Improvement log

Append one line per observation under Open observations. During traversal, don’t fix anything — just record it. Explicit maintenance may reconcile entries; keep closed episodes under Historical observations with the evidence for closure.

Format: - CATEGORY: path/or/subject: observation

Use a short uppercase category prefix that names why the item belongs in the log, such as ABSTRACTION, SYNTHESIS, DUPLICATION, CONTRADICTION, or FIX. Use a file path when one artifact is the clear target; otherwise use a bracketed subject list.

Open observations

  • FIX: [kb/theory/a-derived-copy-of-recomputable-truth-must-be-checked-or-absent.md] — its duplicated-build-artifact example still describes two tracked AGENTS.md.template copies, although only the root copy remains tracked. Update the example's status without treating the realised general theory as unwritten; no theory-note edit was made during this log reconciliation.

  • SYNTHESIS: [ROME weight edit and SAE feature clamp, 2026-11-03] — both locate intervention points and demonstrate behavioral influence. Locating a candidate does not by itself establish that changing it preserves required invariants. Even weak interactions can violate an invariant; the Golden Gate result establishes neither coupling strength nor, without a stated requirement, an invariant violation. The SAE localizes concepts in activations under an instrumented path, not in the weights. Untested experiment: state the intended effect and required invariants, clamp a feature only under a stated condition, and measure effects relevant to those invariants inside and outside that condition. Sources: kb/sources/rome-locating-and-editing-factual-associations-in-gpt.ingest.md, kb/sources/scaling-monosemanticity-claude-3-sonnet.ingest.md.

  • QUESTION: [natural-language meaning and formal semantics, 2026-11-02] — the authority distinction is already developed in kb/theory/symbolic-form-holds-consumers-to-a-semantic-reference.md: an account of English answers to linguistic practice, while an adopted model has authority within the adopting system; change through use does not by itself preclude a formal account. Only the mechanism question remains deferred: when do repair, change through use, uptake of redefinition or vague boundaries maintain a particular convention? Keep conventional meaning, author intent and local agreement distinct; repair is not established as the usual mechanism. Reopen only when a note needs that explanation.

  • QUESTION: [kb/work/open-ended-work-economics, codify-versus-llm-decision-heuristics.md "Stake and odds"] — whether routines are adopted and relaxed because evidence shows the extracted pattern is or is not a spec (the epistemic reading, under which theory building is the shared layer of open-ended knowledge work) or because a cheaper or better method arrived (the resource-rational reading, Lieder and Griffiths, not ingested). Deferred 2026-10-30 on Astra's objection that relaxing triggers do not discriminate: discovering a mistake changes costs, and discovering a cheaper alternative requires evidence. No extraction depends on it; a discriminating test is not yet designed.

  • GAP: [kb/instructions/cp-skill-snapshot-web] — no PostScript route; the Russell–Wefald ingest (2026-10-30) converted the author's .ps copy with ps2pdf -dSAFER before pdftotext, recorded as capture: ps2pdf+pdftotext. Ligatures and equations degrade. Add the route if a second PostScript source arrives.
  • METHOD: [kb/articles/conjectural-learning-with-fixed-models.md, workshop lead-article-learning-first (deleted 2026-09-22)] — the workshop's 3,900-word front-half draft was superseded by a 600-word proposal the operator wrote outside the KB, which became the lead on 2026-09-22 (9670ede7); the operator judged the workshop a failure. Observed difference: the draft opened with a case and a definition walk-through and reached the bet late, while the proposal led with the bet and stated each claim once; the draft's length came partly from added reasons and hedges. Two claims from the workshop are now in no live article and are recoverable only from git history at the workshop path: the second bet (the loop is easier when the builder also writes code; supported by a-fixed-model-house-must-retain-missing-procedures-for-theory-use.md) and the unsupported candidate that a prose methodology improves as base models improve. Watch whether operator-written compact proposals outperform agent-drafted restructurings on other articles before drawing a rule.
  • ABSTRACTION: [agents-md-should-be-organized-as-a-control-plane, context-engineering, why-ai-systems-dont-learn-and-what-to-do-about-it source] share unnamed structure: control-plane abstractions recur at repository, prompt, and learning-architecture levels, but the transfer conditions between those levels are not yet named
  • SYNTHESIS: [three-space-agent-memory-maps-to-tulving-taxonomy (content-type axis), multi-agent-memory-computer-architecture-perspective source (hierarchy-level axis)] — two independent decompositions of agent memory from different traditions (cognitive science vs computer architecture) that together predict a two-axis taxonomy: content type x hierarchy level
  • SYNTHESIS: [superarc source (recursive compression), esolang-bench source (code generation OOD), pathway-sudoku source (constraint satisfaction), ebrahimi induction-bias source (state tracking)] — four sources now converge on "LLMs score near zero on well-specified problems that strip training-distribution shortcuts"; SuperARC adds a fourth independent domain with the strongest methodology (AIT-grounded, mathematically proven equivalence between compression and prediction)
  • ABSTRACTION: [context-efficiency-is-the-central-design-concern-in-agent-systems.md, tool-loop-index.md, process-structure-and-output-structure-are-independent-levers.md, kb/agent-memory-systems/agentic-memory-systems-comparative-review.md] share an unnamed claim that filesystem-native tool use is a high-reach long-context strategy because it externalizes semantic overflow into executable process structure
  • SYNTHESIS: [Huxley-Godel Machine connect, diagnostic-richness/oracle-strength cluster]: outer-loop harness search may need archive-objective alignment in addition to oracle strength and diagnostic richness; HGM frames the model as oracle strength, diagnostic access quality, and archive-selection objective
  • SYNTHESIS: [LLM memory connect, agent-memory-requirements/evaluate-memory-by-effects.md]: memory evaluation may have a ground-truth horizon, degrading from direct recall toward arc change, supersession, temporal correctness, and significance recognition
  • ABSTRACTION: [Stash connect, service-owned memory backends]: service-owned memory backends may need typed-role analysis by storage substrate, extraction schema, validation gate, activation mechanism, and typed operational objects
  • ABSTRACTION: [Stash connect, hypothesis records]: hypothesis records may be a distinct memory lifecycle; the Stash review captures hypothesis lifecycle as a borrowable idea, but the KB lacks a general workshop/library treatment
  • GAP: [kb/reference/proposals/checked-inline-blocks-for-shared-instruction-text.md] — the frontloaded-contract drift problem now has a general theory in kb/theory/a-derived-copy-of-recomputable-truth-must-be-checked-or-absent.md and a concrete unadopted design for marked literal blocks, deterministic comparison and adoption measurements. The remaining question is whether to adopt and implement that design for a live consumer, not how to formulate it. The 2026-06-10 skills chose live contract reads; the former two-copy AGENTS.md.template equality check has no current target. Retain live reads until a checked-copy path is adopted.
  • GAP: [process verification / process reward, external inquiring-line] — the generalize-versus-preserve question is already addressed by kb/theory/abstract-an-experience-only-when-you-can-state-the-boundary.md and kb/theory/checked-outcome-licenses-episode-retention-not-abstraction.md: a checked outcome does not establish an abstraction's transfer scope. The original workshop closed without promotion. What remains is the specific process-reward result cited as 32%→87% in the external discussion: the existing treatment does not establish a dedicated grounded account of that result. Reopen when a concrete verification question needs it; do not write another general abstraction discriminator.
  • GAP: [Claude Code dynamic workflows, structured-output-codec-for-review-protocol.md] — the 2026-06-12 run's argument-passing failure and shell-free script sandbox have not been retested; its per-run token attribution was absent. Current review finalization can store supplied telemetry and check a supplied model against its partition, but cannot by that check authenticate the executing model or derive missing per-run usage. The structured-output proposal remains undecided; reassess its adoption trigger against a review medium actually in use, rather than treating the historical agent({schema}) observation as a current capability test. Completed API-side reconciliation is retained under Historical observations.
  • FIX: [constraining cluster, llm-output-deviation-requires-three-way-diagnosis.md] — the original observation found four notes using the two-phenomena model as their deviation foundation. The checkpoint and debugging-triage notes, their immediate citing passages, and the triage navigation were reconciled on 2026-09-18; the wider two-source framing still needs a scoped audit against the three-way diagnosis. Do not treat the original four-note count as current.
  • FIX: [full-pass 20260821T115349Z-a7c2, bounded-context-orchestration-model.md] — the pass retracted the note's universal head ("the model captures the full space of such architectures" became a scoped conditional) but initially reconciled zero citing files, despite 323 historical mentions and no citer-reconciliation record in the pass reports. The follow-up sweep audited 75 files holding 119 direct Markdown links and reconciled 26 live artifacts: five importing universality or unsupported effectiveness, then twenty-one importing retired P/K + r/scalar-M notation, unbounded-state language, or prescriptive force. The latter include the select/call lemma, tool-loop and context-engineering definitions, context- and scheduler-facing notes, relevant source ingests, and active workshops. The remaining live citers use the retained conditional normal form or its components. The dated Cordis original and premise-cohort capture retain the old language as historical evidence. Fourteen Open items originally stood unactioned under the one-cycle protocol. The most consequential, a GLOBAL closing defeat showing that the singular loop serializes an independent parallel batch and doubles barrier latency, was cleared by operator direction on 2026-08-22: select(K) now emits a finite nonempty independent batch B, call_all(B) preserves concurrent execution through an explicit barrier, and transition(K, B, R) incorporates the aligned results; singleton batches recover sequential workflows. The conversion lemma preserves call specifications, batch membership, and barrier order and is now titled for the barrier-delimited class it proves. Eighteen live dependents were reconciled while the closing report and dated captures retain the defeated form as historical evidence. Thirteen Open items remain
  • QUESTION: [kb/theory/artifacts-must-preserve-named-choice-scope.md] — can a check cheaper than semantic review detect missing reference/range or choice-role context at an artifact's intended consumption boundary? The current note supplies the scope requirement and a bounded opening-surface audit, not a verdict on the whole corpus. The earlier quantifier-based publishability/portability conjecture is retained as superseded under Historical observations; do not revive it as the required form.
  • METHOD: [proposal current-state sections] — the repaired consumed-cost status in Exceeding a review budget splits the task illustrates how duplicated implementation status becomes stale. Whether proposal current-state sections should point to commits or symbols rather than restate status remains a design question, not an outstanding defect in that archived proposal; the specific incident is retained under Historical observations.
  • METHOD: [kb/work/operator-led-article-clarification] — the original nineteen-commit clarification session suggested that reusable prose rules need worked exemplars, not rule text alone. The workshop now records later bootstrap and unit-of-training article runs, including application of fourteen of seventeen proposed candidates; its evidence is no longer limited to one session. Those runs do not establish the stated closing condition: the bootstrap run lacked phase 1 and individual acceptance records, while closure requires two further articles through both phases with taxonomy revision. The exemplar-slot hypothesis and the decision not to change the training article yet remain; record further evidence against that specific condition rather than treating these runs as completed replication.
  • GAP: kb/articles/an-automated-software-house-as-a-second-test-of-a-theory-builder.md: the four witness conditions (holding and application, coherent revision, automated continuation, practical reliability) are defined only in this article; no note develops them and no survey assesses systems against them, so the article can say only that no reviewed system is shown to be an automated software house. Write the note when the software-house line resumes (theory-builder-from-popper workshop, closed 2026-09-25).
  • SIZE: kb/theory/definitions/theory-builder.md: 2,387 whitespace-delimited words at this check, compared with the approximately 2,200 recorded on 2026-09-25. The earlier decision left it intact because moving Qualifiers or Boundary cases would affect about 25 anchor links (#qualifiers, #boundary, #boundary-cases); that link count has not been rechecked. The growth warrants reassessing the definition's cost, not an automatic split; any split first needs a current anchor-link audit.
  • METHOD: [analyse-agentic-system, Dynamic Cheatsheet runs AAS-2026-10-09-dynamic-cheatsheet-2529cdc2795b-01 (artifact-run engine) and df264b32e488 (previous workflow), both gpt-6-luna/medium at source 5cfe3c37] — one same-model comparison (2026-10-09). The new run fixed the old run's main runtime error: a missing </cheatsheet> admits the rest of the response (extractor.py splits on the closing tag), which the old run's embedded record verification had passed and the new run's record-verification blocker forced the analyst to correct. The new run introduced two memory-profile errors that profile verification endorsed: read-back called pull although an automatic selector the generator never requests is push under the profile type, and vectors called symbolic although representational-form puts embedding spaces in the parametric form (contradicting its own RT-OBJ-embeddings). The new overview.md is an index only; the old one was a self-contained reader entry point. Candidate changes, deferred until more runs: put the synthesis opening in the overview, and give the profile verifier the push/pull and representational-form rules explicitly. Watch whether later same-model runs repeat the two profile errors and whether verification passes them.
  • METHOD: [analyse-agentic-system, Dynamic Cheatsheet runs AAS-2026-10-19-dynamic-cheatsheet-2390eec63b75-01 (headless workflow, ADR 118) and the four surviving artifact-run engine worktrees 0043, 41e8, 43cd, 11e3, all pi-sol at source 5cfe3c37] — one same-profile comparison across the two routes (2026-10-19). The headless run published with 11 attempts, 0 corrections and 39:56 wall against the engine runs' 10–13 attempts, 0–1 corrections and 39:40–46:52; its members agree on every classification and consequential claim and differ in record count (22 against 35) and memory-profile unit judgments. One run per route, with the port's kept deviations and the model's variance not separated; no attribution to orchestration is warranted. Watch the record-count gap over the next batch before reading it as a method effect.
  • METHOD: [analyse-agentic-system, Dynamic Cheatsheet runs AAS-2026-10-20-dynamic-cheatsheet-523b6948294d-01 (headless workflow on the Workjig runtime, commit 14fa721) and AAS-2026-10-19-dynamic-cheatsheet-2390eec63b75-01 (ADR 118 pilot), both pi-sol at source 5cfe3c37] — one same-profile comparison across the runtime port (2026-10-20). Operationally identical: both published with 11 launches, no failures, corrections or refusals, all three verifiers accepting first time; 37:26 against 39:56 wall. Every storage, representational-form, admission, read-direction, push-selection and learning value agrees, as do the extractor defects (a missing </cheatsheet> still admits the rest), the eval scorers, post-adoption scoring and the synthesis conclusions. Differences are of the kind the pilot showed against the engine runs: memory-profile lineage, authority, curation and trace-source units, several epistemic ledger values, whether adapter chat history is retained memory (afforded and out of scope against wired), two reconciliation supersessions only in the new run, and each run citing a few code details the other omits. On lineage and on promote/synthesize at claimed basis the pilot is better supported by its own quotes; the rest turn on vocabulary interpretation. One run per side, so no attribution to the runtime is warranted; the port's failure paths (relaunch, usage stop, retry, acknowledge) were not exercised.
  • ABSTRACTION: [working-context memory-management analogy] — the 2026-10-13 survey of 17 retained analysis sets found mainly positional/budget trimming, runtime summarization or uninspected host compaction; AgeMem was the paper-only foreground-agent eviction case and gbrain the pointer-stub case in that survey. These are dated coverage findings, not a current corpus census. Method uptake is complete: kb/agentic-system-analyses/instructions/agentic-analysis-records.md records isolation, eviction, return paths and reload, using the stack-frame lens while rejecting heap-style pointer-following liveness. Remaining research is whether proposed generational eviction, barrier finalizers, explicit roots or evict-unless-declared-live policies improve a specified workload. Their transfer is not established by the analogy; collect mechanisms before promoting general rules.
  • METHOD: [analyse-agentic-system-amendments workshop, closed 2026-10-21] — no active follow-up is commissioned. Reopen trigger: a recurrence of the Graphiti Sol-run error classes (input discovery, quotation, formatting, classification) in two or more audited runs. The former handoff to analysis-audit-evidence is no longer active; both workshop closures are retained under Historical observations.
  • SYNTHESIS: [kb/sources/borretti-no-man-is-an-island.ingest.md, agent-harness-large-language-model-agents-survey.ingest.md, asi-bench-at-the-dawn-of-artificial-superintelligence.ingest.md] — a new field's depth and the depth of its public discourse can diverge: during a commercially hot transition, amplification favours marketing and onboarding material, so a practitioner sampling the visible discourse may judge the field shallow while technical work on harness design proceeds elsewhere. Borretti's essay is one instance of the inference; the cited papers show the problems admit deep treatment but not how common that treatment is. Not yet a mechanism: needs a way to measure the gap (for example, comparing technical and promotional shares across channels) before a note.

  • ABSTRACTION: [kb/sources/why-is-ai-generated-language-so-hard-to-understand.ingest.md, reverse-compression-is-when-llm-output-expands-without-adding.md] — LLM prose may fail readers at two separate levels: selection (too much content: restatement, hedging, exhaustive coverage — the walls of text operators see) and encoding (each included point worded too densely — what Clark, Martin and Mahowald measure, with content held fixed by round-trip paraphrase). The paper's "compression" is therefore not the opposite of reverse compression; both can hold in one text. Unconfirmed conjecture: a thoroughness reward plus per-token cost keeps content and squeezes wording. Editing consequence: word cuts that keep all content push toward dense constructions; cut by selection. Watch whether a KB rewrite under word-count pressure produces reduced relatives, noun stacks or denominal verbs before writing a note.

  • OBSERVATION: [operator report, session 01a18bfe-6954-7689-aada-f20bfc7b9aea, 2026-10-30] — the operator reports finding answers to their own LLM questions interesting and finding another person's generated report more engaging after a short introduction: “my mind kind of imports the state of mind of the person requesting the LLM answer.” Candidate explanation: the introduction supplies a question, uncertainty, and reason for caring, letting the reader assess the report as an answer rather than an unexplained collection of statements. This is one person's report, not a demonstrated mechanism or a population claim. Proposed comparison: hold a report fixed and compare authorship disclosure only, a content summary, and an introduction stating the requester's question and stakes; assess engagement and understanding separately. If the question-and-stakes introduction adds no benefit over the summary or disclosure, this comparison would not support the proposed explanation. No comparison has been run.

  • METHOD: [correction-record pilot, session 01a18bfe-6954-7689-aada-f20bfc7b9aea, 2026-10-30, kb/theory/intent-framed-delegation-is-a-control-regime-not-a-short-prompt.md] — candidate defect: the assistant presented intent-driven writing investigation as a methodology opportunity from the HN discussion without recognizing its existing role in Commonplace. Human intervention: the operator identified intent-driven investigation as a core procedure with Auftragstaktik as its methodology anchor. Changed result: the assistant reclassified the comments as examples of an existing method and retained only diagnosis-only versus replacement-prose feedback as a narrower workflow choice. Check: the linked note already explains the bounded delegation mechanism and names the writing workflow as an operationalization; this checks the duplication correction, not the effectiveness of either feedback mode. The human supplied the decisive correction; the assistant checked the retained methodology and revised its recommendation. This is one explicit record, not evidence that later sessions avoid the mistake. For a few further consequential corrections, capture the candidate defect, intervention and its supplier, changed result, supporting check, and unresolved uncertainty while they are available; trace harvesting can recover additional cases but must mark inferred reasons as inferred. Do not make this a universal reporting requirement before the trial shows whether the records are useful.

  • IDEA: kb/instructions/critique-note.md: consider bounded Internet search for counterevidence to a note's central commitment, beyond the critic's own knowledge. Candidate safeguards: state what would defeat or narrow the claim before searching; seek inspectable evidence rather than mere disagreement; check apparent counterexamples against the claim's scope and strongest response; report sources and search limits; preserve “no surviving attack” without treating an unsuccessful search as confirmation. Deferred: how to bound search cost and resist cherry-picking when opposing material is easy to find. No instruction change made.

  • DESIGN: [workflow code changes during unfinished runs] — problem: changing a workflow implementation can invalidate replay or interpretation of intermediate state. Options: pin each run to its old implementation until completion; maintain replay-compatible version branches; migrate explicit persisted state for new workers; or abandon the old execution and restart from declared inputs. “Long-lived data, short-lived code” moves compatibility obligations into state schemas and transitions; it does not remove them, and does not by itself require choreography rather than central orchestration. Operator preference for Commonplace: restart, accepting repeated token usage as the main repeat cost rather than paying for seamless continuation. Candidate safeguards: retain abandoned runs for inspection, isolate new runs from stale intermediate outputs, and reconcile already-published results or external effects to avoid duplicates. Motivation came from a user-supplied Twitter comment about a 24-hour prototype; author and URL were not supplied, and neither its general claim nor the restart policy has been tested here. Record the trade-off, not an implemented policy; revisit continuation or selective reuse if restart costs become unacceptable.

Historical observations

  • PROCESS: [Claude Code scanner test reporting, 2026-11-07; retained 2026-11-09] — session 3929c733-2dfa-4ee3-8405-dfdd1afe7c0d, L198/L213/L214/L225/L249: the claim that general and tooling suites passed referred to a run before the final scanner edit; only 16 scanner tests ran afterwards. No output defect was demonstrated. Commit 91810b276 records the repair and a post-edit general-plus-tooling rerun in its workshop README. Evidence: I6 in kb/work/hidden-failures-review-fixes/review-2026-11-07.md (retained at 0aff2ac4f). Closed as an episode, not a new testing rule; those tests were not rerun for this retention.
  • PROCESS: [scanner exit-status reporting, 2026-11-07; retained 2026-11-09] — the same session, L138/L139/L225: find | xargs returned 123, but the operator reply reported scanner exit 2 as though directly observed. The scanner's two input issues supported that inference, not the claimed observation. Commit 91810b276 added the xargs warning to scripts/scan_tools/README.md. Evidence: I7 in the same retained review at 0aff2ac4f. Closed as an episode; no additional rule or scanner change is needed.

These entries retain completed episodes or superseded workflow failures; they are not pending fixes.

  • ABSTRACTION: [mark-semantics.md, stale-indexes-are-worse-than-no-indexes.md, status field on notes (unenforced trust mark), ADR 025 (retired denormalized indexes)] — candidate theory underneath mark semantics: derived assertions over recomputable ground truth share one shape (hint-vs-guarantee: false-absence costs a query, false-presence is silently unbounded, so assertions are enforced-or-omitted, and dropping one degrades to recomputation). Likely siblings outside the KB: caches and materialized views, gradual typing's hint/guarantee divide, checksums, freshness stamps. Second in-corpus instance found by connect (2026-06-10): keep-compiled-views-aligned.md states the compiled-view source-of-truth requirement the mark rule instantiates (edge now authored from mark-semantics). Needs more instances before naming — watch the next mark added to the type system, the unenforced status: maturity field (it earns trust without a validator — does it bite?), and external systems in agent-memory reviews carrying confidence/validation metadata — Reconciled: the named theory now exists in kb/theory/a-derived-copy-of-recomputable-truth-must-be-checked-or-absent.md, which develops the asymmetry, deterministic-check preconditions and four applications. The original naming task is complete. The status-field question remains in that note's Open Questions rather than a duplicate research task here; its obsolete two-template example is a separate maintenance item.
  • ABSTRACTION: [oracle-strength-spectrum.md, weakly-discriminated-qualities-tend-to-be-underselected.md, reflection-buys-addressability.md, accumulation-counts-dependence-through-the-retained-result.md, self-improvement-is-relative-to-a-declared-objective.md] — the cluster's profile dimensions (addressability, cumulativity, reflective coverage, computational autonomy) look like objectives but are proximate targets: proxies adopted because the improvement objective is not measurable at the moment of decision. Nobody wants coverage; they want the artifacts to get better, and coverage is held to serve that. The KB already theorizes proxy-ness for checks — oracle-strength-spectrum names oracle strength as "the gradient underlying the exact-spec/proxy-theory distinction", and weakly-discriminated-qualities gives the failure mode (weakly discriminated qualities lose to strongly verified ones) — but only for things that yield a verdict. A structural property adopted as a target has no oracle at all: "all tests pass" is checked, "stay addressable" is not. That is the gap. Two consequences already visible. (1) It narrows the objective-revision open question in self-improvement-is-relative-to-a-declared-objective.md: revising a proximate target is settled by the objective it stands in for — does the new gate track artifact quality better? — an ordinary empirical question with no meta-criterion, so the regress bites only on endogenous revision of the objective itself. (2) It makes the dimensions-as-goals move contestable rather than licensed: whether coverage serves artifact quality is a claim that can be false. Live exposure in this repo: if artifact quality is weakly discriminated by the actual gate while structural validity and link health are strongly verified, the strongly-verified properties win regardless of intent. Route not taken: a second "adaptation goal" objective level borrowed from the self-adaptive-systems domain-goal/adaptation-goal split (Petrovska et al.), deflated because those goals are instrumental to the objective rather than a second objective — the salvage from the profile note's Petrovska paragraph, which is otherwise redundant and pending deletion. Before naming, settle what makes a structural property a good proxy and how you would discover it was not; and note that "proximate target" would read as a near-synonym of "oracle" unless the check-versus-property daylight is stated. — Reconciled: kb/theory/a-proximate-target-is-checked-for-achievement-not-for-warrant.md now states the target-versus-oracle distinction, separates achievement from warrant and explains how a mistaken target can be detected. The original theory gap is closed. Prospective target choice and delayed-outcome calibration remain that note's Open Questions; no new investigation is commissioned by this closure.
  • SYNTHESIS: [externalization unified review connect, context-efficiency-is-the-central-design-concern-in-agent-systems.md, deploy-time-learning-is-the-missing-middle.md, agent-runtimes-decompose-into-scheduler-context-engine-and-execution.md, skills-derive-from-methodology-through-distillation.md]: externalization may reframe context engineering as cognitive burden relocation; the KB has the pieces but not this mechanism as a named frame — Reconciled: kb/theory/semantic-work-can-be-relocated-but-not-eliminated.md supplies the general mechanism, distinguishing upstream amortization from offloading a loaded context. The mechanism is no longer unnamed. A comparison of the original cases is deferred unless a concrete design question needs it; no additional general note is required.
  • SYNTHESIS: [what-is-stored-in-this-kb workshop (belief/choice boundary), llm-executed-methodologies-are-metacircular-interpreters.md, a-framework-rule-with-a-boundary-preserving-rival-is-not-inherited.md, kb/articles/COLLECTION.md] — the publishability requirement on kb/notes/ states itself in basic computation vocabulary: a note may name a Commonplace choice only as a bound variable, universally quantified ("for any system that chooses X, Y follows") or existentially as a witness ("at least one system does X, so X is feasible"); a free occurrence is a claim whose truth conditions silently depend on a local selection, and it does not survive publication to the articles collection's stated audience of readers with no KB context. Deferred conjecture, scoped out of the storage workshop on 2026-08-22 as a side quest: binding a choice-variable may be exactly what makes a theory executable by a system other than the one that authored it, so publishability and portability-of-execution would be the same property under two names — which would also give the free/bound distinction an operational reading rather than a stylistic one. Lands beside the metacircular-interpreter note, not in the workshop; the rival test is the existing detector for the defect and demote-to-defaults is one worked binding move. Open before it could be promoted: whether a free choice-variable is detectable more cheaply than a Level B semantic judgment, and whether existing notes already pass or the requirement prescribes a migration — Superseded formulation: kb/theory/artifacts-must-preserve-named-choice-scope.md requires a proposition-relative reference rule or range and the choice's role at the intended consumption boundary; literal quantifiers are neither necessary nor sufficient. ADR 070 records the collection decision. Publishability is not established as identical to portability of execution; the narrower detection question remains open.

  • ABSTRACTION: [kb/agentic-system-analyses/retained/* (17 sets), agentic-analysis-records.md working-context control dimension, llm-context-is-composed-without-scoping.md, session-history-should-not-be-the-default-next-context.md] — survey of the 17 retained sets (2026-10-13) for how systems manage the model-visible context: eviction is code trimming by position or budget on every call (agentfly, atomic, atlas, agent-r, arigraph, agent-workflow-memory, agentic-local-brain), a runtime summarizer under pressure (letta-code, dynamic-cheatsheet), or host compaction left uninspected (gbrain, arscontexta, ai-research-os); only AgeMem (2601-01885v1) lets the foreground agent evict through its own tools, and only as a paper. Only gbrain leaves an in-window pointer stub after compaction; no set records liveness analysis beyond positional pins or prompt requests, and no set records skill or instruction bodies ever being unloaded. Candidate theory from the conversation that prompted the survey, by analogy to memory management in classical software: sub-agents are a stack discipline over context editing, and the analogy is close (a frame is pushed with selected inputs, popped whole, and only the return value escapes; liveness inside a frame is exact by construction; stack shape also keeps the prompt-cache prefix stable). The heap half does not translate, because context has no references: everything in the window is present by value, and anything outside it is reachable only by a named lookup the model must choose to perform, so there is no pointer to follow, no reachability to compute and no designated liveness; a tombstone or path is a convention the model may act on, not a dereference. What survives of the garbage-collector comparison is therefore partial: compaction today is a stop-the-world, lossy collector run under pressure; generational eviction, finalizers at the barrier, explicit roots, incremental collection and escape analysis via typed returns transfer; reachability-based liveness does not, which with non-destructive eviction to retained state inverts the default to evict-unless-declared-live. The memory lens had no dimension for this; a freeform working-context control dimension was added to the records contract, covering isolation in separate contexts (sub-agents, forked skills, task frames) as well as eviction and reload, so the next runs collect mechanisms before any vocabulary is derived. — Reconciled: the working-context control dimension and its stack-versus-heap distinction are present in kb/agentic-system-analyses/instructions/agentic-analysis-records.md. The historical survey and proposed transfers are retained here; empirical policy questions remain separately open.

  • STALE-CACHE: [kb/reference/proposals/archive/exceeding-a-review-budget-splits-the-task.md] — a dated "Current state" bullet said consumed-cost measurement was not yet built after implementation had shipped. The retained proposal now says "Consumed-cost capture ships as soft per-pair telemetry." It is also explicitly archived as retired by ADR 082. The specific stale-status defect is closed by the corrected passage, not by archival alone; the broader question about duplicating implementation status remains separately open.

  • FIX: [ADR 066 validation series, kb/work/popperian-maintenance-episode/adr-066-test-runs.md] — the modality machinery ran 23 full passes covering 21 of 22 surveyed candidates and accepted zero in-pass mode conversions. Both guards bound and rejected (statistical at run 9, ideal-type adequacy at run 14); instance-shaped defeats dominated and routed scope, category and conditionality repairs rather than mode conversion. The survey's per-note class labels failed as predictions wherever tested. Run 23 exposed another route: the prose gate stripped an unlabelled scalar idealization before any modality assessment. The series lacked an accepted statistical landing with a numeric refuter and an accepted in-pass ideal-type conversion. — Closed as an in-pass testing objective by ADR 080, which removed that conversion path. These missing cases remain limits of the historical experiment, not pending tests of the current full pass. ADR 066 is retained as amended: claim modes, in-text declaration, mode-aware premise review and author-applied conversion guards remain active; ADR 091 explicitly preserves the rest while tightening the universal refuter. Reconciled ADR 066's description, decisions and consequences with those amendments. No new positive conversion evidence is claimed.

  • REVIEW: kb/articles/which-existing-self-improving-systems-are-theory-builders.md: rechecked Recuris and Wheelhouse against their ingests' retained quotes. Both remain unsettled on criticism, but their rationales were corrected. Recuris's §3.4.4 passage explicitly reports two revised skills; growth does not establish absence of revision. The retained passage does not identify a stated consequence tested by the paired gate or a patch answering criticism of a skill's content. Wheelhouse's obsolete, changed and cancelled rulings establish incomplete curation rather than absence of revision; the report does not trace a particular criticism through revision and later use. Corrected both article passages without changing classification counts; the article validates with zero failures and warnings. No new grounding or broader system reassessment was performed.

  • FIX: [ADR 030 batch endpoints, Claude Code dynamic workflows] — the first harness-orchestrated review run executed end-to-end on 2026-06-12: selector, two prepared batches, parallel workflow reviewers (~52s), then ingest; four (note, gate) pairs recorded without Python changes. At the time, workflow arguments required inlining, deterministic endpoints ran in the parent because the script sandbox had no shell, ingest accepted no telemetry, and model partition provenance was an orchestrator assertion. Current finalization accepts opaque harness telemetry and optional runner/model/effort provenance; src/commonplace/review/finalization.py checks the supplied model's derived partition against the job before writing state. The old API limitation is superseded, not evidence that usage attribution or model authentication is solved. Untested harness limitations and the codec adoption question remain separately open.

  • FIX: [ADR 056, kb/reference/types/adr.md] — a dated-heading census found 69 ADRs dated 2026-07-25 or later, with every numbered ADR date parsed. Only ADR 115 lacked the exact required heading; it already retained the alternatives under ## Alternatives. Renamed that heading to ## Considered alternatives, with no inbound links to the old anchor found. All 69 now carry the required heading. The audit checks heading presence, not substantive completeness; no validator was added, and any future enforcement must preserve the older-ADR exemption.

  • FIX: [full-pass 20260826T115728Z-a18058, naur-binds-theory-to-humans-via-premise-that-machines-follow-rules.md, kb/instructions/run-full-improvement-pass-on-note.md] — editorial drift traced across one note in one day (2026-08-26). Motivation: Naur's human-only theory-holder rests on a premise an LLM contests. First draft (cp-skill-write, against the ingest's five retained quotes) overgeneralized to all "irreducibility" arguments and asserted the human premise was "assumed", a source-wide claim. The pass correctly defeated both (premise GLOBAL via a computability counterexample; grounding-alignment FAIL) and reframed to "Naur's retained passages do not establish a human-only theory bearer" — the claim shrank to the evidence base because the old pass could narrow a claim but could not extend grounding. Reading the full snapshot afterwards showed the human binding is argued, from machine = rule-following (§2, §5, §6, §8); four grounding runs (5 → 18 quotes) let the note return to its motivation, stronger than the first draft. Commits def1280a, 4d707961 hold the drifted versions. Distinct from the subject-narrowing escape in narrowing-bought-to-survive-review-is-paid-for-in-content.md: here the subject stayed, the evidence scope became the qualifier. — Closed at the protocol level by ADR 080: claim-bearing grounding failures route to a pending revise without changing the note; closing claim-level or evidential decisions require rollback and hand-back. Grounding expansion remains outside the pass, not an outstanding repair to this episode. Evidence: kb/work/popperian-maintenance-episode/fifth-episode-critique-without-repair.md and the current full-pass instruction; this closure does not establish observed compliance in later runs.

  • FIX: [full-pass 20260826T154716Z-3fa27b, unformalized-improvements-need-a-pre-formal-stage-in-the-loop.md, kb/instructions/run-full-improvement-pass-on-note.md] — the pass reframed the note to "Formal-only gates have distinct admission and warrant limits" on four GLOBAL premise defeats (L* automata learning, a grammar-enumerating generator, a formal disjunction) whose counterexamples all place the concept inside an admitted formal language, while the note's antecedent was a concept with no expression in any admitted language; the defeats equivocate on "unformalized". The bite rule adopted at 18:53 would not have caught it (premises were DEFEATED, not DOUBTFUL). Reframe reverted; thesis restored with the antecedent sharpened and the reframe's gains merged. A temporary guard required counterexamples to meet the antecedent under the note's own definitions. — Closed at the protocol level by ADR 080: the temporary guard was removed after a coherence audit found it could license inventing a narrower definition after seeing a counterexample. The pass now quotes the counterexample and targeted passage for the author's antecedent judgment; a false defeat can cause a hand-back but cannot authorize an in-pass claim rewrite. Evidence: the postscript in kb/work/popperian-maintenance-episode/sixth-episode-critique-defeats-by-equivocation.md and the current full-pass instruction; this closure does not establish observed compliance in later runs.

  • FIX: [kb/sources/prompt-stability-code-llms-emotion-personality-variations.ingest.md] — a bounded audit against the source- and checksum-verified snapshot found unsupported interpretations of framing sensitivity as underspecification, repeated sampling as a separate dispersion estimate, AUC-E as proof of a permanent language property, and non-monotonic scaling as evidence distinguishing theory defects from capacity limits. Repaired these passages, the attribution of a 40% result from prior work, and claims of calibrated or directly transferable perturbation distances. Untested mechanisms are now conjectures; the quote pool and existing correlation cautions are preserved. Tightened the sampling-spread footer in llm-output-deviation-requires-three-way-diagnosis.md. Both files validate with zero failures and warnings. This closes the bounded repair, not an independent replication or a check of every generated variant's meaning.

  • FIX: [kb/theory/definitions/self-improving-system.md] — all four formerly missing literature ingests exist, and the definition's Ailon, Weyns, and Petrovska citations now point to local ingests. Weyns and Petrovska were checked against retained quotes. Ailon grounding verified the name-paired snapshot's source and checksum and retained two passages describing input-driven training followed by fixed distribution-informed data structures; this supports the cited self-tuning architecture without a described reject-capable adoption gate, not a proof that every possible gate is absent. The definition and Ailon ingest validate with zero failures and warnings. The proposal-selection note already uses local ingests.

  • FIX: [Ma et al. prompt-stability citations] — the logged five-note count was stale: five external citations remained across three theory notes. All three now cite the local PromptSE ingest in body prose; duplicate or overstated footer citations were removed. The ingest no longer treats a nonsignificant correlation in 14 models as establishing independence or decoupled optimization objectives. All four edited files validate with zero failures and warnings. The subsequent bounded source-interpretation audit and repairs have a separate Historical observation.

  • FIX: [analyse-agentic-system-amendments workshop, closed 2026-10-21] — the Sol follow-up plan of 2026-10-03 closed with items 0 (evidence retention) and the remaining parts of 3–5 and 7 undone and uncommissioned. The empirical question (does in-place report correction under ADR 108 reduce total work and correction failures) and the unexamined blocked run AAS-2026-10-05-dynamic-cheatsheet-56ba8ab235b8-01 were handed to analysis-audit-evidence, which was subsequently deleted without extraction. That handoff is no longer an active destination; only the recurrence trigger remains under Open observations.

  • FIX: [analysis-audit-evidence workshop, closed 2026-10-21] — deleted without extraction: the operator judged its 21 run audits (Napkin, Graphiti, Letta, Letta Code, GBrain batch, OpenResearch, Flopwire, harness2) not worth retaining once the advisory trace audit moved into the analyse-agentic-system workflow. The in-place-correction question (ADR 108) handed over from the amendments workshop is dropped with it; the committed reports remain in git history at kb/work/analysis-audit-evidence/, the six uncommitted ones (OpenResearch, Flopwire, harness2) are gone.

  • FIX: [kb/reference/proposals/, ADR 056] — proposal disposal was unspecified: ADR 028 named adoption but no disposal operation, so eight July 2026 decisions split between six deletions (each with a properdocs.yml redirect to the adopting ADR, which the first audit missed for five decisions) and two ad-hoc retention banners. Resolved by archiving decided proposals to proposals/archive/ as a link sink, with extraction obligatory before the move, re-extraction rather than an inbound link as the repair, and provenance demoted from link to title mention. Discussion summary in kb/work/lifecycle-management/lifecycle-model.md. Archive-link enforcement was confirmed on 2026-09-18 and remains implemented by validate_proposal_archive_links in src/commonplace/lib/validation.py. The subsequent dated-heading census and conformance repair are recorded separately in Historical observations.

  • ABSTRACTION: [AGENTS.md python3 rule, ephemeral-computation-prevents-accumulation.md, spec-mining-as-codification.md] — agents run stdlib throwaway scripts as inline heredocs and discard them; a designated home for ephemeral helper code would create the accumulation substrate the ephemerality notes predict is lost (a recurring script is spec-mining input for a future commonplace-* command per progressive-constraining). Needs design before adopting: location (workshop-style scratch dir vs scripts/), committed vs gitignored, lifecycle/expiry so it stays a substrate rather than a junk drawer, and what recurrence signal trips promotion to a real command — Status checked 2026-09-18: Resolved by ADR 040 and scripts/README.md: reusable ad hoc tooling has a committed home, a cleanup norm, and a promotion signal.

  • FIX: [full-pass 20260818T115050Z-b196, kb/instructions/run-full-improvement-pass-on-note.md] — the pass's keep-reframe repaired a defeated universal title ("conjecture is seeing the particular as an instance of the general") by narrowing the subject into the predicate's own extension, producing an analytic title that passed every closing gate including claim-strength, which reads the title against the note's own local definition. The report recorded warranted contribution as "weakened" but nothing fails on that field, so the loop had no brake. Inbound citers revealed the real contribution — recognition cost and naming-amortization, both cut or hedged by the repair — so the note was rebuilt as recognition-not-linking-is-the-hard-problem-in-knowledge-systems.md and the failure named in narrowing-bought-to-survive-review-is-paid-for-in-content.md. Two mechanizable guards fall out and are not yet built: a refuter test in the closing cycle (after narrowing, name a case that would still refute the claim — an analytic claim cannot answer), and a citer check (a repair that drops what inbound citers import has repaired the wrong thing). Open design question: whether a "weakened" warranted-contribution verdict should be able to fail a pass the way a gate does — Status checked 2026-09-18: Superseded by the current full-pass protocol: claim changes are handed back; weakened, changed, or undetermined contributions trigger rollback. The older keep-reframe repair path is no longer operative.
  • FIX: [full-pass 20260818T132531Z-e11d99, narrowing-bought-to-survive-review-is-paid-for-in-content.md] — first application of the refuter and citer guards to a second keep-reframe, and they discriminated rather than firing on everything. Refuter test passes: "LLM output deviation requires three-way diagnosis" keeps an independently identifiable subject, and a two-question-suffices result or a fourth in-pipeline relation would refute it — the KB even holds a two-source rival in llm-debugging-starts-with-retry-versus-rewrite-triage.md. Citer test found two summaries still asserting the globally defeated "properties of different objects" premise (computational-model-README.md, llm-reliability-README.md) plus one "only error correction addresses this source" overclaim, all now reconciled. One residual the guards do flag and this pass routed but did not settle: non-substitution was weakened from "acts on exactly one object" to "primary target" plus "does not guarantee", and if "complete repair" just means "repairs the primary defect" while each remedy is identified by its primary target, that supporting claim drifts toward analytic even though the title does not. Worth one worked comparison rather than another qualifier — Status checked 2026-09-18: The remaining repair-versus-masking question was closed by the worked comparison recorded in the following observation.
  • FIX: [llm-output-deviation-requires-three-way-diagnosis.md] — the analytic-versus-operational residual left open by pass 20260818T132531Z-e11d99 is closed with a worked comparison rather than a qualifier. Writing it surfaced a third outcome the note had not separated: besides repair (the intervention reaches the defective relation) and failure to reach it, an intervention can mask a defect — lowering temperature concentrates D on an admissible point inside I while V stays wide, so behavior looks acceptable and nothing was repaired. Masking is separable from repair by perturbation: a masked defect returns when its masking condition lapses, a repaired relation does not. That is what makes non-substitution empirical, and the refuter test is what caught it — the first draft claimed an intervention "restores acceptable behavior only if it reaches the defective relation", which the note's own temperature example refutes — Status checked 2026-09-18: Closed by the worked comparison already recorded here.
  • FIX: [machine-studying-ontology workshop, machine-studying.ingest.md] — workshop closed 2026-08-19 with all seven ledger candidates dispositioned (full ledger record in commit 25895a9f, directory deleted after). Adopt: evidence-timing boundary, as the capture signal-timing sub-axis in memory-design-adds-operational-axes-to-artifact-analysis.md (corpus-only vs task/reward/demonstration/trace-conditioned; self-posed questions stay pre-task, self-mined traces do not). Translate: machine studying (deploy-time adaptation restricted by evidence timing — a regime and protocol, not a new category), expertise (activation-constrained action capacity; retrieval-controlled evidence and the across-budget open question promoted to knowledge-storage-does-not-imply-contextual-activation.md), three intervention families (representational-form placements of the retained change). Merge: agent as model plus harness (convergent evidenced-by on the-deployed-system-not-the-model-is-the-unit-of-learning.md), studying algorithm (the definition-update relation of instantiation-alone-cannot-model-agent-learning-across-sessions.md). Defer: studying intelligence (second-order scalar with no current consumer; the across-budget open question is the revisit hook). Side outputs exceeded the charter: the second ledger row spawned the instantiation note, the domain-pricing criterion note, ADR 066's claim modes, and six OO-attestation ingests — Status checked 2026-09-18: Closed workshop; this entry records its dispositions, not an outstanding fix.
  • FIX: [ADR 082, kb/work/atomic-step-adoption (closed 2026-08-27)] — grounding is now bounded on the artifact: at most five distinct tracked sources per note without a paired verbatim quote, enforced by commonplace-validate (FAIL); the three link-following gates lost their reader-side caps for a head-first rule; ADR 079 superseded; four review-side proposals archived. The eight notes over the bound at adoption were conformed the same day from already-retained quotes (2–10% word growth, no splits, no new grounding runs). — Status checked 2026-09-18: Completed adoption and conformance; this entry records the outcome, not an outstanding fix.
  • OBSERVATION: [goedel-machines-are-a-proof-governed-case-of-self-modification.md] — pre-existing verbatim-quote FAIL at line 32 (span not found in goedel-machines-schmidhuber.ingest.md's Quotes section); the only failing note in kb/notes as of 2026-08-27. — Status checked 2026-09-18: The rewritten note passes commonplace-validate with zero failures and zero warnings on 2026-09-18.
  • CONTRADICTION: [llm-code-boundaries-are-natural-checkpoints.md, llm-output-deviation-has-three-sources-with-non-substitutable.md] — checkpoint note claims both LLM-side properties collapse at an LLM→code crossing, but under the three-object model interpretation error crosses intact: a wrong-but-well-typed value passes the boundary, the very case the note's own debugging advice depends on catching — Resolved 2026-09-18: Rebuilt the checkpoint argument around inspection and conditional replay, with a wrong-but-well-typed booking example; reconciled three live citing notes.
  • FIX: [link-graph-plus-timestamps-enables-make-like-staleness-detection.md] — footer uses labels foundation and motivates, which are not in the current kb/notes/COLLECTION.md authorized set; surfaced by the domain-pricing note's connect pass. Reconcile the labels or the contract — Resolved 2026-09-18: Replaced the three unauthorized footer labels with grounds; removed the obsolete relationship-label example.
  • OBSERVATION: [src/commonplace/lib/quote_verification.py _citation_ranges] — the citation-range heuristic scans back to the nearest ( before a link, so (a) a bare inline link after an earlier parenthetical citation in the same paragraph swallows any quotation between them, and (b) a parenthesis inside a quoted span ((conservative), Ŝ(E)) mis-anchors the range and orphans the verbatim marker onto a later link. Two conformance edits on 2026-08-27 had to restructure paragraphs to work around it; a parser fix would remove the sharp edge. — Resolved 2026-09-18: Replaced nearest-parenthesis lookup with balanced enclosing ranges that ignore link and quote contents; added regression cases for swallowed quotations, misattribution, and nested citation locators.
  • FIX: [cp-skill-ingest, cp-skill-connect backlink lifecycle] — connection discovery runs on the immutable snapshot before its durable ingest exists, so it explicitly defers bidirectional and reverse-edge candidates, but ingest has no post-creation pass that reconsiders those links against the now-existing ingest; the 2026-08-31 Beyond Fixed Representations ingest therefore received outbound links while plausible note-to-ingest backlinks were systematically skipped. The workflow needs either future-ingest backlink candidates during snapshot connect or a mandatory post-ingest backlink pass. — Resolved 2026-09-18: The ingest skill now requires connection discovery on the accepted durable ingest, reports inbound candidates without authoring them, and retains the validated ingest if discovery fails. Re-ingest carries that failure separately from its existing-link audit.
  • STALE-CACHE: [the-risks-of-invariant-risk-minimization.ingest.md, concept-bottleneck-models.ingest.md] — both snapshots were "extended in place with a body extraction" after capture, so they are condensations, not captures: their retained quotes are verbatim to the snapshot but not to the paper (the IRM snapshot's abstract is an agent's summary, not the arXiv abstract). Snapshots are immutable, so the fix is a re-ingest under a new snapshot basename; until then a note quoting either source discharges the ADR 082 bound against summary text. Found 2026-08-27 while conforming the eight notes. — Resolved 2026-09-18: the existing rosenfeld-risks-of-invariant-risk-minimization.ingest.md records a full PDF capture from 2026-08-29, and its paired checksum matches. Concept Bottleneck Models now has concept-bottleneck-models-paper-v3.ingest.md from a complete version-pinned PDF, with seven paper extracts retained through grounding; both citing notes use it and distinguish the two OAI ablations. The older Concept Bottleneck Models ingest is explicitly marked superseded, with its checksum and Quotes preserved.
  • FIX: [commonplace-validate snapshot schema, kb/sources/.snapshots/{scaffold-not-vocabulary-popperian-code-generation-skill, form-not-content-placebo-controlled-self-repair, selection-without-signal-recovery-through-expression}.md] — the snapshot body-date check fails all three with body_dates.1: '2288-13-91' is not a 'date'. It is a false positive: the match is the tail of the bibliography DOI 10.1186/1471-2288-13-91, not a date. The captured bytes are correct and were kept; the ingests validate clean. First recorded in the closed conjectural-learning workshop's source register (2026-09-21) and still reproducing on that date. — Closed 2026-09-21: extract_body_dates in src/commonplace/lib/note_parser.py now accepts a date only when it stands alone, not as the tail of a longer hyphenated, slashed, or dotted token. All three snapshots validate clean. Across kb/ the change drops only date-shaped pieces of run IDs, URL paths, slugs, and this DOI; a regression test covers them