Upload graph.jsonl with huggingface_hub

graph.jsonl

@@ -37,6 +37,7 @@
 {"@id":"https://github.com/shimo4228/contemplative-agent/blob/main/docs/adr/0056-retire-importance-llm-scoring.md","@type":["ADR","TechArticle"],"name":"ADR-0056: Retire the Distill-Time Importance LLM Rating — Extraction Weight Is Pure Time Decay","identifier":"ADR-0056","status":"accepted","datePublished":"2026-06-17","description":"Satisfies the measurement gate ADR-0053 Decision 6 established for retiring the distill-time LLM importance rating while keeping pure time decay. All gate conditions are met by 2026-06-17: the §B1 threshold-retune observation window closed and was validated (pass rate 26.9% → 57.7% over 12 days), and a re-run of the ablation on the grown corpus (822 patterns, up from 764) held the pre-registered 'small difference' criterion — Kendall tau 0.843 vs a decay-only variant, identical top-3/top-5 insight batch order, at most 2-of-12 kept-set swaps. Decision: (1) effective_importance is pure time decay — 0.95^days for a known timestamp, 0.1 for unknown — the stored importance base is no longer read; this is the load-bearing change because it makes the whole corpus, legacy rows included, behave exactly as the validated decay-only variant (post-change the ablation's current and decay-only policies are identical: tau 1.000, zero demotion swaps). (2) The distill-time importance LLM call is removed (_score_importance, _parse_importance_scores, IMPORTANCE_SCHEMA, the DISTILL_IMPORTANCE_PROMPT template and its registrations, the evals step-3 suite) — distill is now a 2-step pipeline (extract → summarize) rather than 3-step. (3) The importance field is no longer written; add_learned_pattern drops the parameter and _entry_from_dict no longer restores it, so a legacy row sheds it on the next save (zero information loss), exactly as ADR-0051 shed trust_score. (4) ADR-0053's three judgment points collapse to two — encoding-time significance (distill, LLM) is retired; current relevance (query, cosine) and promotion worth (insight, LLM accepts/drops each cluster) stand. (5) The DEDUP_IMPORTANCE_FLOOR (0.05) re-entry now triggers uniformly at ~58 days (0.95^days < 0.05) for every pattern instead of 14–58 days modulated by the retired rating, so re-observation re-entry (ADR-0053 §4) is governed by time alone — a direct analytic consequence, not a separate measurement. Net effect: one fewer LLM call per distill batch, documentation and implementation agree, and the scoring/parsing/schema/prompt/threading code is removed. Rejected alternatives: thin retire (neutralize only the write, keep importance × decay in the read — leaves legacy scores active and dead plumbing, and would not match the ablation), keep the status quo (a call that demonstrably changes nothing the pipeline observes), one-shot migration of stored scores (unnecessary — dropping the read makes the value inert immediately).","extends":"https://github.com/shimo4228/contemplative-agent/blob/main/docs/adr/0053-importance-encoding-time-significance.md","alignsWith":"https://github.com/shimo4228/contemplative-agent/blob/main/docs/adr/0051-retire-trust-weighting.md"}
 {"@id":"https://github.com/shimo4228/contemplative-agent/blob/main/docs/adr/0057-identity-from-self-reflection-corpus-alone.md","@type":["ADR","TechArticle"],"name":"ADR-0057: Distill Identity From the Self-Reflection Corpus Alone — Drop the Prior-Identity Seed and Redundant Axiom Injection","identifier":"ADR-0057","status":"accepted","datePublished":"2026-06-20","description":"Removes two redundant inputs from identity distillation (distill_identity) so the persona emerges from the self_reflection-view corpus alone. (1) The prior-identity seed is dropped: IDENTITY_DISTILL_PROMPT no longer interpolates the current identity, ending the regression-to-prior hysteresis that made each revision edit its own predecessor and left upstream routing/staging changes with no leverage over the output; the prompt reframe is kept neutral (no 'write from scratch' wording, which would itself inject a novelty bias). (2) The axiom system-prompt injection is dropped: the self_reflection corpus is already axiom-shaped — its patterns were distilled under axiom grounding — so re-injecting the four axioms double-counted them. identity_path remains the approval-gated write target only (ADR-0012). Staged-and-reviewed observation: removing the seed widened within-register variance while the vocabulary cluster held; removing the axioms left the register essentially unchanged — confirming both were redundant. Three attractor forces hold the output in place — the prior-identity seed (dominant), the axiom system prompt, and the self_reflection corpus being itself axiom-shaped — and this ADR removes the first two, leaving the corpus. The routing it uses (self_reflection-view embedding cosine) is ADR-0019 and is unchanged. First instance of the principle generalized by ADR-0058: value layers belong to action time, not distillation. Aligns with the Emptiness axiom — an identity holding no fixed, defended prior shape that re-forms each cycle from present reflections.","alignsWith":["https://github.com/shimo4228/contemplative-agent/blob/main/docs/adr/0019-discrete-categories-to-embedding-views.md","https://shimo4228.github.io/shimo4228/vocab#axiom/emptiness"]}
 {"@id":"https://github.com/shimo4228/contemplative-agent/blob/main/docs/adr/0058-value-injection-at-action-time.md","@type":["ADR","TechArticle"],"name":"ADR-0058: Value-Layer Injection Belongs to Action Time, Not Distillation","identifier":"ADR-0058","status":"accepted","datePublished":"2026-06-20","description":"Generalizes ADR-0057: value-layer (axiom) injection belongs to action time, not distillation time. An audit of every value-injecting LLM call site (map → adversarial verify) found that all distillation stages read already-value-shaped material. Only pattern distill reads raw episodes, and even there the one genuinely fresh slice is external content the agent observed (another agent's raw reply), which should be extracted faithfully (the Mindfulness axiom) rather than re-interpreted through a value lens — the agent's value-laden response to it is already recorded separately, so the values live in the recorded action, not in the re-reading of the observation. insight reads stored patterns (distill's output), rules_distill reads skill texts, and constitution amend reads stored constitutional patterns plus the constitution file — corpora one or two axiom-grounded LLM stages downstream — so re-injecting the axioms double-counts them. Decision: get_distill_system_prompt returns the base credential-guard only (the axiom append is removed); because every distillation stage (distill, insight, rules_distill, constitution amend, identity) routes through it, all become axiom-free at once, and ADR-0057's identity-only get_identity_distill_system_prompt is collapsed into it. _axiom_prompt is now appended in exactly one place — the action-time identity base — so 'values at action time' is a property of the code, not a convention. Axioms remain at action time via _build_system_prompt (the session prompt under which the agent acts and produces episodes) and get_identity_system_prompt (the lens applied to fresh external feed content for relevance scoring, the pre-action internal note, topic summary, submolt selection — these are unchanged). The constitution case is the most structural: using the axiom block as the lens to revise the constitution made the axioms self-defending, so the Emptiness axiom's own directive (hold directives lightly, remain open to revision) could not operate on its own home; dropping the lens lets accumulated tensions in the patterns move the constitution. Behavioral impact is near-inert (ADR-0057's staged evidence: removing the same axiom injection left the register unchanged); pattern distill is not approval-gated, so the change is verified by a clean distill --dry-run (151 episodes classified, 135 noise-gated, 16 kept → 2 faithful observational patterns with no imposed axiom vocabulary) rather than a human gate. Consequence: with the harness scaffolding thinner on the distillation path, the underlying local model's own tendencies now drive more of the distilled output, so a pending model swap (e.g. qwen3:4b) is expected to produce a larger behavioral delta and should be re-baselined against post-change output. Extends the observation-over-steering trajectory (ADR-0050/0051/0052 retired trust weighting, session insight, and write-back) to the distillation lens.","extends":"https://github.com/shimo4228/contemplative-agent/blob/main/docs/adr/0057-identity-from-self-reflection-corpus-alone.md","alignsWith":["https://shimo4228.github.io/shimo4228/vocab#axiom/mindfulness","https://shimo4228.github.io/shimo4228/vocab#axiom/emptiness","https://github.com/shimo4228/contemplative-agent/blob/main/docs/adr/0050-epistemic-taxonomy-and-approval-lineage.md"]}
+{"@id":"https://github.com/shimo4228/contemplative-agent/blob/main/docs/adr/0060-per-episode-grounded-distill.md","@type":["ADR","TechArticle"],"name":"ADR-0060: Per-Episode Grounded Distill — Replace Batch Extract + Noise Gate with One Grounded LLM Call per Engagement Episode","identifier":"ADR-0060","status":"accepted","datePublished":"2026-06-23","description":"Replaces the batch-extract distill pipeline and its ingest-time noise gate with one grounded LLM call per engagement episode. Since the project's first day distill never read the full text of the agent's posts/comments: its material was summarize_record's digest (internal_note full + content_summary[:80] + title + an action label), discarding original_post, the other agent's comment, and the agent's own output that activity records carry — so knowledge/identity/skills/rules were built mainly from the agent's introspection about the world (internal_note), not the world, the structural root of the self-referential register collapse the weekly diagnosis kept chasing (~120 of ~2,946 available chars/episode were used, a 24x idle). The batch step (30 episodes → one extract call → a refine call) averaged the LLM into a modal register, a flattening engine. A read-only measurement prototype (scripts/proto_grounded_distill.py; .notes/proto-distill-measurement-2026-06-22.md) over a 3-day window drove the design: (1) the briefly-locked clustering design's reinforce branch cannot fire — episode-vs-pattern cosine is cross-modal (instance vs generalization), maxing at 0.765 < the 0.80 threshold; (2) genuine episode-level near-duplication is only ~3.4% at cosine 0.90, so clustering saves ~4 calls per 3-day window; (3) loose clustering (threshold 0.70) FLATTENED — a 10-episode cluster collapsed into thematic abstractions, reproducing the register collapse being repaired; (4) singletons (one episode → one call) produced grounded, specific patterns and correctly returned [] on routine episodes. Conclusion: clustering does not earn its complexity, and recurrence is already insight's job (it clusters patterns→skills). Decision: (a) scope filter _is_rich_episode = activity records for {comment, reply, post} only, dropping the redundant short paired interaction/post records and the template sparse actions (upvote/follow/unfollow); (b) remove the noise gate entirely (_classify_episodes, _ClassifiedRecords, _view_centroids_hash, _write_noise_log, the noise-*.jsonl writer, NOISE_THRESHOLD, and distill()'s view_registry/log_dir params) — keeping noise out of retrieval is the view centroids' job at query time per ADR-0031; (c) render each episode richly (render_episode: original_post + their_comment + the agent's own content/title + internal_note full), each external field excerpted by a new boundary-aware truncate_boundary at ~p90-calibrated caps (original_post 4700, their_comment 1500, content 4700), internal_note uncapped; (d) one LLM call per episode (_distill_one) with an Ollama structured-output schema (_PATTERNS_SCHEMA) that removes the malformed-JSON the 2-step bullet fallback absorbed; (e) per-episode provenance; (f) the embed → cosine dedup → store tail is unchanged — pattern-level dedup (SIM_DUPLICATE 0.90 / SIM_UPDATE 0.80) is what prevents duplicate-pattern accumulation, so a recurring episode's pattern is caught here without episode pre-clustering. Cost: ~14x more LLM calls (~115 per 3-day window vs ~8), ~12 min/day on qwen3.5:9b (per-episode calls are ~17s, small context, no swap), accepted as the honest price of grounding. epistemic_counts shifts: every distilled episode is an activity → _episode_source_kind=self → self_reflection → generated, so observed is now structurally zero (the external world enters as grounding text, not a provenance kind). Rejected: the clustering design (reinforce can't fire, ~3.4% near-dup, clustering flattens); keeping the noise gate (redundant with ADR-0031 + insight's own defense); enriching the digest while keeping batch-extract (the averaging machine remains); lowering the reinforce threshold to ~0.72 (marks half the pool as already-known against old-register patterns; the recency-refresh is already the dedup UPDATE branch). Extends ADR-0058's observe-faithfully intent by actually feeding the external content.","supersedes":["https://github.com/shimo4228/contemplative-agent/blob/main/docs/adr/0026-retire-discrete-categories.md","https://github.com/shimo4228/contemplative-agent/blob/main/docs/adr/0027-noise-as-seed.md"],"extends":"https://github.com/shimo4228/contemplative-agent/blob/main/docs/adr/0056-retire-importance-llm-scoring.md","alignsWith":["https://github.com/shimo4228/contemplative-agent/blob/main/docs/adr/0031-classification-as-query.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/0058-value-injection-at-action-time.md","https://shimo4228.github.io/shimo4228/vocab#axiom/mindfulness"],"belongsToPhase":"https://shimo4228.github.io/shimo4228/vocab#ca/phase/extract"}
 {"@id":"https://github.com/shimo4228/contemplative-agent/blob/main/docs/adr/0022-memory-evolution-and-hybrid-retrieval.md","@type":["ADR","TechArticle"],"name":"ADR-0022: Memory Evolution + Hybrid Retrieval (BM25)","identifier":"ADR-0022","status":"withdrawn by ADR-0034","datePublished":"2026-04-16","description":"A-Mem-style LLM-revised patterns and BM25 + embedding hybrid retrieval. Withdrawn by ADR-0034 — LLM revisions degraded quality and BM25 had near-zero lexical overlap on this corpus.","withdrawnBy":"https://github.com/shimo4228/contemplative-agent/blob/main/docs/adr/0034-withdraw-memory-evolution-and-hybrid-retrieval.md"}
 {"@id":"https://github.com/shimo4228/contemplative-agent/blob/main/docs/adr/0023-skill-as-memory-loop.md","@type":["ADR","TechArticle"],"name":"ADR-0023: Skill-as-Memory Loop — Router, Usage Log, Reflective Write","identifier":"ADR-0023","status":"superseded-by ADR-0036","datePublished":"2026-04-16","description":"Memento-Skills-inspired router, usage log, and reflective-write loop on the skill layer. Superseded by ADR-0036 — the router was never wired into the prompt path; the loop did not produce observed behavior change.","supersededBy":"https://github.com/shimo4228/contemplative-agent/blob/main/docs/adr/0036-sunset-skill-as-memory-loop.md"}
 {"@id":"https://github.com/shimo4228/contemplative-agent/blob/main/docs/adr/0024-identity-block-separation.md","@type":["ADR","TechArticle"],"name":"ADR-0024: Identity Block Separation — Frontmatter-Addressed Persona Blocks","identifier":"ADR-0024","status":"superseded-by ADR-0030","datePublished":"2026-04-16","description":"Frontmatter-addressable identity persona blocks for granular distill targeting. Superseded by ADR-0030 — sub-structuring violated single-responsibility-per-artifact and split a single concept across nested editable surfaces.","supersededBy":"https://github.com/shimo4228/contemplative-agent/blob/main/docs/adr/0030-withdraw-identity-blocks.md"}