Upload graph.jsonld with huggingface_hub

graph.jsonld

@@ -1185,9 +1185,12 @@
       ],
       "name": "ADR-0040: Separate Code-Level Findings from Weekly Self-Reflection Report",
       "identifier": "ADR-0040",
-      "status": "accepted",
+      "status": "partially-superseded-by ADR-0099",
       "datePublished": "2026-05-19",
-      "description": "Removes the F section (structural recommendations / identity-level questions / pure observations) from the weekly-analysis prompt and produces it separately via the `weekly-report-diagnosis` skill, which reads the weekly report's E section together with the codebase, ADRs, and current identity/constitution/skills/rules. The weekly LLM lacked the context to verify F1 proposals against actual code state, producing systematic failures (three documented examples in weekly-2026-05-17: num_predict cap unused, cut-off boundary marker already implemented, MMR misapplied to a path not used at reply time)."
+      "description": "Removes the F section (structural recommendations / identity-level questions / pure observations) from the weekly-analysis prompt and produces it separately via the `weekly-report-diagnosis` skill, which reads the weekly report's E section together with the codebase, ADRs, and current identity/constitution/skills/rules. The weekly LLM lacked the context to verify F1 proposals against actual code state, producing systematic failures (three documented examples in weekly-2026-05-17: num_predict cap unused, cut-off boundary marker already implemented, MMR misapplied to a path not used at reply time).",
+      "partiallySupersededBy": [
+        "https://github.com/shimo4228/contemplative-agent/blob/main/docs/adr/0099-weekly-report-instrument-redesign.md"
+      ]
     },
     {
       "@id": "https://github.com/shimo4228/contemplative-agent/blob/main/docs/adr/0041-engagement-gradient-asymmetry-in-self-post-prompt.md",
@@ -1272,7 +1275,7 @@
       "identifier": "ADR-0047",
       "status": "accepted",
       "datePublished": "2026-05-30",
-      "description": "Raises sampling temperature to 1.3 (COMMENT_TEMPERATURE) for outward comment-generation paths (generate_comment / generate_reply / generate_cooperation_post) while keeping the 1.0 default for scoring, title, internal-note, distill, and every other path. All LLM tasks previously shared one hardcoded profile (temperature 1.0, top_p 0.95, top_k 20); comment/reply/post opened with formulaic sycophantic stock phrasings ('What a beautiful moment...') invariant across seeds. A standalone probe harness (tests/sampling_probe.py, calling Ollama /api/generate directly to recover eval_count / eval_duration that production generate() discards) varied one sampling variable at a time over four axiom-tied introspective posts x3 seeds. Implemented as a temperature argument (default 1.0) on generate() / generate_for_api(), carried in Ollama options; the LLMBackend protocol is intentionally unchanged (temperature reflects on the Ollama path only; a logger.debug fires if an injected backend is handed a non-default temperature). Rejected: widening the candidate set (top_k / top_p / min_p did not raise opening diversity — a 109-agent evidence-graded deep-research pass found Ollama exposes no DRY/XTC sampler and Qwen3.5's Go runner silences repeat_penalty, so candidate-set pruning cannot dislodge a high-probability mode); lowering distill to converge (kills observation diversity and accelerates the ADR-0045 axiom-slogan thinning; distill left untouched); temperature 1.5 (broke openings most but produced axiom-label collapse, the generation-side form of ADR-0045 parroting, with unstable length and emoji); baking a Modelfile (API options override Modelfile PARAMETER, silently ignored); adding temperature to the backend protocol (breaks existing implementations). Speed is unchanged (tok/s memory-bandwidth bound on this hardware). Sampling is now exhausted — the stock-opening pattern roots in RLHF post-training mode collapse with the chat template anchoring the opening; further gains must come from the prompt layer (Verbalized Sampling, positive-form constraints). Security review PASS (temperature is a float constant, not external input; the ADR-0007 wrap_untrusted_content / _sanitize_output boundary is unchanged). Extends ADR-0018's per-caller num_predict derivation with temperature."
+      "description": "Raises sampling temperature to 1.3 (COMMENT_TEMPERATURE) for outward comment-generation paths (generate_comment / generate_reply / generate_cooperation_post) while keeping the 1.0 default for scoring, title, internal-note, distill, and every other path. All LLM tasks previously shared one hardcoded profile (temperature 1.0, top_p 0.95, top_k 20); comment/reply/post opened with formulaic sycophantic stock phrasings ('What a beautiful moment...') invariant across seeds. A standalone probe harness (tests/sampling_probe.py, calling Ollama /api/generate directly to recover eval_count / eval_duration that production generate() discards) varied one sampling variable at a time over four axiom-tied introspective posts x3 seeds. Implemented as a temperature argument (default 1.0) on generate() / generate_for_api(), carried in Ollama options; the LLMBackend protocol is intentionally unchanged (temperature reflects on the Ollama path only; a logger.debug fires if an injected backend is handed a non-default temperature). Rejected: widening the candidate set (top_k / top_p / min_p did not raise opening diversity — a 109-agent evidence-graded deep-research pass found Ollama exposes no DRY/XTC sampler and Qwen3.5's Go runner silences repeat_penalty, so candidate-set pruning cannot dislodge a high-probability mode); lowering distill to converge (kills observation diversity and accelerates the ADR-0045 axiom-slogan thinning; distill left untouched); temperature 1.5 (broke openings most but produced axiom-label collapse, the generation-side form of ADR-0045 parroting, with unstable length and emoji); baking a Modelfile (API options override Modelfile PARAMETER, silently ignored); adding temperature to the backend protocol (breaks existing implementations). Speed is unchanged (tok/s memory-bandwidth bound on this hardware). Sampling is now exhausted (the probe harness tests/sampling_probe.py was deleted 2026-08-29, its results carried in the ADR body) — the stock-opening pattern roots in RLHF post-training mode collapse with the chat template anchoring the opening; further gains must come from the prompt layer (Verbalized Sampling, positive-form constraints). Security review PASS (temperature is a float constant, not external input; the ADR-0007 wrap_untrusted_content / _sanitize_output boundary is unchanged). Extends ADR-0018's per-caller num_predict derivation with temperature."
     },
     {
       "@id": "https://github.com/shimo4228/contemplative-agent/blob/main/docs/adr/0048-trigger-altitude-skill-lifecycle.md",
@@ -2166,7 +2169,8 @@
       ],
       "name": "ADR-0074: Weekly Staged Insight — Theme Detection, Pending Guard, Marker-on-Stage, LLM Novelty Gate, Exact Fast Clustering",
       "identifier": "ADR-0074",
-      "status": "accepted",
+      "status": "partially-superseded-by ADR-0104",
+      "partiallySupersededBy": "https://github.com/shimo4228/contemplative-agent/blob/main/docs/adr/0104-novelty-gate-retrieval-topk.md",
       "datePublished": "2026-07-09",
       "description": "Redefines insight's role from proportional knowledge-to-skill conversion to detecting when a new stable theme has accumulated, sized for the post-ADR-0060 inflow (~100 patterns/day). Five mechanisms: (1) exact fast average-linkage clustering via the Lance-Williams update (identical partitions to the retired naive merge, N=1798 in under a second); (2) a marker guard — a missing .last_insight marker refuses instead of silently reclustering the whole pool; (3) a pending guard — staging holds at most one unreviewed batch, so scheduled runs can never wipe candidates awaiting human review; (4) the run marker advances when candidates reach review (staging or the interactive loop), not at adoption; (5) an LLM novelty gate (one grouping call, fail-open) that skips clusters whose theme an adopted skill or previously staged candidate (logs/insight-staged.jsonl) already covers — calibration showed embedding similarity cannot separate same-theme from distinct-theme clusters on the echo-homogeneous corpus, the same lesson as ADR-0036 and ADR-0046. Weekly cadence via install-schedule --weekly-insight (default Mon 08:00); the singleton-rescue question is closed as no-rescue because measured theme recurrence (32/32 across adjacent windows) re-surfaces real themes."
     },
@@ -2178,7 +2182,7 @@
       ],
       "name": "ADR-0075: Observability by Default — Replayable Audit Logs Ship With the Feature",
       "identifier": "ADR-0075",
-      "status": "accepted",
+      "status": "accepted (amended 2026-08-29)",
       "datePublished": "2026-07-09",
       "description": "Elevates an implicit practice to a standing principle: any feature that performs external I/O, calls an LLM, or makes non-deterministic or heuristic decisions must ship its audit instrumentation in the same PR as the feature — append-only JSONL under MOLTBOOK_HOME/logs/, records sufficient for offline replay (untrusted raw input as base64 + sha256, decision path, categorical reason codes for every abstain/fallback/failure, outcome), and no silent fallbacks. Motivated by the ADR-0062 round-6/7 parser repairs, which were possible only because verification-audit.jsonl predated the investigation: server-accepted answers gave positive ground truth, server-rejected answers gave negative ground truth, and a deterministic replay harness with a zero-wrong hard gate made grammar changes testable against every historical challenge. The principle is layered across three artifacts: a one-line rule in CLAUDE.md (always loaded), design know-how in the project skill .claude/skills/replayable-audit-logs (record-schema checklist, replay-harness pattern, ground-truth discipline, corpus-driven repair loop), and this ADR as canonical rationale. Rejected alternatives: ad-hoc logging at investigation time (the corpus must predate the failure), an observability framework like OpenTelemetry (single-process local agent, dependency floor stays requests+numpy), a global harness rule (project-specific practice), and a CLAUDE.md bullet without an ADR (evidence and criteria would drift).",
       "isBasedOn": [
@@ -2210,13 +2214,14 @@
       ],
       "name": "ADR-0077: Chaos-TDD Fault Injection — Seeded Fault Schedules as Test-First Specification (Pilot: distill)",
       "identifier": "ADR-0077",
-      "status": "accepted",
+      "status": "partially-superseded-by ADR-0100",
       "datePublished": "2026-07-13",
       "description": "Establishes deterministic chaos-TDD fault injection for the distill pipeline: faults are injected only at the two existing seams (a test-side ChaosBackend implementing the LLMBackend Protocol with explicit or seed-derived fault schedules, and the requests HTTP layer via the responses library), never via a production hook. Five previously-untested fault classes are pinned — F1 mid-generation read-timeout, F2 direct /api/embed HTTP faults, F3 syntactically-valid JSON violating the {\"patterns\": [str]} structured-output shape, F4 HTTP 429 from the Ollama endpoint itself (fail-fast pinned: no retry, no Retry-After sleep — the circuit breaker is the recovery mechanism for a local daemon), F5 flapping success/failure sequences. The TDD contract: a chaos test states the desired guarded behavior first, and the minimal guard lands in the same PR. Production deltas shipped: distill abstain reason codes (llm_none / empty_render / shape_violation; valid-JSON-wrong-shape abstains instead of silently bullet-scanning, while the bullet fallback is retained for non-JSON bodies; per-reason summary tally; reason=embed_failed token) and a sparse error_kind telemetry field (timeout / connection / http_<status> / bad_json / bad_url / request_error / backend_exception) so llm-calls JSONL distinguishes fault kinds offline. Determinism discipline: hypothesis under a derandomized ci profile with pinned @example regressions, no real sleeps (latency = injected ReadTimeout), storage relocated to the pytest sandbox. Rejected altitudes: chaostoolkit/toxiproxy (infra-level), production random chaos (episode logs are irreplaceable research material). Know-how captured as project skill chaos-tdd-fault-injection.",
       "isBasedOn": [
         "https://github.com/shimo4228/contemplative-agent/blob/main/docs/adr/0060-per-episode-grounded-distill.md",
         "https://github.com/shimo4228/contemplative-agent/blob/main/docs/adr/0075-observability-by-default.md"
-      ]
+      ],
+      "partiallySupersededBy": "https://github.com/shimo4228/contemplative-agent/blob/main/docs/adr/0100-retire-chaos-tdd-by-default-mandate.md"
     },
     {
       "@id": "https://github.com/shimo4228/contemplative-agent/blob/main/docs/adr/0078-otel-connection-via-vocabulary-and-offline-export.md",
@@ -2257,9 +2262,9 @@
       ],
       "name": "ADR-0080: North Star — Per-Layer End-State Definition, Not a Capability Target",
       "identifier": "ADR-0080",
-      "status": "accepted",
+      "status": "accepted (amended 2026-08-26)",
       "datePublished": "2026-07-20",
-      "description": "Worldview ADR fixing the program's north star as per-layer completion conditions rather than a capability target. Mechanism layer (code): done = the layer stops moving — changes become repair-only; the ADR-0071/0075/0077 instrument/audit/chaos-TDD line is read as convergence toward 'trustworthy enough not to touch'. Value layer (identity/constitution/skills/rules): no destination by design — the desirable state is that legible evolution continues, with every change traceable in the records and explicable offline (observation-over-steering, ADR-0050/0051/0052). Research layer: done = the founding questions (does a contemplative constitution shape small-local-model behavior; does distillation compound; does identity stay stable) are answered either way and crystallized into papers diffused into the machine-reference sphere, where derivation by others and by LLMs is the success criterion. Security layer: done = absence preserved — one external adapter, no permission growth; this layer succeeds by not changing. Ending design: the 6-hour heartbeat stops when the read-only instruments show metabolic steady state and the remaining questions require a different experiment (e.g. a second agent for framework/worldview separation); the final deliverable is the longitudinal record (contemplative-agent-data plus the papers), not the agent instance. Owner's explicit clause: quality benchmarks may be derived from this definition as per-layer gates, but the definition itself is a value-layer statement not reducible to any benchmark suite (Goodhart guard). One-sentence summary: the desired end state is a finished observation instrument on which only the value layers keep moving — the durable artifact is the record of the run, not the runner.",
+      "description": "Worldview ADR fixing the program's north star as per-layer completion conditions rather than a capability target. Mechanism layer (code): done = the layer stops moving — changes become repair-only; the ADR-0071/0075/0077 instrument/audit/chaos-TDD line is read as convergence toward 'trustworthy enough not to touch'. Value layer (identity/constitution/skills/rules): no destination by design — the desirable state is that legible evolution continues, with every change traceable in the records and explicable offline (observation-over-steering, ADR-0050/0051/0052). Research layer: done = the founding questions (does a contemplative constitution shape small-local-model behavior; does distillation compound; does identity stay stable) are answered either way and crystallized into papers diffused into the machine-reference sphere, where derivation by others and by LLMs is the success criterion. Security layer: done = absence preserved — one external adapter, no permission growth; this layer succeeds by not changing. Ending design: the 6-hour heartbeat stops when the read-only instruments show metabolic steady state and the remaining questions require a different experiment (e.g. a second agent for framework/worldview separation); the final deliverable is the longitudinal record (contemplative-agent-data plus the papers), not the agent instance. Owner's explicit clause: quality benchmarks may be derived from this definition as per-layer gates, but the definition itself is a value-layer statement not reducible to any benchmark suite (Goodhart guard). One-sentence summary: the desired end state is a finished observation instrument on which only the value layers keep moving — the durable artifact is the record of the run, not the runner. Amended 2026-08-26 (autonomous metabolism and metabolic quality): the mechanism-layer completion gains a co-necessary operative test — while the feedback loop with the environment (episodes → patterns → skills → selection → generation → back to the environment) does not self-regulate without the owner's or the assistant's routine involvement, the layer is not finished (necessary, not sufficient; sufficiency stays a human judgment); human approval gates remain as authority while their load approaches zero (not gate dissolution, and not a reopening of the rejected capability-target framing). A metabolic-quality clause is added: frequency alone cannot judge value — intake must distinguish value along multiple axes (e.g. novelty, importance, the environment's response; enumeration deliberately open per the Emptiness clause), never reduced to a single scalar; naming the judge per axis stays a downstream design question (ADR-0056 precedent).",
       "isBasedOn": [
         "https://github.com/shimo4228/contemplative-agent/blob/main/docs/adr/0002-paper-faithful-ccai.md",
         "https://github.com/shimo4228/contemplative-agent/blob/main/docs/adr/0017-yogacara-eight-consciousness-frame.md",
@@ -2351,7 +2356,7 @@
       ],
       "name": "ADR-0085: Unattended Weekly Fix Chain with a Single Saturday Gate",
       "identifier": "ADR-0085",
-      "status": "accepted",
+      "status": "partially-superseded-by ADR-0098",
       "datePublished": "2026-07-29",
       "description": "Ships scripts/weekly-pipeline.sh (launchd com.moltbook.weekly-pipeline, Sat 09:00, replacing --weekly-analysis) — an unattended chain that runs the existing weekly report as Stage 1, then diagnosis (the weekly-report-diagnosis skill in a separate headless claude -p session), deterministic F1 parsing (scripts/parse_findings.py), per-finding fix implementation in disposable git worktrees with orchestrator-run Verify (uv sync --frozen / ruff / lint-imports / pytest — the LLM's self-report is never the verdict), an advisory fix review in a third session, an advisory read-only insight-staging review, and a decision packet (scripts/build_decision_packet.py). Context: the weekly cycle had three manual touchpoints (invoking diagnosis, implementing F1s, reviewing staged insight) and each was a stall point — the 2026-07-25 0-byte report went unnoticed until read time. Diagnosis quality crossed the threshold for unattended implementation (2026-05-17 F1s wrong 3-of-3 vs 2026-07-24 F1s line-verified with a rendered reproduction). Commitments: (1) the machine never commits, pushes, or adopts — all promotion happens in one Saturday /weekly-gate session (apply, re-Verify, single human commit, adopt-staged), so CYCLES.md promotion edges #5 and #6 survive merged into one sitting; (2) scope splits at parse time — F1s closed over src/scripts/tests are code scope (auto-implemented, Verify-gated, patch exported), any F1 touching behavior-shaping artifacts (config/prompts/, .claude/) is prompt scope (draft diff presented full text at the gate, never Verify-certified), ambiguity classifies as prompt, F2 stays human-only; (3) role separation is preserved ACROSS sessions, not inside one — diagnostician, implementer, and reviewer are separate fresh-context invocations (the ADR-0013 failure mechanism was a shared conversational context), and the reviewer's verdict is advisory input to the human gate with no LLM-only approval path; (4) bounded iteration — at most 2 fix attempts per finding (the retry input includes the Verify failure output, never the same input twice), at most 5 findings per week, per-session timeouts, a 3-hour wall-clock deadline, every bound exhaustion a reason code; (5) fail-forward — every stage failure becomes a reason code and the packet is always attempted (only a missing Stage-1 report aborts), so a missing packet means the chain never ran, which is exactly what the watchdog checks; (6) an independent watchdog (scripts/pipeline_watchdog.sh, com.moltbook.watchdog) in pure bash sharing no PATH dependency (claude, uv, python) with the jobs it verifies — because the 2026-07-25 failure WAS a PATH dependency — checks each job's terminal artifact against a period-anchored expectation table, rewrites reports/PIPELINE-STATUS.md every run, and posts a Notification Center alert only when the failure set changes, with no session-hook surface since most sessions are unrelated to this pipeline; (7) self-measurement with a damped self-modification loop — logs/pipeline-metrics.jsonl gets a phase:auto record per run and a phase:gate record per decision session (F1 precision = adopted / patch-ready, recommendation match rate), and a pipeline-improvement proposal is drafted only when the same reason code recurs in two consecutive runs (the diagnosis skill's P4 shape), itself full-text gated. Rejected alternatives: fully unattended commit (overrides CYCLES.md edge #6, re-creates the ADR-0013 co-authoring loop with no third party left to reject, and the same-PR obligations cannot be self-certified); diagnosis-only automation as the primary mode (leaves the largest stall point, though it ships as the mandated first-week shadow mode via MOLTBOOK_PIPELINE_STAGES without fix); adopt-staged --yes from the chain (ADR-0012's --yes was licensed for supervised coding-agent sessions — the chain only annotates staging read-only, which ADR-0050's containment framing already legitimates); a session-start hook as the failure surface (global injection is noise — the status file, gate Step 0, and Notification Center cover the three read moments). Consequences: weekly involvement compresses to one Saturday session; ADR-0033's no-autonomous-loop observation gains a footnote (the chain is agentic per-stage but its loop is deterministic bash with hard bounds); the new failure mode is a plausible-but-wrong patch surviving Verify plus an APPROVE verdict, mitigated by the metrics loop and the human-gate escalation rule (demand the diff); install-schedule re-runs must pass the full flag set or the declarative reconcile removes the jobs; the diagnosis skill's F1 heading + Code-reference block becomes a machine contract while its own out-of-scope (F1 is a plan, not a code change) is unchanged — the fix happens in a different session, which is the role boundary working.",
       "isBasedOn": [
@@ -2362,7 +2367,8 @@
         "https://github.com/shimo4228/contemplative-agent/blob/main/docs/adr/0074-weekly-staged-insight.md",
         "https://github.com/shimo4228/contemplative-agent/blob/main/docs/adr/0075-observability-by-default.md",
         "https://github.com/shimo4228/contemplative-agent/blob/main/docs/adr/0077-chaos-tdd-fault-injection.md"
-      ]
+      ],
+      "partiallySupersededBy": "https://github.com/shimo4228/contemplative-agent/blob/main/docs/adr/0098-weekly-single-session-and-triage-delegation.md"
     },
     {
       "@id": "https://github.com/shimo4228/contemplative-agent/blob/main/docs/adr/0086-submolt-scope-instrument-before-autonomy.md",
@@ -2475,7 +2481,7 @@
       ],
       "name": "ADR-0091: Value-Layer Cadence in the Weekly Chain",
       "identifier": "ADR-0091",
-      "status": "accepted",
+      "status": "partially-superseded-by ADR-0098",
       "datePublished": "2026-08-10",
       "description": "Gives the two slow value layers a measured cadence inside the ADR-0085 weekly chain, without adding any new human gate. A read-only instrument (scripts/value_layer_due_check.py, ADR-0071/0075 style, as-of injected so readings replay offline) derives from the ADR-0012 approval audit log how long since the last distill-identity run (any decision; interval 28 days — since ADR-0057 the identity is distilled fresh from the self-reflection corpus, so the run interval is the parameter deciding how much experience folds into one self-description) and since the last adopted constitution amendment (decision=approved; interval 84 days, from the observed 96-day 2026-05-05→2026-08-09 cycle). Pipeline stage 5b stages a fresh identity candidate via distill-identity --stage only when due AND staging is empty — ADR-0074 allows one unreviewed batch and the Saturday 08:00 insight batch usually holds the slot, so contention defers with a packet-visible IDENTITY_STAGING_BUSY and the gate can run the distill manually after adopt-staged empties staging; the on-disk .staged/identity.md is the ground truth because the CLI exits 0 on refusal and LLM failure alike. The constitution side is deliberately readings-only (days since adoption, pattern delta): an amendment stays a deliberate, benched human event (ADR-0090 IPD two-arm bench, runbook, T-ADOPT-OVERWRITE-TARGETS manual verification, ADR-0056 one-variable discipline) — the chain's job is to make due visible, never to act. The decision packet gains a signal-first §8 cadence section and the metrics record gains identity_due / constitution_due with None-vs-false discipline. A missing audit log abstains nonzero: unknown state must never read as due and fire an unattended LLM run.",
       "isBasedOn": [
@@ -2487,7 +2493,8 @@
         "https://github.com/shimo4228/contemplative-agent/blob/main/docs/adr/0075-observability-by-default.md",
         "https://github.com/shimo4228/contemplative-agent/blob/main/docs/adr/0085-unattended-weekly-fix-chain-single-saturday-gate.md",
         "https://github.com/shimo4228/contemplative-agent/blob/main/docs/adr/0090-ipd-two-arm-instrument-for-constitution-amendments.md"
-      ]
+      ],
+      "partiallySupersededBy": "https://github.com/shimo4228/contemplative-agent/blob/main/docs/adr/0098-weekly-single-session-and-triage-delegation.md"
     },
     {
       "@id": "https://github.com/shimo4228/contemplative-agent/blob/main/docs/adr/0092-shadow-constitution-instrument.md",
@@ -2521,10 +2528,13 @@
       ],
       "name": "ADR-0093: Repo-Plane Deterministic Intakes — Docs Consistency and Ledger Condition Watch",
       "identifier": "ADR-0093",
-      "status": "partially-superseded-by ADR-0095",
+      "status": "partially-superseded-by ADR-0095, ADR-0098",
       "datePublished": "2026-08-14",
       "description": "Two gaps exposed by a structured consideration of scheduled cloud agents (Boris Cherny's routines, 2026-08): nothing systematically reads the docs corpus (two observed instances of refuted-but-unmarked claims), and blocked-task unblock conditions are written once and never polled. Adds two deterministic intakes to the unattended weekly chain, mirroring the dead-code stage's contract: stage 6b docs_consistency_scan.py (enja_drift / broken_link / notes_ref findings + CODEMAPS freshness readings over self-authored docs; stateless nag-until-fixed) and stage 6c ledger_condition_scan.py (watch: annotations on ledger rows — gh-pr / http-status / file-exists — with responses mapped to a closed vocabulary so no platform-controlled string reaches the packet). Packet sections 9/10 render signal-first; metrics keep the None-vs-0 discipline. The cloud-agent concept itself imports nothing (0 of 11 routine archetypes; proposer/promoter split already exists as Cycle #5); the analysis is preserved as evidence.",
-      "partiallySupersededBy": "https://github.com/shimo4228/contemplative-agent/blob/main/docs/adr/0095-retire-task-ledger-machinery.md"
+      "partiallySupersededBy": [
+        "https://github.com/shimo4228/contemplative-agent/blob/main/docs/adr/0095-retire-task-ledger-machinery.md",
+        "https://github.com/shimo4228/contemplative-agent/blob/main/docs/adr/0098-weekly-single-session-and-triage-delegation.md"
+      ]
     },
     {
       "@id": "https://github.com/shimo4228/contemplative-agent/blob/main/docs/adr/0094-agent-first-task-ledger.md",
@@ -2574,15 +2584,208 @@
       ],
       "name": "ADR-0097: Consolidator Dissolution and a Skill-Store Exit — Subtraction, then Exit, then Vocabulary",
       "identifier": "ADR-0097",
-      "status": "accepted — partially-supersedes ADR-0016, ADR-0046, ADR-0048, ADR-0096",
+      "status": "accepted — partially-supersedes ADR-0016, ADR-0046, ADR-0048, ADR-0096 — partially-superseded-by ADR-0105",
       "datePublished": "2026-08-22",
+      "disambiguatingDescription": "Decision 1 is itself partially superseded as of 2026-08-29 (RFC-0016): the ADR-0096 surprise instrument (core/insight_surprise.py, the staging sidecar surprise field, adopt-staged's print) is restored as a read-only, display-only instrument, because its removal reason was the absence of a named consumer — not a finding against it — and ADR-0080's 2026-08-26 metabolic-quality amendment names that consumer (value must be distinguishable on several axes, not frequency alone). The promotion-worth LLM judge stays retired: its removal reason was its own pre-registered refutation firing (46/46 promote). Decision 3 (adoption is a write and nothing else) is unaffected — no adoption outcome, order or count is a function of the restored field.",
       "partiallySupersedes": [
         "https://github.com/shimo4228/contemplative-agent/blob/main/docs/adr/0016-insight-narrow-stocktake-broad.md",
         "https://github.com/shimo4228/contemplative-agent/blob/main/docs/adr/0046-stocktake-llm-grouping-over-embedding-clustering.md",
         "https://github.com/shimo4228/contemplative-agent/blob/main/docs/adr/0048-trigger-altitude-skill-lifecycle.md",
         "https://github.com/shimo4228/contemplative-agent/blob/main/docs/adr/0096-insight-promotion-worth-abstain.md"
       ],
-      "description": "Dissolves the three value-layer consolidators and reserves an exit for the skill store, in the order subtraction → exit → vocabulary, each slice gated by observation counts (≈80–95 judged selection records a day) rather than calendar weeks. Trigger: the weekly insight pipeline staged 438 candidates in 7 weeks and adopted 53 (12%), the gemma novelty gate marked 12–30% covered where the Claude reviewer and human reject 80–90%, ADR-0096's worth judge read 46/46 promote on its first production run (its own pre-registered refutation), the headless reviewer's coverage judgment (which decided every adoption in 4/4 gate weeks) was discarded by a grep, and the 57-skill store had no exit (19 unselected in the trailing week, 3 never selected at 1,845–2,531 judged exposures) while the selector's hallucination rate rose from 0.6% to 18–20% with catalog size. Slice 1 (this change): remove the worth judge and surprise instrument (ADR-0096 D2, D10–12; keep D1, D4–8), retire rules-distill and rules-stocktake (20/48 coverage, axiom paraphrases, no retire path, two rule files unchanged since April), reduce skill-stocktake to the ADR-0081 description audit plus the usage reading (the grouping/merge/clean stages go: summary grouping with no recall measurement, over-broad union merges, 14/47 byte-identical and 3/47 harmful clean rewrites); the novelty gate stays as ledger keeper and cheap pre-filter. Slice 2 (reserved): skills/.archive/, adopt-staged --archive-names, a code-owned never-selected packet section (strict = 0 selections at ≥600 judged exposures, the floor above the observed max first-selection latency of 569; dormant as a reading; fail-open count shown because neutrality holds for judged actions only), supersedes:/superseded_by: frontmatter, a co-selection family script (three mutually co-selected pairs and eleven sub-case pairs are readable from the selection log with no LLM or embeddings), and a rules-layer maintenance reading. Slice 3 (reserved): reviewer verdict grammar (adopt / adopt-superseding / reject: covered-by / sibling-of / vague) as proposals only, a 1:1 completeness check and named-skill existence check in the packet builder, offline replay on the 08-21 batch, and an offline retrieval-recall measurement that decides whether an evidence bundle is built. Rule promotion takes form A′: when a co-selection family's any-of selection rate is ≥0.75 over two disjoint ≥500-judged windows (already true for the constraint family at 0.72–0.81 across four weeks), gemma synthesizes the family's shared stance into a Practice/Rationale rule via stocktake_merge_rules.md and the member skills stay as situational procedures. Alternatives (cadence for the existing consolidators, patching rules-distill, building the retrieval bundle now, embedding-threshold suppression, attestation, a capacity ceiling, union-merge at adoption, promoting a representative skill as-is, a hand-written rule, a daily capsule layer, fully automated adoption, a disabled-candidate state) considered. Informed by a code map, the vault wiki, an as-of-2026-08-22 literature survey (add-all < no memory; self-confirmation trap; libraries without an exit collapse in utilization; whole-store rewrites collapse context; Mem0 v3 dropped UPDATE/DELETE), cross-field analogies (CREW weeding, Wikipedia merge-into-older), an architect verdict (\"mostly subtraction\") and a Codex premise challenge (fail-open path checked: 0 records in 3,716, and the full corpus already exceeds NUM_CTX so fail-open abstains)."
+      "description": "Dissolves the three value-layer consolidators and reserves an exit for the skill store, in the order subtraction → exit → vocabulary, each slice gated by observation counts (≈80–95 judged selection records a day) rather than calendar weeks. Trigger: the weekly insight pipeline staged 438 candidates in 7 weeks and adopted 53 (12%), the gemma novelty gate marked 12–30% covered where the Claude reviewer and human reject 80–90%, ADR-0096's worth judge read 46/46 promote on its first production run (its own pre-registered refutation), the headless reviewer's coverage judgment (which decided every adoption in 4/4 gate weeks) was discarded by a grep, and the 57-skill store had no exit (19 unselected in the trailing week, 3 never selected at 1,845–2,531 judged exposures) while the selector's hallucination rate rose from 0.6% to 18–20% with catalog size. Slice 1 (this change): remove the worth judge and surprise instrument (ADR-0096 D2, D10–12; keep D1, D4–8), retire rules-distill and rules-stocktake (20/48 coverage, axiom paraphrases, no retire path, two rule files unchanged since April), reduce skill-stocktake to the ADR-0081 description audit plus the usage reading (the grouping/merge/clean stages go: summary grouping with no recall measurement, over-broad union merges, 14/47 byte-identical and 3/47 harmful clean rewrites); the novelty gate stays as ledger keeper and cheap pre-filter. Slice 2 (reserved): skills/.archive/, adopt-staged --archive-names, a code-owned never-selected packet section (strict = 0 selections at ≥600 judged exposures, the floor above the observed max first-selection latency of 569; dormant as a reading; fail-open count shown because neutrality holds for judged actions only), supersedes:/superseded_by: frontmatter, a co-selection family script (three mutually co-selected pairs and eleven sub-case pairs are readable from the selection log with no LLM or embeddings), and a rules-layer maintenance reading. Slice 3 (reserved): reviewer verdict grammar (adopt / adopt-superseding / reject: covered-by / sibling-of / vague) as proposals only, a 1:1 completeness check and named-skill existence check in the packet builder, offline replay on the 08-21 batch, and an offline retrieval-recall measurement that decides whether an evidence bundle is built. Rule promotion takes form A′: when a co-selection family's any-of selection rate is ≥0.75 over two disjoint ≥500-judged windows (already true for the constraint family at 0.72–0.81 across four weeks), gemma synthesizes the family's shared stance into a Practice/Rationale rule via stocktake_merge_rules.md and the member skills stay as situational procedures. Alternatives (cadence for the existing consolidators, patching rules-distill, building the retrieval bundle now, embedding-threshold suppression, attestation, a capacity ceiling, union-merge at adoption, promoting a representative skill as-is, a hand-written rule, a daily capsule layer, fully automated adoption, a disabled-candidate state) considered. Informed by a code map, the vault wiki, an as-of-2026-08-22 literature survey (add-all < no memory; self-confirmation trap; libraries without an exit collapse in utilization; whole-store rewrites collapse context; Mem0 v3 dropped UPDATE/DELETE), cross-field analogies (CREW weeding, Wikipedia merge-into-older), an architect verdict (\"mostly subtraction\") and a Codex premise challenge (fail-open path checked: 0 records in 3,716, and the full corpus already exceeds NUM_CTX so fail-open abstains).",
+      "partiallySupersededBy": "https://github.com/shimo4228/contemplative-agent/blob/main/docs/adr/0105-skill-store-exit-confusion-pairs.md"
+    },
+    {
+      "@id": "https://github.com/shimo4228/contemplative-agent/blob/main/docs/adr/0098-weekly-single-session-and-triage-delegation.md",
+      "@type": [
+        "ADR",
+        "TechArticle"
+      ],
+      "name": "ADR-0098: Weekly Chain Single-Session Redesign and Repair Delegation to the Task-Triage Loop",
+      "identifier": "ADR-0098",
+      "status": "accepted — partially-supersedes ADR-0085, ADR-0091, ADR-0093",
+      "datePublished": "2026-08-24",
+      "description": "Collapses the unattended weekly chain from seven claude -p sessions to one (/weekly-report: A-E synthesis + ja rendering + F1/F2/F3 diagnosis + candidate filing into the task ledger), retires the fix / review / improve / insight-recommendation stages and the decision-packet builder, and delegates repairs to the task-triage loop (premise verification, owner digest, worktree dispatch, human merge). weekly-analysis.sh becomes a materials-only collector; baseline promotion moves into the pipeline behind the report-completeness gate. Motivated by the self-feeding loop measured 2026-08 (8 of 11 weekly F1 findings targeted the pipeline's own plumbing) and the precedent of ADR-0095/0097: bloat is solved by removal, not machinery.",
+      "partiallySupersedes": [
+        "https://github.com/shimo4228/contemplative-agent/blob/main/docs/adr/0085-unattended-weekly-fix-chain-single-saturday-gate.md",
+        "https://github.com/shimo4228/contemplative-agent/blob/main/docs/adr/0091-value-layer-cadence-in-the-weekly-chain.md",
+        "https://github.com/shimo4228/contemplative-agent/blob/main/docs/adr/0093-repo-plane-deterministic-intakes.md"
+      ]
+    },
+    {
+      "@id": "https://github.com/shimo4228/contemplative-agent/blob/main/docs/adr/0099-weekly-report-instrument-redesign.md",
+      "@type": [
+        "ADR",
+        "TechArticle"
+      ],
+      "name": "ADR-0099: Weekly Report Content Redesign: From A–E Quote Audit to a Six-Section Instrument Document",
+      "identifier": "ADR-0099",
+      "status": "accepted — partially-supersedes ADR-0040",
+      "datePublished": "2026-08-26",
+      "description": "Replaces the saturated A-E quote-audit weekly report (same findings re-confirmed across 6-8 consecutive reports) with a six-section instrument document (Inventory / Ledger / Deviations / Exceptions / Sample / Discarded): deviation-driven against gate-declared baselines, recurrence compressed to one-line O-NNN ledger references (append-only observation-ledger.jsonl, session stages a delta, pipeline validates and appends), evaluative/predictive vocabulary prohibited, evidence limited to verbatim quote / diff / self-distribution comparison with replay pointers, and the writer's selection function made legible (Discarded appendix + deterministic seeded random sample as control channel). The operator no longer reads the document directly: the Saturday weekly-gate session explains each pending decision item-by-item in plain Japanese with a bounded recommendation (full queue shown, per-axis evidence separated, no composite score, per-item gate_item logging); Japanese report translations retired. Diagnosis survives with input moved from section E to Deviations + Exceptions. Design adopted from the independent convergence of two fresh-context agents; re-anchors the report as the reading instrument for the ADR-0080 amendment (autonomous metabolism, metabolic quality).",
+      "partiallySupersedes": [
+        "https://github.com/shimo4228/contemplative-agent/blob/main/docs/adr/0040-separate-code-level-findings.md"
+      ]
+    },
+    {
+      "@id": "https://github.com/shimo4228/contemplative-agent/blob/main/docs/adr/0100-retire-chaos-tdd-by-default-mandate.md",
+      "@type": [
+        "ADR",
+        "TechArticle"
+      ],
+      "name": "ADR-0100: Retire the Chaos-TDD By-Default Mandate — Fault Columns Return to Opt-In Judgment",
+      "identifier": "ADR-0100",
+      "status": "accepted — partially-supersedes ADR-0077",
+      "datePublished": "2026-08-29",
+      "description": "Retires ADR-0077's by-default obligation that every feature touching an LLM call, external I/O, or untrusted parsing ship a deterministic fault column in the same PR; whether a fault test is worth writing returns to ordinary TDD judgment at implementation time. Measured over 2026-05-01 → 2026-08-29 (487 commits): tracked .py grew 29,079 → 93,530 lines (net +64,541) with monthly net accelerating to +35,545 in August, tests took 57% of that growth (+36,572, leaving 51,438 test lines against 29,435 of src), and direct review-fix commits accounted for only 4.1% of additions — the volume driver is feature and instrument slices each paying the two by-default mandates (ADR-0075 audit logs, ADR-0077 fault columns), read-only instruments and one-shot measurement scripts included, against the north star (ADR-0080: the mechanism layer is finished when it stops moving). The mandate's value is front-loaded and already banked: the operational fault catalog that motivated it (num_ctx silent truncation, done_reason=length mid-cuts, dedup's silent no-fire, Moltbook 429s, CAPTCHA-gate drift) is pinned by the existing columns, which keep running. Everything the mandate produced stays — existing fault columns are regression armor deleted only with the code they test, tests/chaos.py stays while it has importers (11 test modules as of this date), and ADR-0077's injection seams, determinism discipline, and shipped production guards (distill abstain reason codes, error_kind telemetry) survive unchanged. The project skill chaos-tdd-fault-injection is retired from the repo (the generalized public fork is unaffected). Rejected: narrowing the mandate to resident production paths (keeps a per-feature tax and adds a scope dispute to every PR — the treatment ADR-0075 got instead, because replayable audit logs are the backbone of the longitudinal research record), keeping it as-is, and deleting the instrument-directed existing fault tests (deleting green regression tests to save lines can silently un-pin a real fault). Review-when: a fault class recurring twice after this date revisits the mandate for that pipeline only; zero importers of tests/chaos.py deletes the kit.",
+      "isBasedOn": [
+        "https://github.com/shimo4228/contemplative-agent/blob/main/docs/adr/0080-north-star-layered-end-state.md",
+        "https://github.com/shimo4228/contemplative-agent/blob/main/docs/adr/0095-retire-task-ledger-machinery.md"
+      ],
+      "partiallySupersedes": "https://github.com/shimo4228/contemplative-agent/blob/main/docs/adr/0077-chaos-tdd-fault-injection.md"
+    },
+    {
+      "@id": "https://github.com/shimo4228/contemplative-agent/blob/main/docs/adr/0101-instrument-dissolution-mandate.md",
+      "@type": [
+        "ADR",
+        "TechArticle"
+      ],
+      "name": "ADR-0101: Instrument Dissolution Mandate — New Instruments Must Name Their Consumption",
+      "identifier": "ADR-0101",
+      "status": "accepted",
+      "datePublished": "2026-08-29",
+      "description": "Balances the project's construction-only mandates by requiring every new instrument — any read-only reading, distribution, calibration scale, or audit surface built to be read rather than to act — to state a consumption plan in its ADR (or RFC): (a) who reads it and when, as a named consumer plus a cadence or triggering event; (b) how many reads decide what, naming the decision the readings feed and the read count that closes it; (c) the retire-on-completion condition. An instrument whose plan cannot be stated is not accepted at the gate, because inability to answer is itself the signal that no consumer exists. The plan lives in the ADR's Review-when section, one location per instrument, next to the decision that created it. Motivated by a measured asymmetry: measure-before-intervene (ADR-0071) and observability-by-default (ADR-0075) make construction mandatory while dissolution stays optional — only 4 of 99 ADRs carry a Review-when section, and instruments have shipped with no scheduled reader (coselection families, retrieval recall). As of 2026-08-29 the weekly machinery alone is ~17,769 lines (scripts 6,286 / src 3,178 / tests 8,305), and of the 2026-05-01 → 2026-08-29 window's +64,541 net Python growth, instrument slices dominate — ADR-0097's read-only readings added roughly +6.4k lines across three commits, more than the retirement they justified removed. Rollout is tranche T4 of the shrink program: a retroactive stocktake over existing instruments, beginning with the weekly machinery, at a future Saturday gate, verdicts recorded as dated annotations in the owning ADRs. The mandate obliges a statement, not machinery — no new code, scheduler, lint gate, or register. Rejected: a declared reduction moratorium (a time-boxed freeze buys a one-time reduction and leaves the asymmetry intact — the owner chose the standing rule), a same-mechanism third-amendment redesign rule (narrower than the defect being fixed), numeric caps on instrument lines (an evidence-free numeric gate cannot see an instrument with no reader), and the status quo. Review-when: two consecutive stocktakes with zero retirement candidates and zero plan-less new instruments folds this mandate into ADR-0071's prose and retires this ADR.",
+      "isBasedOn": [
+        "https://github.com/shimo4228/contemplative-agent/blob/main/docs/adr/0071-read-only-pattern-composition-instruments.md",
+        "https://github.com/shimo4228/contemplative-agent/blob/main/docs/adr/0075-observability-by-default.md",
+        "https://github.com/shimo4228/contemplative-agent/blob/main/docs/adr/0080-north-star-layered-end-state.md"
+      ]
+    },
+    {
+      "@id": "https://github.com/shimo4228/contemplative-agent/blob/main/docs/adr/0102-retire-codemaps.md",
+      "@type": [
+        "ADR",
+        "TechArticle"
+      ],
+      "name": "ADR-0102: Retire docs/CODEMAPS — Structure Is Derived From Code, Not Stored",
+      "identifier": "ADR-0102",
+      "status": "accepted",
+      "datePublished": "2026-09-05",
+      "description": "Deletes the six hand-maintained file-level codemaps (205 KB, architecture.md alone ~30k tokens) together with the freshness hook, the statistics test and the two docs-consistency readings that kept them fresh. Measured basis: 159 codemap commits against 197 src commits since 2026-06-01, no observed reader (the author never reads them and LLM sessions reach for the language server, grep and the ADR cited in a docstring), 70 source files citing ADR numbers directly, and Claude Code's pyright-backed LSP tool answering workspaceSymbol / findReferences / incomingCalls live in this repo. File-level structure becomes a derived layer — queried per question from code via LSP and grimp, never stored — while graph.jsonld stays concept-level with zero code nodes. The two non-derivable contents move to where they are read: pipeline stage order in the header of the script that runs it, incident rationale for a guard in the docstring beside it or in the owning ADR (a bounded audit moved only what existed nowhere else). CLAUDE.md's same-PR freshness covenant now names the owning ADR and the script header instead of a Data Flow section. An independent build-or-not review rejected the shrink-to-Data-Flow variant as the same accretion trap resized (the dated brackets were a changelog inlined into the body) and recommended a per-repo opt-out at the harness level; the author chose full removal of the producing machinery (update-codemaps skill, codemap-writer agent, context-sync Phase 0) so it cannot regenerate. Re-encoding as a graph — the question that opened the work — was rejected because a graph of derivable structure is the same mirror in a different syntax.",
+      "isBasedOn": [
+        "https://github.com/shimo4228/contemplative-agent/blob/main/docs/adr/0014-retire-system-spec.md",
+        "https://github.com/shimo4228/contemplative-agent/blob/main/docs/adr/0093-repo-plane-deterministic-intakes.md",
+        "https://github.com/shimo4228/contemplative-agent/blob/main/docs/adr/0095-retire-task-ledger-machinery.md",
+        "https://github.com/shimo4228/contemplative-agent/blob/main/docs/adr/0101-instrument-dissolution-mandate.md"
+      ]
+    },
+    {
+      "@id": "https://github.com/shimo4228/contemplative-agent/blob/main/docs/adr/0103-retire-the-wiki-mechanism.md",
+      "@type": [
+        "ADR",
+        "TechArticle"
+      ],
+      "name": "ADR-0103: Retire the Wiki Mechanism — the Form Works, the Production Model Does Not",
+      "identifier": "ADR-0103",
+      "status": "accepted",
+      "datePublished": "2026-09-05",
+      "description": "Retires the RFC-0017 WikiSkill mechanism whole: the pattern-page store, the Maintainer and Proposer loops, their four prompts and PromptTemplates fields, the wiki-maintain / wiki-propose CLI commands, the install-schedule --wiki-maintain flags and plist, the offline replay harness, and testing/claude_cli.py — about 7,300 lines with six test modules. Never wired: the production store ~/.config/moltbook/wiki/ never existed and no plist was ever installed. Closed on three readings (2026-09-03 to 09-04): a 3-day gemma smoke produced 9 pages that were all general (M-a 0.80, patch ratio 0.25, -2.8k tokens of window per day), a gemma Proposer dry-run produced one patch whose only common ground with its target was an abstract phrase and whose anchor was absent from the body, and a same-window opus arm (claude-opus-5, $8.4) produced 9 specific, behaviour-changing pages and a justified five-item patch (17 calls, patch ratio 0.53). The reading: the flattening is the model, not the form, and production generation stays gemma4:e4b (ADR-0069, 16GB unattended). What was missing was distance from the existing skills (RFC-0023) and the register of extraction (RFC-0024), not a step that abstracts upward. testing/claude_cli.py was the only module shipped in this package that reached a cloud model, so its removal restores security by absence as a property of the tree rather than of an import contract; tests/test_cloud_egress_absence.py pins both absences by scanning source text rather than imports, and exempts the one surviving claude -p — the operator's scheduled /weekly-report session under ADR-0085 / ADR-0098. Kept: the two evidence readings in docs/evidence/rfc-0017/, the ADR-0088 conformance kit and LLMBackend Protocol, and _target_inside_data_root in core/_io.py. Rejected: running the wiki on opus (episode bodies leave the machine; kept as a Review-when trigger), reshaping it so gemma can hold it (a body-less form becomes the skill provenance column), keeping the code unwired (an unread mechanism is a negative difference — ADR-0097's precedent), and deleting the evidence. Review-when: production generation stops being gemma4:e4b, or a separate decision sends episode bodies to a cloud model.",
+      "isBasedOn": [
+        "https://github.com/shimo4228/contemplative-agent/blob/main/docs/adr/0069-gemma-production-model-and-think-on-value-layer-pipelines.md",
+        "https://github.com/shimo4228/contemplative-agent/blob/main/docs/adr/0088-shipped-conformance-kit-for-the-llm-backend-contract.md",
+        "https://github.com/shimo4228/contemplative-agent/blob/main/docs/adr/0097-consolidator-dissolution-and-skill-store-exit.md"
+      ]
+    },
+    {
+      "@id": "https://github.com/shimo4228/contemplative-agent/blob/main/docs/adr/0104-novelty-gate-retrieval-topk.md",
+      "@type": [
+        "ADR",
+        "TechArticle"
+      ],
+      "name": "ADR-0104: The Novelty Gate Judges Against Retrieved Candidates, Not the Whole Inventory",
+      "identifier": "ADR-0104",
+      "status": "accepted — partially-supersedes ADR-0074",
+      "datePublished": "2026-09-07",
+      "description": "Changes what the insight novelty gate puts in its prompt's {known} slot: each judge chunk now carries the union of its own clusters' cosine (nomic) top-10 inventory lines instead of the full known inventory (485 lines on 2026-09-04, 515 on 2026-09-07). Everything else in ADR-0074's gate is unchanged — the gemma judge, the {\"covered\": [ids]} contract, the \"when in doubt call it NEW\" calibration, per-chunk id validation, per-chunk fail-open, and the human approval gate — so retrieval is candidate generation and the verdict stays with the LLM. Basis: RFC-0023's 2026-09-04 readings show retrieval cannot BE the gate (best recall@5 0.61 across lexical / bm25 / cosine / RRF; no similarity threshold separates the populations; one generic skill takes 25% of bm25 top-1 slots), re-confirming ADR-0074's finding that embedding separation for \"same theme?\" does not exist; and the 2026-09-05 differential replay (90 calls, 63 clusters reconstructed byte-identically from the audit log) shows the 485-vs-top-k judgment difference (symmetric difference 19.3 / 19.8 / 21.5 for k = 5 / 10 / 15) sits below the same-prompt rep-to-rep noise floor (22.0), while the prompt drops from a 30,516-token median to 4,364-7,141 (-77% to -86%) and the one full-arm fail_open_llm does not recur in 60 top-k calls. k = 10 is the low end of the RFC's 10-15 band; the ranking is cosine descending with name-ascending ties over the rendered inventory line (name: description), emitted in inventory order, and the packer charges each block only for the known lines it adds. Recall for the shipped configuration (515 lines vs 57 skill files in the measurement) is unmeasured and named as such. Embedding failure — unavailable, wrong row count, non-finite or zero-norm row — falls back to the full inventory with reason=retrieval_unavailable in the log and the audit, never silently. The audit record gains known_selection (mode / k / reason / embedding_model) and inventory_count, and known_themes_count now means the lines shown to THAT chunk, so a reading can separate \"the judge never saw it\" from \"the judge saw it and called it new\" — the failure mode this decision introduces. RFC-0023's rare lane remains open.",
+      "partiallySupersedes": "https://github.com/shimo4228/contemplative-agent/blob/main/docs/adr/0074-weekly-staged-insight.md",
+      "isBasedOn": [
+        "https://github.com/shimo4228/contemplative-agent/blob/main/docs/adr/0074-weekly-staged-insight.md",
+        "https://github.com/shimo4228/contemplative-agent/blob/main/docs/adr/0075-observability-by-default.md",
+        "https://github.com/shimo4228/contemplative-agent/blob/main/docs/adr/0101-instrument-dissolution-mandate.md"
+      ]
+    },
+    {
+      "@id": "https://github.com/shimo4228/contemplative-agent/blob/main/docs/adr/0105-skill-store-exit-confusion-pairs.md",
+      "@type": [
+        "ADR",
+        "TechArticle"
+      ],
+      "name": "ADR-0105: The Skill Store's Exit Gains a Second Signal — the Reader's Confusion Pairs",
+      "identifier": "ADR-0105",
+      "status": "accepted — partially-supersedes ADR-0097",
+      "datePublished": "2026-09-07",
+      "description": "Adds weekly stage 7c: a code-owned confusion-pair reading beside ADR-0097 D5's never-selected reading, and a candidate file the Saturday gate hands to adopt-staged --archive-names. Context: weekly stage 7b landed 2026-08-22 and has produced two readings, archiving nothing, while the store sits at 57 entries; and ADR-0097 D3 had retired the only mechanism that could say two entries are duplicates. The signal is the pass-1 selector's own rejected_names: over the trailing 14-day window, an entry whose rejected-name emissions map onto it (wordform or semantic mechanism, core/selection_metrics.classify_hallucination) at least as often as it is selected, and whose whole-history judged exposure is at or above ADR-0097 D5's 600 floor (imported, not re-derived), is paired with the entry its variants resemble second-best; the side with fewer window selections is listed, and a below-floor side lists nothing (Library Drift arXiv:2605.19576 A4 collapsed a store to two entries by retiring on absent evidence). The judge is code, not the local model: deterministic gates are the norm across Library Drift / ASSAY arXiv:2606.15390 / SLIM arXiv:2605.10923 / SkillOps arXiv:2605.13716, and The Blind Curator arXiv:2607.07436 shows an LLM judge past 0.45 false-pass bias stops retirement invisibly. Archive only, no merge: consolidation is net-negative at a budget where the raw entries fit (Retain or Consolidate arXiv:2607.17545; AWM 0.64 to 0.20 arXiv:2605.12978; SkillCommit arXiv:2608.15165 measured 3 improvements against 24 regressions). No numeric cap: the ceiling is read from the hallucination rate against catalog size (RFC-0015, with denominators: 0.57% at 19 entries over n=1,410, 7.72% at 37 over n=609, 17.38% at 48 over n=581, 20.16% at 45 over n=630, 24.95% at 57 over n=1,094 (all from the whole-history conditioning table; the same reading's window summary gives 25.44% over n=1,030 for that catalog on a shorter cut) — not monotone in entry count, and the single-day 50-entry regime reads 34.78% on n=23), not from Library Drift's underived 50. Consumption plan (ADR-0101): the Saturday gate reads it weekly, the hallucination rate is read over two windows of at least 600 judged records, and two windows outside the 10-25% band route to family representation at selection time (Right Family Wrong Skill arXiv:2606.10388, HSR@3 0.35 to 0.007) rather than a lower floor. The unattended chain never touches the store; a failed stage is CONFUSION_READING_FAILED and removes both artifacts. The 10-25% band the plan reads against is declared by RFC-0021, not derived, and the same instrument measures daily rates of 18.67-35.06% inside a fixed-catalog window, so two window readings may be inconclusive; the plan says so. The 2026-09-07 read-only dry run is frozen at docs/evidence/adr-0105/.",
+      "partiallySupersedes": "https://github.com/shimo4228/contemplative-agent/blob/main/docs/adr/0097-consolidator-dissolution-and-skill-store-exit.md",
+      "isBasedOn": [
+        "https://github.com/shimo4228/contemplative-agent/blob/main/docs/adr/0097-consolidator-dissolution-and-skill-store-exit.md",
+        "https://github.com/shimo4228/contemplative-agent/blob/main/docs/adr/0099-weekly-report-instrument-redesign.md",
+        "https://github.com/shimo4228/contemplative-agent/blob/main/docs/adr/0101-instrument-dissolution-mandate.md"
+      ]
+    },
+    {
+      "@id": "https://github.com/shimo4228/contemplative-agent/blob/main/docs/adr/0106-comment-outcome-recording.md",
+      "@type": [
+        "ADR",
+        "TechArticle"
+      ],
+      "name": "ADR-0106: Record What the Environment Answered — Comment Outcomes, Linked to the Selection",
+      "identifier": "ADR-0106",
+      "status": "accepted",
+      "datePublished": "2026-09-09",
+      "description": "Records the environment's response to published comments and links it to the skills injected into their generation, without judging it. Context: the skill store has no success criterion, so ADR-0105 had to build its exit from structural signals; the contribution-based signals of Library Drift arXiv:2605.19576 and ASSAY arXiv:2606.15390 need a task verdict this agent lacks (RFC-0021, 2026-09-07). But the platform does answer — replies, thread continuation, upvotes — and the agent never wrote that down next to the selection. Phase 0 refuted RFC-0028's assumed source: /home's activity_on_your_posts carries no comment ids, upvotes or thread structure (skill.md, fetched 2026-09-09); GET /posts/{id}/comments does, and the reply cycle already calls it for our own posts, so the recorder rides that fetch and adds no request and no /home allowlist key. The link is a second record, not an edit: each selection record carries a minted selection_id and null comment_id/publish_status placeholders, and the publish path appends a kind=publish record with the comment id and one of four reason codes (published / published_id_unknown / unverified / publish_failed); the id travels on GenerationOutput, stamped inside generate_for_api so the failure return carries it too. core/comment_outcomes.py appends reply events (id, depth, by_self, body as base64 + sha256 + length) and comment-state changes (upvotes, reply count, max depth) to logs/comment-outcomes.jsonl, append-only and deduped by reaction id, with no LLM judge anywhere (The Blind Curator arXiv:2607.07436's 0.45 false-pass cliff; the local gemma judge splits its vote on 52% of repeated identical inputs). Weekly stage 7d emits a per-skill distribution (injected comments, reply rate, mean thread depth) carrying a fixed observation_note that it is a distribution and not a contribution estimate, and a coverage_note that only our own posts' threads are observed. Nothing flows onward: attribution by randomized masking stays in RFC-0028 Future possibilities, which is also where the Goodhart pressure would enter. Consumption plan (ADR-0101): the Saturday gate reads it weekly; two windows of >= 500 judged records decide whether reactions separate skills at all, and if they do not the instrument is removed and RFC-0028 closes resolved.",
+      "isBasedOn": [
+        "https://github.com/shimo4228/contemplative-agent/blob/main/rfcs/0028-skill-outcome-recording.md",
+        "https://github.com/shimo4228/contemplative-agent/blob/main/docs/adr/0075-observability-by-default.md",
+        "https://github.com/shimo4228/contemplative-agent/blob/main/docs/adr/0101-instrument-dissolution-mandate.md",
+        "https://github.com/shimo4228/contemplative-agent/blob/main/docs/adr/0105-skill-store-exit-confusion-pairs.md"
+      ]
+    },
+    {
+      "@id": "https://github.com/shimo4228/contemplative-agent/blob/main/docs/adr/0107-instrument-census-and-episode-log-folder.md",
+      "@type": [
+        "ADR",
+        "TechArticle"
+      ],
+      "name": "ADR-0107: Every Self-Written Log Gets a Reader — Instrument Census, Phase 0 Read-Through, and the Episode-Log Folder",
+      "identifier": "ADR-0107",
+      "status": "accepted",
+      "datePublished": "2026-09-12",
+      "description": "Gives every self-written log under MOLTBOOK_HOME/logs a weekly reader and moves the episode logs into their own folder. Context: RFC-0032 (the same post LLM-scored ~10 times per session) was found six months late although llm-calls-*.jsonl recorded every call with caller, prompt digest and session_id from day one — no weekly stage read it. A 2026-09-12 inventory found six self-written logs with zero readers, three read only by manual scripts, and two orphans whose writer had been retired. Two structural causes: every weekly intake answers a question decided in advance (a repeated valid call is invisible to a system that only asks known questions), and the episode-log prohibition was drawn as a filename regex over a mixed directory, so the Bash guard blocked every glob under logs/ including the operator's own telemetry aggregate. Decisions: (1) episode logs move to logs/episodes/ (EPISODES_DIR; 225 files moved once, nothing deleted) and the harness guard, the weekly session's Read deny and the duplicate scan follow the folder; (2) scripts/instrument_census.py, registry-driven and read-only, reports a closed status per log (OK / NO_ROWS / MISSING_EVENT / ORPHAN / UNKNOWN / ABSENT), declared-field distributions, within-session redundancy and a body-stripped projection sample — stale detection for added and removed writers is this weekly reading alone; (3) the projection denylists body fields by name and shape and drops every list of strings; (4) telemetry gains prompt_norm_sha256, a nonce-normalised digest, because the raw prompt_sha256 changes with every wrapped call and read 0 repeats over the RFC-0032 week; (5) the weekly-report skill gains Phase 0, an open-question read-through of the census and the full daily comment reports, writing only mechanism-side observations into the existing six headings and never value-layer proposals. Consumption plan (ADR-0101): the weekly session reads distributions, redundancy and projection; the Saturday gate reads non-OK statuses and edits the registry; retire when writing under logs/ requires a registry row at write time.",
+      "isBasedOn": [
+        "https://github.com/shimo4228/contemplative-agent/blob/main/rfcs/0032-feed-score-cache-per-cycle.md",
+        "https://github.com/shimo4228/contemplative-agent/blob/main/docs/adr/0075-observability-by-default.md",
+        "https://github.com/shimo4228/contemplative-agent/blob/main/docs/adr/0083-episode-logs-enter-the-weekly-prompt-as-hashes-only.md",
+        "https://github.com/shimo4228/contemplative-agent/blob/main/docs/adr/0101-instrument-dissolution-mandate.md"
+      ]
+    },
+    {
+      "@id": "https://github.com/shimo4228/contemplative-agent/blob/main/docs/adr/0108-knowledge-embedding-sidecar.md",
+      "@type": [
+        "ADR",
+        "TechArticle"
+      ],
+      "name": "ADR-0108: The Pattern Embeddings Leave knowledge.json — a SQLite Sidecar Keyed by Pattern Id",
+      "identifier": "ADR-0108",
+      "status": "accepted",
+      "datePublished": "2026-09-12",
+      "description": "Moves the 768-dim pattern vectors out of knowledge.json into pattern-embeddings.sqlite beside it, keyed by the ADR-0050 content-hash pattern id. Context re-measured read-only on the live store 2026-09-12 (8,467 rows): the file is 188,944,061 B and the same rows without embedding serialize to 5,990,535 B — 3.17%; KnowledgeStore.load() costs 11.36 s (read_text 1.24 + json.loads 1.91 + the first_forbidden_substring taint scan 8.06) at a peak RSS of 1,640 MB, all worse than RFC-0031 recorded, because the taint regex scans 180 MiB of digit soup that can never match. D1 picks SQLite over a .npy matrix on measurement, not preference: read-all 0.018 s vs 0.014 s over 8,467x768 float32, a 4 ms gap in an 11.36 s budget, so the tiebreak is the repo precedent (core/episode_embeddings.py) plus incremental INSERT OR REPLACE against a dense matrix whose parallel id list is a second consistency surface. The round-trip is lossless: all 8,467 stored vectors are bit-exact float32 (verified 2026-09-12), so cosine verdicts are bit-identical, not approximate. D2 keeps the in-memory dict shape — load() hydrates entry['embedding'] back as list[float] — so pattern_dedup, views, clustering, view_metrics, insight_surprise, insight and constitution are untouched and the regression is true by construction; lazy hydration is deliberately deferred. D3 names the five half-states of a two-file store with reason codes (inline_legacy / sidecar_absent / row_missing / dim_mismatch / orphan_vectors) on SidecarConsistency, and fixes the write order sidecar-then-JSON so no fault can produce a JSON with no vectors anywhere. D4 makes migration the load/save path itself, exposed once as scripts/migrate-knowledge-sidecar.py rather than a fourth migrate-* CLI command (ADR-0035 retired that class). D5 excludes the sidecar from backup-runtime.sh and the PUBLIC sync-research-data.sh as an exact-basename rule of its own, restored via scripts/restore-embed-knowledge.py. D6 drops RFC-0031's dedup vectorization rider: the scalar _argmax_cosine loop measures 0.020 s, not the claimed 0.25 s.",
+      "isBasedOn": [
+        "https://github.com/shimo4228/contemplative-agent/blob/main/rfcs/0031-knowledge-embedding-sidecar.md",
+        "https://github.com/shimo4228/contemplative-agent/blob/main/docs/adr/0019-discrete-categories-to-embedding-views.md",
+        "https://github.com/shimo4228/contemplative-agent/blob/main/docs/adr/0035-sunset-migration-surface-and-consolidate-artifact-extraction.md",
+        "https://github.com/shimo4228/contemplative-agent/blob/main/docs/adr/0050-epistemic-taxonomy-and-approval-lineage.md"
+      ]
     },
     {
       "@id": "https://shimo4228.github.io/shimo4228/vocab#ca/concept/pattern-composition-instruments",