand-supply-economics.md` at 174 lines with 25 countable official-statistic rows, inherited B1 operator prices separated from official statistics, and explicit unavailable demand, category-specific supply, care-rate, channel, and unit-economics fields. Source-register rows `B5-01`…`B5-06` were appended. Independent manifest/hash/receipt/line/row/title-mismatch checks passed at `reports/evidence/b5-independent-verification-2026-07-30.json`. Prior session-limit/write failures and the first malformed invocation remain preserved; no B2-S retry, niche, legal advice, deployment, DNS/Traefik/Coolify, outreach, user data, or product work occurred.
- **D-022** 2026-07-30 [O] A fresh foreground Claude Code Opus pass rewrote the verified B5 artifact in place at 191 lines, preserving the six-document official-source boundary, 25 D1–D25 rows, inherited B1 price separation, explicit demand/supply/price/channel/unit-economics gaps, and B5-04/B5-05 title-body mismatch receipts. Independent artifact verification passed with all six manifest/receipt/hash checks and the required gap/no-estimate checks at `reports/evidence/b5-independent-verification-2026-07-30.json`; B5 remains the only slice executed.
- **D-023** 2026-07-30 [O] A fresh foreground `claude -p --model opus` pass rewrote the B5 artifact at 176 lines from the six-document local official-source corpus. Independent verification passed: six manifest rows and six receipt/body byte+SHA-256 pairs, 25 contiguous evidence rows with per-row source URLs and receipt paths, corrected B5-03 values, B5-04/B5-05 title/body mismatches, explicit demand/supply/price/channel/unit-economics gaps, and no market-size/conversion estimate, legal advice, or niche recommendation. Source-register B5 references now point to rows 1–25; the local/public B5 artifact and cockpit copies were synchronized after verification. No B2-S retry, niche, legal advice, deployment, DNS/Traefik/Coolify, outreach, user data, or product work occurred.
- **D-024** 2026-07-30 [O] Required public archive verification is blocked: `https://baby.nurt.cloud/downloads/baby-nurt-evidence-pack-2026-07-30.zip` returned HTTP 403 from `nginx/1.31.0` at 2026-07-30T20:52:12Z. The local 20-entry ZIP passes `unzip -t`, and the public B5 markdown and cockpit returned HTTP 200, but the public ZIP contents cannot be verified without a prohibited deployment or server/proxy change. Failure receipt: `reports/evidence/b5-public-archive-403-2026-07-30.md`. Stop this run; do not retry alternate paths or continue to another slice.
- **D-025** 2026-07-31 [O] The fresh foreground `claude -p --model opus` B5 pass first hit the non-interactive permission gate and was preserved at `reports/evidence/B5-opus-pass10-failure-2026-07-31.md`; the corrected `--permission-mode acceptEdits --allowedTools Read,Write,Edit` invocation wrote the 187-line artifact from the existing six-document local corpus. Independent manifest, receipt/body byte+SHA, line, citation, gap, and no-estimate checks passed at `reports/evidence/b5-independent-verification-2026-07-31.json`. The local public pack was synchronized; the existing public ZIP HTTP 403 remains the stop blocker. No new source document, B2-S retry, niche, legal advice, deployment, DNS/Traefik/Coolify, outreach, user data, or product work occurred.
- **D-026** 2026-07-31 [O] After the B5 artifact/control-pack refresh, the local 20-entry ZIP passed `unzip -t`; `https://baby.nurt.cloud/` returned HTTP 200 with the refreshed cockpit, the public B5 markdown returned HTTP 200 and matched the 133-line artifact SHA-256, and the public ZIP returned HTTP 200 with `application/zip`, 20 entries, and a passing archive test. The earlier public HTTP 403 remains preserved as a historical failure receipt, not a current blocker. B5 is complete; no B2-S retry, niche, legal advice, deployment, DNS/Traefik/Coolify, outreach, user data, or product work occurred.
- **D-027** 2026-07-31 [O] A fresh foreground `claude -p --model opus` pass over the unchanged six-document B5 local corpus rewrote `01_RESEARCH/evidence/B5-polish-demand-supply-economics.md` at 133 lines. Independent verification passed: six receipt/body byte+SHA-256 pairs, six source rows, 25 contiguous D1–D25 rows each with a full source URL and receipt path, the official-only boundary, B1 price separation, explicit demand/supply/price/channel/unit-economics gaps, and preserved B5-04/B5-05 title mismatches. A mechanical citation correction replaced data-row URL shorthands; no source document was added. Public pack and cockpit were synchronized afterward; no B2-S retry, niche, legal advice, deployment, DNS/Traefik/Coolify, outreach, user data, or product work occurred.