Batch 3 raw findings — connect-report mining

Sources covered (15 reports, source: frontmatter field):

  1. large-language-model-agents-are-not-always-faithful-self-evolvers.connect.md → kb/sources/large-language-model-agents-are-not-always-faithful-self-evolvers.md
  2. palantir-ontology-vs-decision-traces.connect.md → kb/sources/before-llms-palantir-was-competing-with-snowflake-and-databricks-po-2006384049485484145.md
  3. problem-first-skill-inverts-solution-jumps-...connect.md → kb/sources/problem-first-skill-inverts-solution-jumps-2063186118409929161.md
  4. prov-overview.connect.md → kb/sources/prov-overview.md
  5. reading-between-the-dots-decoding-hidden-computation.connect.md → kb/sources/reading-between-the-dots-decoding-hidden-computation.md
  6. researchers-asked-llms-strategic-advice-trendslop.connect.md → kb/sources/researchers-asked-llms-strategic-advice-trendslop.md
  7. shi-agenti-chotiri-navichki...en.connect.md → kb/sources/shi-agenti-chotiri-navichki-vazhlivishi-za-dobriy-prompt.en.md
  8. skillopt-executive-strategy-self-evolving-agent-skills.connect.md → kb/sources/skillopt-executive-strategy-self-evolving-agent-skills.md
  9. skillrl-evolving-agents-recursive-skill-augmented-rl.connect.md → kb/sources/skillrl-evolving-agents-recursive-skill-augmented-rl.md
  10. the-best-model-routing-is-task-specific-...connect.md → kb/sources/the-best-model-routing-is-task-specific-2077537847951945742.md
  11. the-github-for-context-doesn-t-exist-yet-...connect.md → kb/sources/the-github-for-context-doesn-t-exist-yet-2077772169455530152.md
  12. towards-causal-representation-learning.connect.md → kb/sources/towards-causal-representation-learning.md
  13. verbalizable-representations-global-workspace-llms.connect.md → kb/sources/verbalizable-representations-global-workspace-llms.md
  14. we-should-take-text-optimization-more-seriously-...connect.md → kb/sources/we-should-take-text-optimization-more-seriously-2064027464926716154.md
  15. run-full-improvement-pass-on-note.connect.md → kb/work/agent-note-improvement/run-full-improvement-pass-on-note.md

1. Synthesis opportunities carried over

  • (1) Faithful self-evolvers: distillation.md + activate-behavior-changing-memory.md + evaluate-memory-by-effects.md + knowledge-storage-does-not-imply-contextual-activation.md converge on "read-back is not activation, condensation is a systematic activation-killer." Well-formed, near note-ready.
  • (2) Palantir: trace-derived-extraction + lineage + raw-accumulation + symbolic-context-engineering imply "trace-first ontologies are derived views, not starting schemas," plus a platform-first-vs-trace-first cost tradeoff. Well-formed, two distinct claims.
  • (3) Problem-first skill: "skill value often comes from forcing skipped reasoning moves under pressure, not automating the domain answer." Vague — author explicitly says it may only become a note "if more practitioner skill examples accumulate."
  • (4) PROV-overview: PROV + in-toto + build-systems-à-la-carte converge as three external provenance/lineage/build-freshness standards that could ground lineage.md and distilled-artifacts-need-source-tracking.md. Well-formed (named source-list), flagged explicitly as "do not create during connect."
  • (5) Reading between the dots: this paper + oracle-strength-spectrum + J-space ingest imply "monitorability is relative to an observation surface." Well-formed, narrow, with an explicit non-adversarial caveat attached.
  • (6) Trendslop (strategic advice): two opportunities — "LLM advice follows discourse priors before task logic" (source + content-effects + structure-activates-training-distributions + first-principles-reach) and "prompt movement is not bias removal" (option-order finding + PromptSE). Both well-formed.
  • (7) Shi agenti (AI Fluency): three opportunities — "operator fluency as human-side complement to context engineering" (vague, mapping exercise); "delegation crosses the operator/orchestration boundary" tying source to frontloading.md, bounded-context-orchestration-model.md, feasibility-heaviest-forks-net-load.md (fairly well-formed); "intent engineering needs a local definition" (vocabulary gap, not a connection).
  • (8) SkillOpt: "validation-gated skill optimization is readable-artifact learning under a hard-oracle boundary"; "skill provenance now has ≥3 routes (methodology distillation, trace-induction, validation-gated optimization)"; "rejected edits are retained learning artifacts." Well-formed, converges with SkillRL (see §2).
  • (9) SkillRL: names the explicit higher-order claim "skill banks can couple readable-artifact and policy-weight learning loops," spanning SkillRL/SkillOpt/AgeMem/AWM/SkillWeaver/SAGE/TIMG. Well-formed, this is the sharpest synthesis candidate in the batch.
  • (10) Best model routing: "task-specific routing earns value only when a workflow has a learned input taxonomy and a verifier strong enough to define cheap-enough/good-enough per subtask." Well-formed.
  • (11) GitHub for context: "reusable context should outlive the agent runtime when reuse value and replacement cost exceed the consuming agent's; portability is a sovereignty property, not an integration convenience." Well-formed.
  • (12) Towards causal representation learning: none beyond the already-created reach-assessment note.
  • (13) Verbalizable representations (J-space): "externalized reasoning trades internal workspace demand for context budget" (mechanistic, well-formed); "probeability of distributed-parametric state is weaker than artifact inspectability" (caveat to the readable/opaque contrast, well-formed).
  • (14) Text optimization manifesto: no new synthesis — three pillars map onto existing notes 1:1 — but flags an unnamed candidate claim, "update-time compute is a scaling axis symmetric to inference-time compute," demoted to a maintenance observation rather than promoted.
  • (15) Run-full-improvement-pass (workshop): none new — already captured by the existing error-correction-decorrelated-checks ↔ synthesis-is-not-error-correction cross-link.

2. Recurring cross-report themes

  • treat-continual-learning-as-substrate-coevolution.md — hit in 4/15 reports (SkillOpt, SkillRL, verbalizable-representations, text-optimization manifesto), all evidence. It's the hub note for "prose artifact + weight update coevolve"; every skill/text-layer-learning source lands here. Strong candidate for promotion off seedling if not already current.
  • deploy-time-learning-is-the-missing-middle.md — hit in 4/15 (SkillOpt, SkillRL, GitHub-for-context, text-optimization). Second hub note in the same cluster.
  • continual-learning-open-problem-is-behaviour-not-knowledge.md — hit in 4/15 (faithful-self-evolvers, SkillOpt, SkillRL, text-optimization).
  • diagnostic-richness-constrains-outer-loop-learning-quality.md and agent-memory-requirements/use-trace-derived-extraction.md — each hit in 3/15 (reading-between-the-dots/SkillOpt/SkillRL; and Palantir/SkillOpt/SkillRL respectively).
  • context-efficiency-is-the-central-design-concern-in-agent-systems.md — hit in 3/15 (shi-agenti, best-model-routing, SkillRL).
  • Skill-evolution cluster convergence (as flagged in the task): faithful-self-evolvers, SkillOpt, and SkillRL do converge on one shared arc — condensed/optimized experience risks being behaviorally inert (faithful-self-evolvers) unless the update loop couples the artifact to policy training or validation (SkillOpt's held-out gate, SkillRL's GRPO). SkillRL's own report states this explicitly: "SkillRL's cold-start SFT plus GRPO makes skill use part of the policy-learning loop, potentially addressing the activation gap." This is a real, already-partly-authored resolution chain, not just topical adjacency — the KB gap is that no single note yet states "activation gap → readable-artifact-loop closes it only when coupled to a validation or policy signal" as its own claim.
  • problem-first-skill is a different "skills" sense than the RL/skill-bank cluster above — and this collision recurs explicitly: skills-are-instructions-plus-routing-and-execution-policy.md is accepted as evidence in problem-first-skill's report but rejected in both SkillOpt's and SkillRL's reports with near-identical language ("title collision on 'skills'"). Likewise skills-derive-from-methodology.md is accepted (problem-first, shi-agenti) and rejected as a "contrasting provenance route" (SkillOpt, SkillRL). Four reports touch this pair of notes with a split accept/reject pattern driven purely by which sense of "skill" is in play (harness-invoked procedure vs. RL/optimization-trained agent skill).
  • skills-are-instructions-plus-routing-and-execution-policy.md total: hit in 4/15 reports (problem-first accept, shi-agenti see-also, SkillOpt reject, SkillRL reject).
  • kb/notes/definitions/context-engineering.md — rejected in 3/15 reports (faithful-self-evolvers, Palantir, verbalizable-representations), each time as "too broad" / explicitly out-of-scope. This looks like healthy behavior of a deliberately narrow current definition resisting keyword-driven edges, not a gap.
  • Agent Workflow Memory (AWM) review/ingest — hit in 3/15 (faithful-self-evolvers off-authorisation, SkillOpt compares-with, SkillRL compares-with/evidence) — a recurring comparison anchor for trace-to-procedure systems.
  • oracle-strength-spectrum.md — hit in 3/15 (reading-between-the-dots evidence, best-model-routing evidence, verbalizable-representations rejected-as-weak).

3. Systemic/maintenance issues

  • Missing .ingest.md companions (recurring, 3/15 reports: faithful-self-evolvers, prov-overview, text-optimization manifesto). Each of these sources is already cited as load-bearing evidence by multiple library notes, yet has no authored outbound surface — connect can only produce reverse-edge candidates, deferring real edge-authoring indefinitely. This is a backlog, not a one-off.
  • "Split candidate" flag (recurring, 4/15: SkillRL, verbalizable-representations, text-optimization manifesto, GitHub-for-context) — each source makes 3-4 independent claims and the report explicitly warns against promoting/citing them all onto one note. The discipline is working, but it signals that recent captures (papers, manifestos) routinely need multi-claim triage before ingest.
  • Index staleness: researchers-asked-llms-strategic-advice-trendslop flags that its own snapshot was not yet present in kb/sources/dir-index.md during the discovery pass — a generated index lagging a fresh capture.
  • Cross-collection authorization gap: faithful-self-evolvers explicitly defers two strong edges (to Dynamic Cheatsheet and Agent Workflow Memory reviews) because agent-memory-systems/COLLECTION.md wasn't loaded in that connect run, routing them to Off-authorisation Candidates. This is a structural connect-workflow limitation (single-collection-scoped runs strand valuable cross-collection edges) rather than a one-off oversight — worth checking whether it recurs when connect is run on sources whose strongest edges land outside kb/notes/.
  • Citation drift: faithful-self-evolvers notes three notes cite the source via a bare arxiv.org/html/2601.22436v2 URL while the snapshot is now v3 — local-snapshot links vs. external URLs falling out of sync.
  • Long filename blocks ingest slug: palantir-ontology flags that the captured snapshot's basename is too long to reuse for a validator-compliant .ingest.md filename — a naming-convention friction for long social-media-style capture slugs.
  • Recurring vocabulary-gap flags (not the same term, same shape): "intent engineering" (shi-agenti), "update-time compute" (text-optimization), "text-layer extended-cognition lineage / Hutchins-Clark-Chalmers-Bush" (text-optimization). Three instances of connect surfacing an externally-named concept with no KB home, each correctly deferred rather than forced into an edge.
  • Bilingual source pair: shi-agenti flags that both the Ukrainian original and English translation snapshots exist; downstream ingest should standardize on the English translation for corpus consistency while keeping the Ukrainian as provenance.
  • Workshop source with no frontmatter: run-full-improvement-pass-on-note's report is marked "provisional" because the source text file has no type/tags/description, forcing description-level-only discovery; flags re-running connect after promotion adds real frontmatter.

4. Candidate kb/log.md entries

  • - SYNTHESIS: [skillrl-evolving-agents-recursive-skill-augmented-rl source, skillopt-executive-strategy-self-evolving-agent-skills source, large-language-model-agents-are-not-always-faithful-self-evolvers source, agent-workflow-memory, SkillWeaver, amazon-science--SAGE, trajectory-informed-memory-generation, agentic-memory-learning-unified reviews]: skill-evolution sources now converge on an unnamed claim — condensed/optimized experience (SkillOpt, distillation) risks being behaviorally inert unless the update loop couples the artifact to a validation gate or policy-training signal (SkillRL's SFT+GRPO), i.e. "activation gap closes only under coupled validation/policy pressure." treat-continual-learning-as-substrate-coevolution.md and deploy-time-learning-is-the-missing-middle.md are the two hub notes every one of these sources lands on (4/15 reports each in this batch alone) but neither states the closure condition as its own claim.
  • - ABSTRACTION: [skills-are-instructions-plus-routing-and-execution-policy.md, skills-derive-from-methodology.md, problem-first-skill-inverts-solution-jumps connect, skillopt/skillrl connect reports]: "skill" is overloaded across the KB between harness-invoked procedural packaging (Claude Code / problem-first sense) and RL/optimization-trained agent behavioral memory (SkillOpt/SkillRL sense); both target notes get accepted for one sense and explicitly rejected for the other in the same batch of connect reports, with reports independently reaching for the phrase "title collision."
  • - FIX: [large-language-model-agents-are-not-always-faithful-self-evolvers source, prov-overview source, we-should-take-text-optimization-more-seriously source]: three cited-but-unpromoted sources have no .ingest.md companion despite being load-bearing evidence for multiple current notes; connect on each can only emit reverse-edge candidates. Backlog, not isolated — run cp-skill-ingest on all three.
  • - FIX: [large-language-model-agents-are-not-always-faithful-self-evolvers connect]: distillation.md, evaluate-memory-by-effects.md, and activate-behavior-changing-memory.md all cite this source via the external arxiv.org/html/2601.22436v2 URL (now stale, source is v3) instead of the local snapshot path; standardize on local links to keep v2/v3 drift in one place.
  • - FIX: [researchers-asked-llms-strategic-advice-trendslop connect]: fresh source snapshot was absent from kb/sources/dir-index.md during its own connect discovery pass — index generation lagging capture.
  • - ABSTRACTION: [large-language-model-agents-are-not-always-faithful-self-evolvers connect off-authorisation section]: connect runs are single-collection-scoped by default, so the strongest edges for a source (here, to agent-memory-systems reviews of Dynamic Cheatsheet and Agent Workflow Memory) get stranded in Off-authorisation Candidates whenever the best target lives outside the collection whose rules were loaded. Worth checking whether this recurs enough to justify a routing convention (e.g. running connect a second time scoped to the cross-collection target).