Search results: directions 3 and 5

Raw material for the reconciliation map. Coverage and summary only; no judgment about whether the economic framing adds anything. Paths are KB-relative (kb/...). Descriptions are quoted from frontmatter unless marked "first sentence".

Direction 3 — Recursion and reflective improvement

3.1 Definitions: reflective and self-improving systems

• kb/theory/definitions/reflective-system.md — "Definition — a system is reflective relative to selected aspects when an internal process uses a causally connected self-representation of them in its operation." Reflection is assessed against a declared boundary, represented aspects, a self-representation, an internal process, and two-way causal connection; it is an architectural capacity, not proof that every part is revisable. It separates self-description, introspection, reflection and intercession. Reflection and adaptation are orthogonal: a Smalltalk image reflects without adapting, Ashby's Homeostat adapts without reflecting. "Reflection therefore supplies a possible path for self-directed operation, not an improvement process."

• kb/theory/definitions/self-improving-system.md — "Definition — operative, evidence-responsive change to a system's own behavior-determining organization, read against a declared boundary, horizon, and improvement objective." Uses Ashby's two-loop distinction: operating a system versus modifying the system that operates. Four causal obligations must close: evidence shapes an update, the update changes the system's own organization, the change enters a live authority path, and later operation depends on it. Storage, acceptance or loading are evidence about single links, not the attribution. "An improvement criterion is required semantically; an explicit evaluator is not required architecturally." Reflection, gates and autonomy are not membership conditions.

• kb/theory/the-self-improving-system-definition-classifies-its-boundary-cases.md — "Ten boundary cases run against the self-improving-system definition — each classifies by the stated criteria alone; the stress they apply falls on boundary declaration, not on the membership clauses." Bears on levels because the same substrate is "being improved" under one boundary and "self-improving" under another (e.g. a fine-tuning pipeline). Human-maintained software already counts as self-improving when maintainers are inside the boundary.

• kb/theory/self-improvement-is-relative-to-a-declared-objective.md — "The improvement objective is a declared parameter alongside boundary and horizon, carrying two separable conditions — indexed by the analyst, antecedent in the pathway — whose failures differ in kind." An attribution is elliptical until boundary, horizon and objective are named. "Indexed" is a condition on the claim (naming the objective); "antecedent" is a condition on the pathway (the objective must exist independently of the change it licenses and must have shaped it). Failing the first makes a claim incomplete; failing the second makes it false or empty.

• kb/reference/commonplace-declared-frame.md — "The declared boundary under which Commonplace's self-improvement, reflectivity, and allocation attributions are assessed — what is inside, what is outside, and how to cite or depart from it." Declares the repository, its operative artifacts, consuming software and agents, and designated maintainers in their improvement roles as inside the frame. Any analysis that needs a different boundary must declare its departure.

3.2 Investing in machinery; the return on the method level

• kb/theory/an-optimal-long-run-learning-strategy-invests-in-its-own-machinery.md — "A machinery improvement is paid for once and reused by every later learning episode, so over a long horizon its return can exceed immediate learning; where it does, an optimal strategy diverts effort to the machinery." States the two-level allocation directly: effort is split between learning now and improving the machinery that learns. "Immediate learning yields one result; a machinery improvement yields a stream of results, one per later episode." The optimal allocation includes reflective investment "for as long as its expected marginal net return exceeds that of competing uses of effort", and it shifts back toward immediate learning when harder problems, resource limits or maintenance costs erode the return. Conditional on improvements being findable, persistent and assessable; judging productivity needs a capability measure that stays comparable over time, "which the system may not have."

• kb/theory/a-bootstrap-fits-the-bitter-lesson-only-if-learning-outgrows-it.md — "A hand-crafted starting state fits the Bitter Lesson only if scalable learning displaces the task- and family-specific production knowledge it supplies as claimed reach widens." Distinguishes a bootstrap (the running system using its current machinery to guide search for its successors) from a scaffold or prototype that is discarded. Calling present artifacts a bootstrap is not enough: "A promised path beyond hand-crafting is cheap." The strategy must show search and evidence progressively taking over production, or concede that a more direct computational method is better.

• kb/theory/machinery-persists-by-warrant-not-position-in-a-reflective-loop.md — "Reflection makes selected production machinery challengeable, but placement alone neither warrants nor requires revision; fixed general machinery may persist when its role and scope are earned." Reads Sutton's "build in only the meta-methods" as a fixed outside that gives gradient descent its stability. A reflective path that puts its evaluator or update machinery inside its revision surface "gives up part of that fixed point", and governance must carry the lost stability (no candidate authors its own admission, localized reversible changes). A fixed higher-level gate or kernel may still be the warranted design; fixed placement is not deferred human work when no claimed path needs to revise it.

• kb/theory/computationally-directed-self-improvement-is-a-reallocation.md — "The progress question for self-improving systems is not category membership but which decision-bearing functions humans still supply; the endpoint test is whether the boundary can be contracted to exclude them." Frames progress as reallocation of decision-bearing functions (search, evaluation, retention) between humans and computation within a fixed frame. The endpoint is a "cut set" test: transition is complete when humans can be excluded and the pathway still completes. Contraction is "not a ratchet" — functions can move back when an oracle proves a poor proxy. Reflection makes the cut set legible but is not used by the argument.

• kb/theory/improvements-can-accumulate-without-compounding.md — "Improvements accumulate when later improvement consumes or preserves retained results; compounding requires an earlier benefit to counterfactually improve a later episode, directly or through reinvested savings." Separates a growing stock of retained improvements from improvements that make later improvement cheaper. Benefits reach later improvement directly (better retrieval or evaluation) or indirectly (freed time reinvested in improvement work); an unused saving does not compound. Fixed human attention limits the scale and duration of compounding.

• kb/theory/compounding-is-tested-in-later-improvement-not-by-the-accepting-metric.md — "Compounding evidence must come from later improvement episodes through displaced productivity measures and causal traces, not from the metric that accepted the earlier change." The acceptance record looks identical under "compounding" and "repeated maintenance", so it cannot show that machinery investment paid off. Evidence must come from a later episode, measured on a quantity other than the accepting check, connected by a causal trace.

• kb/theory/oracle-accumulation-improves-the-selection-environment.md — "A failure retained as a lesson helps tasks that retrieve it; retained as a maintained check it improves selection for later candidates in its maintained domain and amortizes validation." Distinguishes two things retained from a failure: a lesson (improves proposals, best-effort via retrieval) and a check (improves evaluation of every later candidate in scope). Building checks is machinery production as a standing channel; the system learns "a better selection environment for later policies."

• kb/theory/increasing-computational-autonomy-relocates-human-effort.md — "In an open-ended system, increasing computational autonomy need not cut total human hours — attention moves to the frontier — so measure improvements per human judgment, not human time." When routine work becomes computational, the system attempts harder work and human attention moves to the new frontier, so total human time stays roughly constant. Bears on the return of machinery investment: savings are reinvested at the frontier rather than banked.

• kb/theory/codifying-predictable-choices-leaves-agents-with-less-predictable-work.md — "Explains the negative-selection mechanism by which preferential codification changes the composition of work retained at an agent boundary." If codified cases are more predictable than those left to judgment, the residual is concentrated at the unpredictable end. A universal result under its stated condition, not a claim that deployed systems satisfy it. Cited in the workshop README as the "frontier does not shrink" weak point.

• kb/theory/warranted-transfer-leaves-people-the-hardest-to-warrant-decisions.md — "When a system preferentially transfers decisions whose premises, criteria, and checks are available, the remaining human decisions become harder to warrant per decision; this predicts a residue composition, not structural computational openness." Transferring warrantable decisions to computation enriches the human residue for unrepresented premises, unsettled criteria and weak checks. Parallel to the codification note, stated for decisions about improvement.

• kb/theory/a-method-ceiling-bounds-the-method-not-the-transfer-already-made.md — "Separates envelope expansion, where a responsibility leaves the residual human work, from performance gains inside a fixed envelope, so a bounded method reaching its ceiling does not retract the transfer it already made." A mechanism's "automation envelope" is the set of responsibilities it carries. Improving a mechanism by its own method eventually stops adding responsibilities; that ceiling does not undo responsibilities already moved out of human work. Relevant to diminishing returns on method improvement.

• kb/theory/progressive-constraining-commits-only-after-patterns-stabilize.md — "Constraining via LLM code generation freezes a single projection of the spec in one shot, but progressive constraining observes behavior across many runs and commits only the interpretations that consistently emerge." The library's mechanism for converting recurring judgment into machinery: observe across runs, commit only the stabilized interpretation. Bears on when method investment is warranted (after stability), not on its pricing.

3.3 Levels: object work, method, method improvement, and their interaction

• kb/theory/reflection-makes-retained-lessons-second-order.md — "Reflection lets a retained lesson target a prior commitment explicitly — rejecting, revising, or rescoping it — while non-reflective correction acts indirectly through the substrate." A second-order lesson operates on a represented commitment (reject, rescope, revise) rather than sitting alongside it. Rescoping requires the commitment's scope to be represented. "The same lever that makes correction cheap makes damage cheap." Open question: "Whether there are third-order lessons worth naming — lessons about the system's rejection and rescoping policies themselves — or whether those collapse into second-order operations on the improvement process's own commitments."

• kb/theory/reflection-buys-addressability.md — "Self-improvement can accumulate without reflection — parametric learners do — but non-reflective retention gives only indirect handles; reflective retention makes the changed object addressable." Separates operative, cumulative and addressable retention. Lists "Improvement of the improvement process itself" as an expected advantage: an explicit evaluator, criterion or update rule is organization like any other, so a reflective pathway "can be turned on its own machinery." Addressability does not supply credit assignment, coherence, retrieval or admission.

• kb/theory/a-proposal-selection-loop-requires-search-evaluation-and-retention.md — "A proposal-selection improvement loop — candidates generated, evaluated with possible non-adoption, and accepted changes made operative — requires search, reject-capable evaluation, and operative retention." Names the three functions whose allocation is studied at each level. It is a subtype; direct-update pathways need no candidate that could be rejected.

• kb/theory/a-methodology-governs-its-own-extension-only-as-far-as-it-settles.md — "A retained methodology governs the consequential extension decisions it supplies or imports; actor competence can carry the process further without making those choices settled by the method." Extending a system by a method raises meta-decisions (artifact form, checks, authority). The method governs them only as far as its content settles them ("closure under its own recommendations"). A capable actor going further shows actor-plus-method capacity, not method closure. Closure and reflection are independent.

• kb/theory/methodological-and-computational-closure-track-different-changes.md — "Methodological closure tracks what a retained method settles; computational closure tracks the absence of human decisions during the assessed operation, without requiring every judgment to have explicit criteria." Two readings of how much a level is settled: whether the method decides, and whether a human must decide. A pathway can have one without the other.

• kb/theory/capable-agents-need-methodology-selection.md — "A capable agent may know many individually relevant but mutually incompatible approaches, so task control requires selecting a governing methodology rather than merely supplying relevant knowledge." Treats method selection as a control problem distinct from knowledge retrieval; more capability makes selection more consequential. It does not treat selection as an investment with costs and returns.

• kb/theory/learning-inside-a-fixed-decomposition-inherits-its-mistakes.md — "Why optimization cannot repair consequential distinctions, responses, or mappings outside the effective update space of a fixed task decomposition." Object-level optimization cannot fix an error in the level above it (the decomposition). A decomposition "helps determine what can count as a candidate at all." This is the library's reason a higher level must sometimes be revised.

• kb/theory/an-omitted-loop-function-and-a-frozen-one-need-different-repairs.md — "Five proposal-selection systems expose frozen functions, while a direct-update contrast shows why absence of a gate is not omission; HyperAgents supplies a preliminary partial unfreezing." When a loop underdelivers, ask which function does no work and whether it is omitted or frozen. Frozen functions are the higher-level machinery left outside the update surface.

• kb/theory/a-repeatable-operative-path-keeps-a-redesign-class-open-to-revision.md — "Operationalizes repeatable operative revision for a named redesign class as a causal path through representation, evidence-bearing determination, admission, installation, dependence, and continuity." Improvement machinery stays open to revision when a complete path connects evidence to an installed successor, and the path stays applicable after the transition. Repeatability is what lets the method level be revised again rather than once.

• kb/theory/goedel-machines-are-a-proof-governed-case-of-self-modification.md — "A Gödel machine admits self-rewrites through proof under its current formalization; this restricts admission without establishing how many useful changes are reachable or how reliably they are found." The limiting formal case of recursion: the machine can rewrite its own proof searcher, and admission requires proof that switching beats continuing the search. Makes "switch now versus keep searching" an explicit allocation. Rigor is relative to a given formalization.

• kb/theory/open-ended-theory-learning-and-factory-learning-close-the-same.md — "In Commonplace's arrangement, theory learning and software-factory learning require one connected reflective path; proof-governed switching alone does not settle criticism." Theory learning about the system's own organization and factory learning from production experience are "the same loop reached from opposite ends"; each needs what the other supplies.

• kb/work/self-revision-design-space/README.md and plain-account.md (peer workshop) — first sentence: "Explain where architectural redesign occurs across Commonplace, five experimental pathways, and the theoretical Gödel machine." Every reported self-improvement loop sits inside a larger builder loop; each reaches some redesign levels and leaves others supplied, and "Adding another fixed controller would only move the boundary upward." Commonplace claims an affordance for moving up another meta-level, not "an achieved infinite level". Surviving form of the kernel thesis: "nothing has to stay outside the loop forever, but incumbent conditions govern each particular transition." It separates followed procedure, good change and safe-to-try.

• kb/theory/evidence/six-reported-self-improvement-paths-expose-bounded-redesign-surfaces.md — "Comparative evidence separates operative redesign, revision of governing machinery, and contributions to later improvement from declared editability across six reported self-improvement paths including HyperAgents." Empirical comparison of how far each reported path reaches up the levels while "consequential parts of the method remain outside the reported update surface."

• kb/theory/evidence/real-self-improving-systems-occupy-combinations-no-rung-captures.md — "Casebook of thirteen placements on selected pathway-profile fields — Homeostat to Commonplace — reflection, cumulativity, allocation, and evidential limit in combinations no rung expresses." Argues against a single ladder: the properties (reflection, cumulativity, allocation) combine independently in real systems.

• kb/theory/evidence/commonplace-as-a-reflective-system.md — "Commonplace witnesses that a human-inclusive KB can be reflectively self-improving on one pathway despite uneven coverage and human-gated design judgment." The case evidence that Commonplace's method level is revised through its own retained artifacts on at least one pathway (ADR 026), with design judgment still human.

• kb/reference/tag-readme-trace-as-self-improving-loop.md — "Maps the ADR-026 tag-readme trace onto the search/evaluation/retention loop, showing which half of each step runs in code and which stays human." The worked instance where operational strain produced validator machinery.

3.4 Discoveries that change goals rather than advance them

• kb/theory/revising-an-improvement-objective-is-licensed-from-outside-it.md — "Objective change is improvement only against a level outside both objectives; proxy revision, re-indexing, and surfaced under-specification subtract most apparent cases." The library's main treatment of goal change. A successor objective's self-approval establishes nothing, and evidence bears on how a system stands to an objective, "never about whether that objective is the one to have." Three subtractions remove most apparent cases: proxy revision (judged against the unchanged objective), re-indexing (redeclaring the analysis), and under-specification surfacing (the narrow objective was a proxy for a wider operative one). The residue — terminal objective change with no wider objective — has no license: "the improvement claim is unavailable rather than merely unproven." Four kinds of outside level are listed; an epistemic standard can settle empirical disputes but not differences in what objectives value. "Commonplace exhibits no instance of the residue."

• kb/theory/a-proximate-target-is-checked-for-achievement-not-for-warrant.md — "Between an improvement objective and its oracles sits a target level — a property pursued because it is held to serve the objective — whose linking claim no check in the loop evaluates." Distinguishes four levels: objective, proximate target, operational criterion, oracle. Oracles sit in the "verdict position" and leave a trace when they misrank; proximate targets sit in the "goal position" and nothing in the loop tests the claim that achieving them serves the objective. Bears on how an intermediate goal can be wrong without any check noticing.

• kb/theory/a-claim-without-external-assessment-carries-three-obligations.md — "Without external assessment a claim needs its own contradiction-and-support rule, a comparison level for objective change, and a performance measure it does not grade itself, plus attribution when it asserts a cause." One of the three obligations is a comparison level for objective change, which external assessment would otherwise supply.

• kb/theory/intent-controls-choices-by-distinguishing-live-alternatives.md — "A stated intent controls a local choice only when it changes which live alternatives are admissible, preferred, worth further search, or sufficient to stop." A goal has force on a choice only if removing it would change the alternatives or their ordering. Distant intents gain local force through bridging sub-intents. Relevant to how a higher goal governs lower-level allocation, including "sufficient to stop".

• kb/theory/llm-generation-relaxes-goals-where-human-writing-stalls.md — "An LLM can ship fluent output after silently relaxing an unmet goal, while human composition may expose the same gap as a stall; a conjectural mechanism for why readers inherit the check." A conjectural mechanism by which a goal changes silently during work rather than by an explicit revision. Marked as a mechanistic hypothesis.

• kb/reference/proposals/automated-note-refinement-as-search-over-source-bundle.md — "Proposal: build the automated note-refinement loop on a source bundle that emits a set of notes, not a single identity-stable note — reframing non-convergence (split, drift, kill) as search outcomes." Refining a note can show that the target itself was wrong: it may split, drift into a different note, or turn out not to deserve existence. "The process is not convergent." The search object is changed to a source bundle, so goal changes become legitimate outcomes.

• kb/sources/crux-open-ended-ai-research-paper.ingest.md — "CRUX shadow evaluations document research critique that fails to redirect projects, separating local correction from project-level learning under a bounded expert-review protocol." External evidence that agents correct locally on criticism but fail to change project direction. Bears on discoveries that should change a higher-level goal but do not.

3.5 Learning software factories

• kb/theory/definitions/software-factory.md — "Definition — in the Greenfield lineage, a software factory is a configured family-specific software-production environment." Base term for the factory-learning cluster.

• kb/theory/definitions/factory-development.md — "Definition — factory development constructs or revises reusable family-level production machinery rather than one product's lifecycle state." The level distinction for factories: factory development (reusable machinery) versus solution development (one family member). "A discovery made during that work becomes factory development only when it changes reusable" machinery.

• kb/theory/factory-learning-is-experience-responsive-retention-that-improves.md — "Experience-responsive retention: production experience determines a retained change to reusable family machinery that later production depends on; factory-level learning is retention that improves the factory relative to a declared objective." Path: production, experience, change to reusable machinery, retention, changed later production. Each link excludes a weaker event (observation, product repair, discarded candidate). Carries a TODO to switch to "software house".

• kb/theory/factory-learning-mechanisms-should-be-compared-on-the-same-causal-job.md — "Compares factory-learning mechanisms on their shared causal job — experience-responsive retention — while separating update mechanisms from the project-theory function needed for open-ended coherent modification." The shared job includes an explicit "allocation of search or update effort" step. Minimal factory learning differs from the harder capacity to modify a long-lived program coherently.

• kb/theory/a-better-factory-claim-compares-operative-states.md — "The improvement claim's relata are predecessor and operative-successor states and its relation is declared before the development it judges; evaluator location is a separate declaration from the learner boundary." Retention has no quality requirement; "better" is a separate claim between two operative states under a relation fixed in advance.

• kb/theory/factory-construction-does-not-establish-knowledge-acquisition.md — "Recursive software-factory construction is prior art, but the demonstrated constructors receive the family definitions, metamodels, mappings, and expertise that determine the produced fact[ory]." Recursive factory-building (factories producing factories) already exists; it realizes supplied knowledge and does not show the system acquired the family knowledge. Separates construction from acquisition.

• kb/theory/broad-software-demands-create-pressure-for-agentic-factory-development.md — "Broad software demands make exhaustive predefinition of useful family-specific production machinery practically implausible, motivating agentic factory development without ruling out a fixed universal substrate in principle." An engineering-scaling argument for investing in machinery that builds machinery, not an impossibility theorem.

• kb/theory/agentic-substrate-needs-family-specific-machinery-to-be-a-factory.md — "Maps the bounded-call agentic substrate to Greenfield's software-factory ontology without calling every generic harness or generated program a factory." Factory = configure(general machinery, family knowledge). The general substrate alone is not a factory.

• kb/theory/addressable-theory-can-coordinate-heterogeneous-factory-development.md — "Tentative natural-language project theory may provide an addressable way to coordinate heterogeneous factory development while search, testing, and backtracking construct and revise it." Minimal factory learning needs no theory; coherent open-ended modification needs a project-theory function (following Naur). Natural-language theory is a candidate carrier.

• kb/theory/universal-software-factory-needs-a-declared-universality-axis.md — "Universal software factory is ambiguous unless the universality axis, covered class, supplied inputs, adequacy relation, and resource bounds are declared." Separates portability, factory-valued output, constructional expressivity and production-knowledge acquisition reach; evidence for one is not evidence for another.

• kb/sources/fluent-self-improving-software-factory-2081823472016335059.ingest.md — "A practitioner account of Fluent, a software factory that turns observations into tested, reviewed, merged changes and reusable project expertise." External practitioner case of a factory feeding observations back into reusable machinery.

3.6 When reflective improvement should stop or return to object-level work

• kb/work/reflective-improvement-divergence/README.md and problem-map.md (peer workshop, owns this question) — first sentence: "Determine when non-convergent improvement processing becomes a failure mode in reflective self-improving systems, which mechanisms are specific to reflection, and what kind of episode closure lets improvement remain open-ended across the system's lifetime without consuming the work it exists to improve." Motivating distinction: "A system may remain indefinitely improvable; an improvement episode must still yield control." Candidate central claim: "An open-ended self-improvement pathway needs episode closure, not convergence." Separates search non-convergence, revision drift, oscillation, scope expansion, criterion drift, operational non-termination and stock divergence; proposes "control allocation under a moving improvement frontier" as the possible unifying object. Candidate controls include hard budgets, incumbent survival, marginal-value decisions, phase boundaries, a stable in-episode contract ("fixed during an episode, addressably revisable between episodes"), external adoption decisions and opportunity-cost accounting. Cases: Yegge's Gas Town ("just two more things"), note refinement, full-pass closure, Gödel machine, deferred traversal improvements. Predicts a stronger critic without a budget lengthens episodes more reliably than it increases completed object work. Nothing promoted yet.

• kb/theory/traversal-improvements-should-be-deferred-via-logging-to-avoid-mid.md — "Loading writing methodology into an already-committed context window is expensive; a one-line log entry preserves the improvement signal at near-zero cost and lets a separate pass do the fix." Separates noticing an improvement from doing it. Fixing on the spot "costs more attention than the original task"; a log line keeps the signal and returns control to object work. The library's one adopted rule for returning to object-level work.

• kb/reference/full-improvement-pass-closure.md — "How the shipped full-improvement workflow reassays final note bytes, routes residual findings, and stops without claiming convergence." Shipped closure: one closing assay against the final bytes, residual findings preserved, then stop. "It does not mean every assay passes, every finding is fixed, or the note has converged to a fixed point."

• kb/theory/open-ended-improvement-allocates-search-before-evaluation.md — "Open-ended improvement must choose which questions, candidates, experiments, or proof paths to develop before decisive evidence about them is available; even a Gödel machine's proof gate retains this prior search problem." Allocation at every level happens before decisive evidence. "This is not merely an economic constraint": evaluation may be impossible until a candidate exists, so more resources do not remove the choice of direction. (Listed in the workshop README under direction 2; included because it governs every recursive level.)

• kb/sources/autoreason-self-refinement-that-knows-when-to-stop.ingest.md — "Autoreason paper showing self-refinement improves only when candidate synthesis is paired with blind comparative judging and incumbent survival, with gains concentrated in the generation-evaluation gap." External evidence for incumbent survival as a stop rule: making "do nothing" a first-class candidate counters the bias of critique prompts toward always editing.

• kb/sources/goedel-machines-schmidhuber.ingest.md — "Schmidhuber's Gödel machine — rewrite of any part of one's own code, proof searcher included, gated on a proof of higher axiomatized utility: the proof-governed case of reflective self-modification." Source for the formal switch-versus-continue rule.

• kb/sources/darwin-godel-machine-open-ended-evolution-self-improving-agents.ingest.md — "DGM declares its own departure from the proof-governed Gödel machine, but its mechanism is looser than a benchmark gate: viability alone admits a child to a monotonic archive, and score only weights reproduction." An empirical recursive self-improvement system whose admission is weaker than proof; its population controller stays fixed.

• kb/sources/poetiq-perspective-on-recursive-self-improvement.ingest.md — "Poetiq's vendor account defines RSI as same-lineage recursive improvement and claims autonomous cross-benchmark harness evolution, but leaves redesign closure, compounding, and safety warrant unestablished." The zero-human-intervention result is scoped to harness construction after people chose task, data, objective, evaluator and outer process.

• kb/sources/bounded-recursive-self-improvement.ingest.md — "AERA demonstrates observation-driven model learning inside a protected architecture, while leaving recursive compounding and scalability untested." A historical case of learning inside a protected kernel whose representation language, scheduler, drives and value calculations are fixed by designers.

3.7 Related, less central

• kb/theory/the-bitter-lesson-selects-production-methods-not-representational.md — "The lesson's axis is production method — hand-crafted versus search-and-learning — not representational form." Supplies the build-mode the machinery note relies on.
• kb/theory/commonplace-builds-a-theory-builder-and-tests-whether-it-learns.md — "Commonplace builds a system meeting the theory-builder conditions and tests whether it learns; …" Learning is the hypothesis under test, not established by membership.
• kb/theory/a-complete-theory-path-does-not-establish-improved-capacity.md — "Mediation, empirical contact, response to criticism, and recurrent mediation support different claims; none alone establishes improved capacity, …"
• kb/theory/unformalized-improvements-need-a-pre-formal-stage-in-the-loop.md — "An improvement whose concepts have no expression in a loop's admitted formal language is reached only through a pre-formal stage, inside the loop or fixed at design time in the choice of language; translation relocates that stage." A limit on how high a formal loop can reach.
• kb/theory/diagnostic-richness-constrains-outer-loop-learning-quality.md — "Outer-loop learning depends on inspectable failure evidence, not only on the oracle used to select winning candidates."
• kb/theory/choosing-what-to-learn-requires-both-validity-and-learning-value-gates.md — "Separates two promotion checks for learning loops: whether a candidate is trustworthy enough to learn from, and whether learning it would improve the current system." The learning-value gate is a value judgment over what to invest learning in.
• kb/theory/warranted-autonomy-is-bounded-by-oracle-domain.md — "Bare autonomy is free, but warranted evaluation autonomy extends only to the candidates an oracle can assess with the required confidence."
• kb/work/self-improvement-cluster-operationalization/README.md (peer workshop) — first sentence: "Operationalize the self-improving-systems cluster as an ontology and a Commonplace-specific overlay on established improvement methods." Its trigger finding is that the cluster "fails its own operativity test": nothing loads it when Commonplace's machinery is changed, which bears on whether method-level theory actually reaches method-level decisions.

Not found (direction 3): - No note states the three-level stack in the brief (investigation as investment, method selection as investment, method improvement as investment) as one recursive structure. The long-run machinery note has two levels (immediate learning vs machinery); the reflection-second-order note leaves third-order lessons as an open question; the self-revision plain account describes a meta-ladder without economics. - No note treats method selection as an investment with cost and return; capable-agents-need-methodology-selection treats it as control only. - No library note on discoveries during object-level work that change the work's goal (as opposed to an improvement pathway's objective). The objective-revision note covers improvement objectives only and states Commonplace has no instance of the residue; note-refinement drift/kill and CRUX are the nearest material. - No adopted library theory on when reflective improvement should stop or yield. Candidates exist only in the reflective-improvement-divergence workshop; the shipped pieces are the log-deferral rule and full-pass closure. No note gives a marginal-value stop rule for method improvement beyond the "expected marginal net return" clause in the machinery note. - No note on how machinery investment interacts with goal change (e.g. machinery built for an abandoned goal) beyond the generic entrenchment and retirement notes in direction 5.

Direction 5 — Cumulative progress and revision

5.1 What retention enables across bounded episodes

• kb/theory/ephemeral-computation-prevents-accumulation.md — "Discarding an intermediate artifact loses that artifact's reuse path, not all learning; cross-run accumulation fails only when no experience-dependent state survives to affect later work." Accumulation across runs requires that some experience-dependent state (code, theory, examples, tests, policy, weights, external records) reach the later run. The unit is the deployed system, not one file.

• kb/theory/accumulation-counts-dependence-through-the-retained-result.md — "Cumulativity counts dependence through the retained result only; counting the evidence channel that result caused would make it coextensive with operativity." A later episode builds on an earlier one when information from the retained change shapes the later candidate, evaluation or update. Dependence that runs only through evidence the change caused does not count, or every closed loop would be cumulative.

• kb/theory/raw-accumulation-does-not-create-usable-memory.md — "Accumulation preserves material, but usable agent memory requires ingress work that adds handles, scope, relationships, provenance, trust signals, and lifecycle pressure." "A store can grow while all three properties get worse": findability, composability and calibrated trust. Retention without ingress work can make later episodes worse off.

• kb/theory/only-explicit-retention-is-durable-writable-and-addressable.md (title: "Explicit retention provides direct targets for selective revision") — "Explicit artifacts give a learner direct targets for inspecting and revising commitments; durability, writability, and effective addressability still depend on the boundary and available operations." Separates durability, writability and addressability as three properties of retained state.

• kb/theory/retained-artifacts-enable-persistent-deployment-time-adaptation.md — "Retaining evaluated changes to behavior-shaping prompts, rules, tools, and tests gives deployed systems a persistent adaptation path outside model-weight updates." Names deployment-time adaptation: a selected change affects later sessions without entering weights.

• kb/theory/retained-theories-may-improve-sample-efficiency.md — "Conjecture: retained theories may reduce target observations under structured shifts; a useful theory's reuse benefit is separate from selecting it by estimated explanatory-reach." When a task shifts but keeps structure, a retained theory may reduce the new observations needed. "Whether retaining that basis beats reconstructing it is an empirical question."

• kb/theory/proposals/retained-theories-compared-with-retained-traces-under-resource-limits.md — "Proposal: distinguish context limits, reconstruction budgets, and the conjectured value of retained theories; trace strategy, effective context, and derived state remain unadopted terms." Unadopted terms for comparing retaining theories versus retaining traces under resource limits. Separates the conditions that require retained state from the conjecture that theories supply it efficiently.

• kb/theory/a-retained-instruction-preserves-what-testing-selected.md — "Explains why an instruction generated from model weights can still add KB value: testing selects a procedure under a criterion and retention makes that choice reusable." Even when the content is regenerable from weights, the test result (which candidate met the criterion) is not. "Reuse avoids repeating candidate generation and evaluation."

• kb/theory/instantiation-alone-cannot-model-agent-learning-across-sessions.md — "The class/instance analogy captures session startup but omits the retained update relation that can revise later agent definitions and reusable-content placement." The cross-session update relation is what makes bounded sessions cumulative.

• kb/theory/active-work-state-is-not-retrospective-memory-or-chat-history.md — "Active work state needs current pointers, evidence gates, and closure; treating it as retrospective memory or chat history preserves the wrong state." Distinguishes three kinds of retained state by the question they answer: what happened (retrospective), the interaction order (chat history), and "what is live now, what must happen next, and when can this work leave the active set?" (active work state). Bears directly on carrying unfinished work across bounded episodes.

• kb/theory/a-functioning-kb-needs-a-workshop-layer-not-just-a-library.md — "Explains workshop as the temporal counterpart to a permanent knowledge library: in-flight state, dependencies, expiry, and promotion bridges, with tasks as an early prototype." Two retention regimes: in-flight state with expiry (workshop) and durable accumulated knowledge (library), joined by promotion bridges.

• kb/theory/learning-is-not-only-about-generality.md — "Per Simon, any capacity change is learning; accumulation is the basic operation, explanatory-reach its key property (facts low, theories high); capacity splits into generality vs reliability/speed/cost." Places accumulation as the basic learning operation and ranks retained items by explanatory reach (facts low, theories high).

5.2 Backtracking, deferral, search state

• kb/theory/backtracking-keeps-lightweight-search-control-provisional.md — "Backtracking preserves the provisional status of a heuristic branch choice by restoring an earlier usable state and redirecting search after contrary evidence." Requires a preserved earlier usable state, recognition of contrary evidence, a return path and redirection. "Without an operative return path, a trial can become a de facto commitment." It does not recover spent effort. Applies only to state the process can actually restore.

• kb/theory/lightweight-search-control-does-not-license-adoption.md — "A search judgment is lightweight when its authority stops at allocating further investigation, probing, continuation, suspension, or abandonment rather than licensing an operative change." "'Worth investigating' and 'warranted to adopt' are different conclusions." Retention of a search judgment does not promote it to an adoption.

• kb/theory/productive-deferral-requires-option-evidence-and-convergence.md — "A qualitative test that separates controlled postponement of a consequential choice from hidden commitment, mere delay, and unmanaged indecision." Deferral needs a preserved option (an alternative present commitment would destroy), discriminating evidence, and a convergence rule. Together they "do not prove that waiting will be worth its cost."

• kb/theory/failure-explanation-changes-later-branch-decisions.md — "An explanation of a failed branch becomes operative search control only when its retention changes a later choice about scope, priority, probing, continuation, or abandonment." A retained failure explanation that changes no later branch decision "has not yet improved search control" and does not separate learning from rationalization.

• kb/theory/program-theory-sustains-search-under-delayed-feedback.md — "Holding a program's theory is tested by whether a partial, tentative account of what the program is for keeps modification search, backtracking, and recovery coherent until delayed evidence arrives, not by whether the first change is right." Retained project theory is what keeps backtracking and recovery coherent between delayed feedback signals.

• kb/theory/ad-hoc-explanation-can-be-rational-when-error-is-cheap-and-local.md — "Explains why a disposable local guess can rationally select the next probe when error is cheap and contained, while retained explanations need reach checks." Sets a retention boundary: disposable guesses need no reach check because they are not retained; retained explanations do. The conjecture-store workshop cites this for "Retention is cheap; promotion pays the toll."

• kb/theory/cost-sensitive-formalisms-for-tentative-theory-search.md — "Exploratory map of backtracking, learning, and complexity models that expose budgets relevant to search guided by tentative theories." Survey of formal models with budgets for backtracking search. (Listed in the README under allocation.)

5.3 Design rationale and decision premises

• kb/theory/design-rationale-must-preserve-unregenerable-decision-premises.md — "Retention test for source-checkout design rationale: keep current decision premises not faithfully recoverable from implementation, git, and general knowledge; treat recoverable, role-free explanation as a cache." Retain what the interpreter cannot regenerate from implementation, git and general knowledge: relevance decisions, applicability limits, which alternatives were actually considered, and epistemic status. Recoverable passages may still deserve retention for independent roles (authority, cross-check, provenance, activation cue, exact record, stable address). Role-free recoverable explanation is judged as a cache. "Greater reasoning strength cannot recover a particular historical fact that no declared source determines." Interpreter-relative: a stronger model shrinks what must be retained.

• kb/theory/specific-intent-may-out-yield-local-rationales-facts-stay-separate.md — "Conjectures that an unrecoverable governing intent yields more local rationale per token than rationale snippets, while contingent design facts need their own record." An unmeasured conjecture about value per token inside the retention boundary: retaining the governing intent may reconstruct more rationale than an equal-size bundle of local rationales, but contingent facts still need their own record. This is the nearest library claim that different retained kinds have different yield.

• kb/theory/use-tests-a-decomposition-locally-rationale-makes-transfer-testable.md — "Running a decomposition confirms only that it sufficed here; because many force-sets fit the same split, rationale retained at design time is what gives a transfer claim an antecedent to test." Use is an outcome check that licenses replay, not a rule. Retained rationale gives a later transfer claim something to test.

• kb/theory/revision-guided-by-rationale-needs-faithfulness-not-just-legibility.md — "When revision of an addressable theory relies on rationale to locate a failed premise, misleading rationale can direct repair to the wrong part; rationale is one optional repair aid." Retained rationale has value only if faithful; unfaithful rationale misdirects revision.

• kb/theory/evidence/missing-rationale-does-not-exclude-a-builder-weights-do-across-runs.md — "The August 2026 Prime Agent, Recuris, and Apodex reports: retained rules without a recorded rationale leave theory-builder membership open, while Apodex carries only weights across runs, where no unit says anything, so no builder spans its runs." What kind of retained object (rules vs weights) determines whether criticism can span runs.

• kb/reference/design-rationale-management.md — "How Commonplace distributes design-rationale records across proposals, ADRs, contracts, and enforcement, without enforcing end-to-end continuity." Describes the shipped distribution: workshops, proposals, ADRs, contracts, type specs, validators, code and git each retain different parts of rationale. "There is no central rationale" store.

• kb/reference/adr/074-git-is-the-change-history-layer.md — "Accepted decision that git holds the change narrative of the source checkout, reference and ADRs retain only premises that a named change operation must read, and instructions declare the git read paths." Retention criterion for reference: a passage earns its place by naming the change operation that must read it; change narrative goes to git.

• kb/reference/adr/077-content-routing-by-regeneration-source-and-consumer.md — "Accepted decision that content placement is decided first by what an interpreter cannot regenerate from implementation, git, and general knowledge, then by the consuming operation's force; content-routing.md is the table; limits get a named section." Generalizes ADR 074 to every content kind: first ask whether it should be written down at all (regenerability), then where (consumer force).

• kb/reference/proposals/decisions-bind-their-consumers-through-site-back-pointers.md — "Proposal: make a recorded decision reach the operations it constrains by having each ADR declare the artifacts it binds and each bound artifact carry a checked back-pointer, rather than a hand-maintained decision index." Retained decisions consulted "by luck" carry full authoring and maintenance cost with unreliable consumption: "Retention without a consumer is storage, not operative change."

• kb/work/decision-lifecycle-evidence/README.md (peer workshop) — first sentence: "Determine how Commonplace should preserve a decision's development from proposal through deliberation, implementation, and later evaluation, while making its current state clear to agents and readers." Governing question: "what must remain available behind a concise account of the adopted architecture so that later work can criticize its grounds, reconsider alternatives and revise the choice without reconstructing the design space from scratch?" Requires comparing what later revision can recover, what routine work must load, and the cost of capture and maintenance; a concise view "must not silently delete retained alternatives, objections or original expectations."

• kb/sources/programming-as-theory-building.ingest.md — "Naur's theory-building view makes maintainability depend on situated design understanding while bounding what retained rationale alone can transfer." Source for the limit that unwritten judgment has no retained form.

5.4 Commitment versus derivation; what to keep after supersession

• kb/theory/commitment-not-derivation-creates-new-ground-truth.md — "Derivation — claims recoverable from the source, nothing added — leaves the source as ground truth; what adds unentailed resolutions becomes ground truth at commit, repaired by supersession." Derived artifacts stay dependent copies (repair by recompute; deletion costs recomputation). Committed artifacts become ground truth for what they add (repair by supersession; deletion is "irrecoverable loss of the only record"). The boundary predicts disposal: generated indexes were deleted and rebuilt (ADR 025); adopted proposals were archived (ADR 056). Regime attaches to regions, not files. Codifying a recurring arbitration commits once, after which regeneration is derivation. Open question: "What retires a commitment when no successor commitment is coming…?"

• kb/theory/superseded-choices-are-retained-superseded-beliefs-are-not.md (title: "Superseded choices need a historical witness; refuted beliefs lose subject-matter standing") — "Retention after supersession follows remaining truth role rather than maintenance operation: preserve a witness to the choice event, while a refuted belief loses subject-matter standing." A refuted belief loses standing; a superseded choice remains a true event ("the system committed to X at t") and needs at least one reliable witness. Two maintenance regimes: choice records grow by appending; belief bodies are reworked holistically, keeping old wording only when a separate historical use earns it.

• kb/theory/theory-and-capacity-building-make-the-same-kind-of-commitment.md — "Theory building and capacity building both retain resolutions their evidence does not entail; an explanatory commitment stays answerable to the object it describes while a constructive commitment changes the object, so retraction differs in kind." Once implemented, a constructive commitment "becomes ground truth: its reasons demote to provenance, and later theory must account for the organization it produced." So project theory has a historical component: understanding a system needs to know which constructive commitments produced it.

• kb/theory/history-has-one-chance-to-become-checkable.md — "An artifact's production history is convertible to later-checkable form only at production time, via records/attestation or re-derivability; after that a bounded reviewer sees only carried state." Retention timing: some history can only be captured when it happens. Uses W3C PROV and in-toto as exemplars.

• kb/theory/prototype-standing-is-revision-cost-binding-plus-lost-investment.md — "A theory's prototype standing is its expected revision cost — external binding plus the investment a revision discards — so natural-language versus symbolic form determines neither component and acceptance status is a separate axis." Revision cost has two parts: external binding (consumers) and intrinsic reconstruction cost ("the investment a revision discards", e.g. proofs, safety cases, trained models). Bears on the cost side of revising retained work.

• kb/theory/abstract-an-experience-only-when-you-can-state-the-boundary.md — "Abstract an episode into a lesson only when you can state its boundary, else preserve the instance; an over-generalized lesson is one that drops the condition clause." Choice between retaining an instance and retaining an abstraction; the boundary clause is the discriminator, not frequency, confidence or recency.

• kb/theory/checked-outcome-licenses-episode-retention-not-abstraction.md — "One result-only check can warrant retaining an episode as evidence, but abstracting its explanation also needs evidence about a faithful producing process and an explicit scope boundary." Outcome checks license keeping the episode; process checks are needed to keep the explanation.

• kb/theory/retaining-episode-evidence-keeps-a-distilled-rule-open-to.md — "Keeping relevant episode evidence and its relation to a distilled rule preserves a route for re-examining that rule; reconstruction, comparative value, and correct generalization still require testing." "This is one useful relation between retained forms, not a requirement to keep every transcript." The test is whether the surviving record keeps the premises needed to re-examine the lesson.

• kb/theory/trace-extracted-memory-earns-authority-per-operation-not-at-capture.md — "Trace memories begin as records; verification, abstraction, and consultation earn authority under progressively harder oracles, while unverified stores accumulate guesses presented as knowledge." A witness ladder: capture grants no authority; each later operation earns a stronger reading.

• kb/work/lifecycle-management/disposal-and-retirement-survey.md and lifecycle-model.md (peer workshop) — first sentence of README: "What life-cycle does a KB artifact travel — from workshop intake through promotion, maturation, and eventual retirement — and where do current conventions under-specify the transitions?" Corrected archive-versus-delete test: "Archive when the residue is irreproducible observation — dated current-state anchors and the corpus measurements a design rested on … Delete when the residue is claims, which a successor either carries or falsifies. Deliberation is never grounds for archiving." The lifecycle model reads tiers as coordinates (attention cost, link-graph role, relation to ground truth) and maps the cache analogy (re-extraction as refill-on-miss, extraction as writeback-before-eviction), separating "recomputable cache" from "committed compression".

• kb/reference/adr/056-adopted-and-retired-proposals-archive-out-of-the-frontier.md — "Adopted and retired proposals move to a link-sink archive after extraction; provenance becomes a title tag, so ADRs from 2026-07-25 must carry a Considered alternatives section." The shipped retention rule for superseded design work.

• kb/reference/adr/085-withdrawn-proposals-are-deleted.md — "Decision that a design proposal whose problem or adoption trigger has lapsed, with no decision foreclosing it, is withdrawn by deletion rather than archived or kept." A third exit: when no commitment rested on a proposal, nothing has a claim on it, and it is deleted.

• kb/reference/adr/036-review-acceptance-is-current-state-not-append-only-history.md — "Review acceptance stores only the current accepted baseline per note/gate/model key and prunes superseded review evidence inline." A shipped case of deciding not to retain history because "no current code consumed that history."

5.5 Store versus recompute economics

• kb/theory/llm-recompute-cost-inverts-the-store-vs-recompute-default.md (title: "LLM recompute cost shifts the store-vs-recompute balance") — "For model-facing derived values, costly model-side recomputation shifts cache economics toward checked materialization, but persistence pays only when its total expected cost beats the alternatives and the copy substitutes for work." Compares three paths over a validity window: model-side derivation, external on-demand derivation, persisted checked materialization, with total costs including creation, reads, source changes, validation, retrieval, context and failure recovery. "The materialized answer must replace work, not merely precede it." Validation keeps the cache subordinate to the source.

• kb/theory/opposed-recompute-factors-do-not-decide-documentation-segmentation.md — "When audiences trade per-reconstruction savings against recurrence, neither factor alone determines cache value; among feasible alternatives, segmentation separately depends on whether specialization repays every cost introduced by the split." Supplies the general cache ledger: "net cache value ≈ (net saving per equivalent reconstruction) × (reconstructions avoided) − (maintenance)", over a chosen horizon in a currency the project uses.

• kb/theory/a-derived-copy-of-recomputable-truth-must-be-checked-or-absent.md — "When an artifact carries a copy of information recomputable from a ground-truth source, the copy must be machine-checked against that source or not exist — hand-maintained-and-trusted is forbidden." Rests on an asymmetry: an absent copy costs one bounded recomputation; a false trusted copy causes silent unbounded error.

• kb/theory/an-insufficient-summary-precedes-the-source-rather-than-replacing.md — "When a summary cannot license a reliability-compliant stop, the authoritative fallback remains in the path; only fallback work the summary removes can offset its own cost." A retained summary pays only if it lets the reader stop; otherwise it adds cost on top of reading the source.

• kb/theory/localized-retention-pays-where-change-is-sparse-in-a-matching.md — "Addressable retention localizes a sparse change when units match its decomposition; total adaptation stays local only when the affected units also have a small, explicit impact closure." Retained addressable units pay off when environmental change touches few units (k ≪ n) and those units have a small dependency closure; dense change erases the advantage.

• kb/reference/storage-architecture.md — "Authority and lifecycle boundaries among Commonplace's authored files, local source copies, derived artifacts, and operational state." The shipped split of which representations are authoritative and which are reconstructable.

5.6 Retention costs, lifecycle and retirement

• kb/theory/cheap-adoption-and-weak-retirement-accumulate-cost.md — "Explains why locally cheap structural additions become routing, maintenance, and migration debt when retirement needs distributed evidence and lacks an equally operative path." Cost comes from lifecycle asymmetry: additions are local, retirement needs distributed evidence. Retirement must be an operative path (signal, target, dependency evidence, owner, disposition, migration). Cites 148 agent-memory systems: only seven perform all seven tracked lifecycle operations. Predicts pressure, not inevitable growth.

• kb/theory/current-task-fit-alone-does-not-warrant-costly-entrenchment.md — "Distinguishes reversible adoption from costly structural entrenchment and confines option reasoning to the timing of a commitment supported by an enduring constraint, scoped transfer warrant, or actual coordination value." Entrenchment needs an enduring constraint, transfer evidence, or coordination value; real-options reasoning affects timing only.

• kb/theory/maintenance-capacity-must-match-harmful-artifact-inflow.md — "Stable quality depends on capacity to prevent, contain, detect, and repair harmful retained artifacts keeping pace with their risk-weighted inflow, for which gross generation volume is only a proxy." A retention-cost invariant: maintenance capacity must keep pace with risk-weighted harmful inflow. One widely reused defect can cost more than many harmless outputs.

• kb/theory/agent-memory-requirements/retire-redact-supersede-relax.md — "Memory systems need lifecycle operations for redaction, decay, supersession, retirement, relaxation, and temporal validity." "Every stale artifact competes for attention, search rank, review time, or behavioral authority." Append-only capture suits provenance; derived and activated surfaces need lifecycle operations.

• kb/theory/agent-memory-requirements/preserve-evidence-without-loading-history.md — "Trace retention should preserve evidence for audit and extraction without making raw history the agent's default context." Separates capture cost from loading cost. Broad trace retention is justified because "the future consumer is often unknown at capture time"; text traces are cheap to keep, while large media, regulated material and multi-user streams change the calculation.

• kb/theory/definitions/lineage.md — "Definition - lineage records the source dependencies needed to invalidate, regenerate, retire, or review retained behavior-shaping artifacts." The dependency record that retirement and invalidation need.

• kb/theory/stale-self-description-conceals-its-own-staleness.md — "What artifact drift adds when it is reflexive: the process that would detect it consults the artifact that drifted, the trigger has no edit event to hook, and synchronization load scales with autonomy." A retention cost specific to retained self-descriptions: synchronization load grows with autonomy.

• kb/theory/link-graph-plus-timestamps-enables-make-like-staleness-detection.md — "Existing links already encode dependency information; comparing note and target timestamps flags notes that may be stale without any new annotation, analogous to make's file-based rebuild logic." A cheap staleness signal for retained notes.

• kb/theory/false-positive-generation-is-filtered-before-retention.md — "False-positive generation faces evaluation before retention, while false-positive acceptance becomes operative and can compound." Retention is where errors become costly: a bad accepted change acquires a consumer and force and is caught by nothing downstream.

• kb/reference/adr/007-reports-directory-for-generated-snapshots.md — "Decision that kb/reports is a collection whose cache, state, and retained areas make report retention policy explicit without conflating generated output with disposable output." Shipped three-tier retention policy (cache, local state, durable retained) for generated reports and experiments.

• kb/reference/proposals/workflow-recovery-history.md — "Proposal: preserve enough workflow recovery evidence to distinguish clean completion from completion after repairs." Current job state resets failure history on accept, so a clean final state cannot show whether refusals occurred. A case where discarded history removes evidence later method evaluation needs.

• kb/work/use-and-outcome-records/README.md and synthesis-pilot-cost-case.md (peer workshop) — first sentence: "Opened on 2026-10-02 at the operator's request. Work out how Commonplace can retain and use observations from ordinary operation to guide later decisions, without requiring a controlled experiment for every idea." Working hypotheses: whether an idea should guide a decision depends on stakes, reversibility and observable consequences; cheap capture may beat up-front assessment; reuse must stay distinct from favorable outcome. Cost comparison "includes capture, retrieval, interpretation, retention, and later correction, not only model calls." The worked case records a decision to replace a 36–48-call pilot with an eight-call check before any outcome existed. No schema adopted.

• kb/work/conjecture-store/README.md (peer workshop) — first sentence: "Move Commonplace from a library that retains mostly corroborated theory to a system whose primary state is a large store of cheap, criticizable conjectures, with the corroborated theory collection derived from that store as a projection at a stated threshold." Retention policy for weak conjectures: "Retention is cheap; promotion pays the toll." Notes that log conjectures go unread by write and connect skills, so a second witness can arrive unnoticed — a retained item with no consumer.

5.7 Different kinds of retained information have different future value

• kb/theory/information-value-is-observer-relative.md — "Information value is observer-relative: prior knowledge, tools, compute, and goals determine extractable structure, grounding use-shaped reshaping and discovery." Grounds value in the data-observer pair and cites value-of-information theory, relevance theory and bounded rationality. The minimum retained content is the gap between what a use needs and what the consumer can supply, which may be substance, a connection the consumer will not make, or "what reconstruction cannot preserve: warrant, provenance, exact wording or membership, a governing version, or an authoritative local choice." The literature survey is marked TODO (from training data).

• kb/theory/open-domain-memory-retention-needs-a-declared-output-spec.md — "Explains why an input stream alone can't answer 'what to store' in open-domain memory design; a declared output spec supplies the missing inclusion criterion." Input-driven retention (novelty, frequency, salience) accumulates whatever comes up and hides gaps; output-driven retention measures candidates against a declared account of what the system must serve.

• kb/theory/choosing-what-to-learn-requires-both-validity-and-learning-value-gates.md — (see 3.7) A learning-value gate asks whether retaining a valid candidate would improve the current system.

• kb/theory/oracle-accumulation-improves-the-selection-environment.md — (see 3.2) Lessons and checks retained from the same failure have different delivery guarantees and so different future value.

• Also bearing on differential value, already listed above: superseded-choices-are-retained-superseded-beliefs-are-not (choice events vs refuted beliefs), design-rationale-must-preserve-unregenerable-decision-premises (unregenerable premises vs recoverable explanation), specific-intent-may-out-yield-local-rationales-facts-stay-separate (intent vs local rationale per token), learning-is-not-only-about-generality (facts low, theories high), the lifecycle-management archive-versus-delete test (irreproducible observation vs claims), and active-work-state-is-not-retrospective-memory-or-chat-history.

5.8 External sources bearing on retention

• kb/sources/agora-git-as-shared-memory-for-collective-autoresearch.ingest.md — "Agora's sustained agent run shows how immutable contribution lineage, downstream evidence, and diversity views can coordinate shared research, while leaving their causal benefit untested." Append-only lineage DAG with rebuildable views for frontier and neglected branches.
• kb/sources/prov-overview.ingest.md — W3C PROV, used by history-has-one-chance-to-become-checkable as an exemplar of carried provenance.
• kb/sources/the-price-of-meaning-why-every-semantic-memory-system-forgets.ingest.md — "Formal no-escape theorem for semantic memory interference, with exact-record and symbolic-verifier escape clauses that sharpen retrieval-vs-verification tradeoffs." A cost of retention: interference grows with what is stored.
• kb/sources/novel-memory-forgetting-techniques-autonomous-ai-agents.ingest.md — "Formula-based adaptive forgetting with constrained optimization for agent memory — … with empirical evidence that uncontrolled accumulation causes false memory propagation."
• kb/sources/palantir-ontology-vs-decision-traces.ingest.md — "Jaya Gupta frames Palantir-style top-down ontology and workflow-first decision traces as two ways to build LLM-facing world models." Retaining decision traces as a rival to retaining a designed ontology.

Not found (direction 5): - No single note that compares the future value of the kinds the brief lists (knowledge, decisions, rationales, experiments, outcomes) under one account. The differentiating claims are scattered across the notes in 5.7 and 5.4, each comparing two kinds. - No note on how the value of retained information decays over time or with change in the system (time-discounting or obsolescence of retained items). Staleness detection and retirement operations exist; a value-decay mechanism does not. - No library theory on retaining experiments and their outcomes as such. The shipped reports retention tiers (ADR 007) are policy, not theory; use-and-outcome-records is open and adopts nothing. - No note gives a retention-cost ledger for decisions or rationale comparable to the cache ledger for derived documentation. The design-rationale note sends recoverable explanation to the cache ledger but gives no cost model for retaining unrecoverable premises (they are kept by rule). - No note treats retention economics for workshops or bounded episodes specifically (what an episode should hand to the next). The workshop-layer note, active-work-state note and archive-versus-delete test come closest. - The commitment note's open question, what retires a commitment when no successor is coming, has no answer in the library.