Reconciliation map: the brief's six directions against the library

Closing condition 1 of the framing. Written 2026-10-30 over three search files under search/ (317 entries, produced by Opus workers from rg sweeps and description reads) and full reads of the central notes named below. Entries the workers marked "description only" were not re-read; a verdict that rests on one is marked.

Each direction gets four lists. Held names notes that already state the result. Restated names the brief's claims that the library holds under other words. Tension names where the brief and the library disagree. Gap names what neither holds. A verdict follows.

Vocabulary

Term decisions are in vocabulary.md; tolerated double senses are in the vocabulary collision ledger. This map uses "search judgment" for an unretained decision about where effort goes, "commitment" for a retained resolution, "effort allocation" for the brief's allocation, and "actor" for a bounded person or agent.

Direction 1: the nature of commitments

Held. - Explanatory vs constructive commitment, and why retraction differs in kind: theory and capacity building make the same kind of commitment. - Commitment vs derivation as the ground-truth boundary: commitment, not derivation, creates new ground truth. - Belief vs choice after retraction: superseded choices are retained, superseded beliefs are not; conventions are adopted by decision, not refuted (epistemic stance, point 2). - "Worth investigating" vs "warranted to adopt" as different conclusions with different authority: lightweight search control; candidacy evidence licenses escalation, not acceptance, which anchors the split in Peirce's economy of research: a pursuit-worthiness judgment is not a truth judgment. - Reversible adoption vs costly entrenchment, with option value confined to timing: current-task fit alone does not warrant costly entrenchment; revision cost as a separate axis from acceptance: prototype standing. - A practical commitment resting on an untested linking claim: a proximate target is checked for achievement, not for warrant.

Restated. "Avoid collapsing a claim about truth into a decision about what is worthwhile" is the candidacy/verdict split. The ladder considering → adopting → acting → maintaining → retracting exists in pieces (discovery lifecycle for beliefs; adoption/entrenchment for structures; supersession for retraction). The economic framing adds only that each rung has a cost, which the entrenchment note already says for the costly rungs.

Tension. The brief wants the guess behind each effort allocation (what the investment will return) made explicit. The library evaluates search control distributionally: a search controller is tested by what it brings to stronger evaluation. Under that note, individual return guesses need not be retained or tested; the controller is. This is a real disagreement about what to retain. Resolving it is the one substantive question direction 1 leaves.

Gap. - No note says what kind of thing an effort-allocation decision is. My reading: it is a constructive-style commitment of limited authority and horizon. It changes what work exists, its reasons demote to provenance, and it is retracted by backtracking rather than by changing one's mind. The library already has every piece of that except the classification. A paragraph in the explanatory/constructive note would hold it; it does not need a note. - Goals and preferences are not treated as commitments. Objectives are declared parameters (self-improvement is relative to a declared objective). See direction 3 for why the library keeps it that way.

Verdict. Mostly held. One tension worth a discriminating test (below) and one classification paragraph worth adding at extraction.

Direction 2: bounded rationality and the economics of investigation

Held. - Allocation must precede decisive evidence, and more resources do not remove the choice: open-ended improvement allocates search before evaluation. The note says this "is not merely an economic constraint", which is stronger than the brief: the problem is structural, not only budgetary. - When to act on a guess: ad hoc explanation can be rational when error is cheap and local. - When deferral is productive: productive deferral requires option, evidence and convergence. - Qualitative judgment is the de facto position. Pindyck's option theory is kept as structure only ("no portable quantitative threshold"); prototype standing glosses "expected" as "anticipated, not probability-weighted"; the cognitive-architecture scan argues selection values are not one number. The exception is the machinery-investment note's "expected marginal net return" wording. - The limiting formal case: a Gödel machine's switch-versus-continue gate is an expected-utility proof, and it still leaves search allocation unsolved (Gödel machines).

Restated. "Do not assume numerical probabilities" is already the library's practice. "How actors decide when to investigate further" has local answers (cheap-error regime, deferral conditions, ADR 120's rule to keep searching while a consequential conclusion could change, full-pass closure) and no unifying note.

Tension. None with the library's claims. There is one with the brief's framing: the library's allocation notes do not derive from economics. They derive from the structure of open-ended search (evaluation unavailable before a candidate exists). The economic framing would be an added layer, not the ground.

Gap. - No treatment of metareasoning or value of computation (Russell–Wefald, Horvitz), resource-rational analysis (Lieder–Griffiths), satisficing or ecological rationality (Simon, Gigerenzer), opportunity cost, or explore/exploit. None of these has an ingest. The bounded-optimality ingest (Russell–Subramanian) exists and no theory note cites it. - No explicit statement of the library's position on expected-utility reasoning. It is consistent across notes but unwritten.

What the missing literature would add. Resource-rational analysis is the important one, because it is a rival explanation of the workshop's main claim. It explains apparently crude heuristics as optimal under computational bounds. On that account a routine is adopted because it is cheap enough and good enough, and its adoption need not track evidence about its stability. That is a non-epistemic account of routinization. The library's practice is on the other side: progressive constraining commits only after patterns stabilize makes adoption track evidence. The two accounts predict different things about when a routine should be relaxed; see the tests below.

Verdict. The mechanisms are held, under a structural rather than economic grounding. The literature check (closing condition 2) needs at least one ingest: resource-rational analysis first, Russell–Wefald second. Bounded optimality is already ingested and bears on direction 6 more than here: it makes the program-plus-architecture, not the reasoner, the unit whose value is assessed.

Direction 3: recursion and reflective improvement

Held. - The investment argument for the method level, with its conditions and erosion: an optimal long-run learning strategy invests in its own machinery. - Why the level above must sometimes be revised: learning inside a fixed decomposition inherits its mistakes. - What reflection buys and does not buy: reflection makes retained lessons second-order; reflection buys addressability. - Machinery at the method level is challengeable but persists by warrant: machinery persists by warrant, not position. - How far reported systems reach up the levels, and that every loop sits inside a larger one: the self-revision design space workshop and its evidence notes. - Goal change: revising an improvement objective is licensed from outside it. Most apparent goal changes are proxy revision, re-indexing, or surfaced under-specification; the residue has no improvement claim available; Commonplace shows no instance of the residue. - Stopping: owned by reflective-improvement-divergence (candidate claim: "episode closure, not convergence"); the shipped rules are log-and-defer and full-pass closure.

Restated. "An investigation is an investment justified by a conjecture; selecting the method is another; improving it is another" is the machinery-investment note generalized by one level. The same three functions (search, evaluation, retention) and the same prior-allocation problem recur at each level; the proposal-selection and allocation notes together imply this without saying it.

Tension. The brief treats "discoveries change higher-level goals" as a normal case of the recursive process. The library says a goal change is not licensed from inside the work governed by that goal, and that Commonplace has no instance. Keep the library's position. The economic framing has no mechanism for terminal goal change either; a guess about return is relative to a goal.

Gap. - The three-level stack is nowhere stated as one structure. Whether it deserves a note depends on whether stating it predicts anything the two notes do not. Candidate: the second-order note's open question about third-order lessons. - Method selection is treated as a control problem (capable agents need methodology selection), not as an investment with a cost of selecting and a cost of selecting wrongly. This is small and real; it folds into the routine-conversion account in direction 6 (selecting a routine is the cheap end of investing in one).

Verdict. Held, with one bounded addition. Nothing here should be extracted separately from direction 6.

Direction 4: general operations for open-ended work

Held. - The six operations (derive, compare, locate, revise, evaluate, apply), their task-level descriptions, the mechanical/interpreted/unavailable scale, and the finding that in a model-mediated system every operation is available by interpretation and none carries a classical guarantee: theory-refinement-interface (working document, source checks not done, not promoted). - The brief's split between "can the operations express the procedure" and "does reuse make it cheaper" exists in one sentence: a fixed-model house must retain missing procedures: "An operation that is possible but too costly is different from one the permitted machinery cannot express." The same note states the economic case for shared operations: reusable retrieval, traversal and checking tools can make a new theory the only retained change. - Warnings against designing from abstract operations: scenario decomposition, holistic rewrite, and compiling coordination preserves primitive, not aggregate, authority. - The guard against universality from examples, for factories: universality axis.

Restated. "Start with the existing interface rather than inventing a competing one" is already that workshop's boundary. "Avoid declaring universality on the strength of examples" is the universality-axis rule.

Tension. None. The interface workshop asks expressiveness and guarantee questions; it does not ask the brief's second question. That is a gap in its scope, not a disagreement.

Gap. - No note expresses a procedure from another domain (science, software architecture, legal drafting, exploration) as a composition of operations, and no cross-domain reuse evidence exists. The library touches each domain once (an experiment-selection note, the software-house article, one legal note, the ARC case in a workshop). - Several operation lists coexist unreconciled: the interface, the discovery lifecycle's six phases, the theory-builder Boundary list, the reflective operation profile, the change-operations catalogue. This is not the brief's concern but it blocks any claim that "a small set" exists.

Verdict. The economic framing's contribution is one question to send to the interface workshop: for each operation, what does it cost interpreted versus mechanical, and does reuse across theories amortize the cost of making it mechanical? Domain examples should not be pursued here; they are evidence for the architecture-level conjecture and belong to its test (direction 6). The unreconciled operation lists are a separate problem for that workshop.

Direction 5: cumulative progress and revision

Held. This is the most complete direction. - Why retention is what makes accumulation possible: ephemeral computation prevents accumulation; only explicit retention is durable, writable and addressable. - Backtracking as the retraction path for provisional choices: backtracking keeps lightweight search control provisional. - What to keep: unregenerable decision premises by rule (design rationale); regenerable explanation by cache ledger (LLM recompute cost); choice events but not refuted beliefs (superseded choices). - Why the two sides get different treatment: the ground-truth inversion in the commitment/derivation note. A commitment cannot be recomputed, so no cache ledger applies. - Retention cost: cheap adoption and weak retirement accumulate cost; value is observer-relative (information value); what to store needs a declared output spec (open-domain memory retention). - A checked outcome licenses episode retention but not abstraction: checked outcome licenses episode retention, not abstraction.

Restated. "Not everything is worth storing indefinitely" and "different kinds of retained information have different future value" are both held, the second across several pairwise notes.

Tension. None.

Gap. - No single account compares future value across the brief's kinds (knowledge, decisions, rationales, experiments, outcomes). A table over the pairwise notes would be a restatement; the brief's discipline says not to write it unless it predicts something. - No model of value decay for retained items. Staleness and retirement operations exist; a decay mechanism does not. Speculative; leave in log. - No theory of what one bounded episode should hand to the next. The nearest are the workshop-layer note and active work state is not retrospective memory. This is the one gap here the economic framing might fill, since a handoff is a retention decision under a context bound. Open; no draft yet. - The commitment note's open question — what retires a commitment with no successor — stays open and is not economic.

Verdict. Held. Record the episode-handoff gap as an open question; extract nothing else from this direction.

Direction 6: why this is an economical architecture for Commonplace

Held. - The strategy and its stated reasons: start from a running system people use because use supplies a selection environment where no fixed oracle exists (system use selects theory fit, description only); the declared objective is useful, better-warranted work per unit of human judgment (Commonplace as an instrument, which also says transfer "has not been demonstrated"). - The failure conditions, verbatim, in a bootstrap fits the Bitter Lesson only if learning outgrows it: including "each new domain requires a bespoke ontology and evaluator" and "human global judgment and maintenance grow with the corpus". - How to measure it: the ablation baselines proposal (five arms; human decisions, maintenance cost, negative transfer). - How work becomes routine and returns: the constraining/codification cluster, with codification as a bet (codification and relaxing), commitment after stabilization, relaxing signals, and the negative selection that leaves the agent boundary with less predictable work (preferential codification). - The closest account of when conversion pays: the repetition section of codify versus LLM decision heuristics: "Compare construction, invocation, checking, retrieval, and maintenance costs under the same outcome requirements"; recurrence is a cost signal and a candidate source, not a correctness criterion; a sound operation can be worth implementing after one case, another never. - What warrants a shared structure: entrenchment warrants and coordination value. One concrete case of a shared engine retired for its maintenance cost: ADR 118 superseding ADR 113.

Restated. "Managing search judgments and commitments recursively is economical under bounded resources" (the work-level conjecture) decomposes into held results: allocation before evidence, candidacy vs verdict, machinery investment, retention by ground-truth status. What is not a restatement is the claim that these are one process; see cross-cutting below.

Tension. The brief asks why "starting with explicit knowledge management" is a good strategy. The library's answer is not about knowledge management; it is that explicit retention is what makes machinery addressable and therefore revisable by the same loop (reflection buys addressability), and that use supplies selection. The brief's phrasing would need to be replaced by those two reasons.

Gap. - The architecture-level conjecture is not stated anywhere. The library has its ingredients (software-house endpoint, universality axis, failure conditions, measurement proposal) but no artifact says: reusable support for commitment management is an economical route to the endpoint, and here is what would defeat that. - No thesis that a small shared interface reduces duplication while permitting domain extensions. The nearest structures are configure(G, K_P), text contract plus profiles (ADR 042), and guarded defaults. The cost side has one direct case (checked inline blocks: "the natural-language equivalent of a missing function boundary"). - No single account of when conversion into routine pays. The repetition section has four of the five terms (construction, invocation, checking, retrieval, maintenance); stability of the pattern is in progressive constraining and relaxing signals; frequency is in ADR 040's promotion signal. Nothing joins them. - No cross-domain reuse evidence, and no measurement has been run.

Verdict. This is where the workshop's new content lives. Two artifacts are warranted: a theory note on routine-conversion economics (work-level, grounded in held notes, with stability as the epistemic term) and a proposal stating the architecture-level conjecture with its defeaters (mostly reusing the failure conditions and the ablation proposal). The "small shared interface" thesis should stay inside the proposal as a sub-conjecture; it has one case for and one against.

Cross-cutting findings

The work-level conjecture is a composition, not a new mechanism

Every mechanism the brief names is held. What the brief adds is the claim that they are instances of one process. The library has not said that, and whether saying it is worth anything depends on the test below. The honest statement: the economic framing re-describes the search-control, discovery-lifecycle, machinery-investment and retention notes with cost attached to each decision. The one place cost is missing from the library is controller evaluation (direction 1's tension).

The operator's reading survives the map, with a rival

The reading (epistemic work is the shared layer because allocation is only as good as its guesses, and epistemic work produces the routines) is consistent with every held note. Constraining is learning in Simon's sense; codification is a bet on an extracted regularity; adoption follows stabilization. The rival is resource-rational analysis: routines as heuristics adopted for cost, not for evidence about stability. The two agree on most behaviour and disagree on relaxing: the epistemic account says relax when evidence shows the committed reading was a wrong proxy; the cost account says relax when a cheaper or better general method arrives. Both appear in the constraining definition's relaxing clause, so the library currently holds both without choosing.

Discriminating tests

1. Controller cost. Add opportunity cost to the evaluation of a search controller and see whether it changes which controller is preferred on recorded Commonplace episodes (connect-report routing, review bundle ordering). If the ranking does not change, the economic layer adds nothing over "tested by what it brings to stronger evaluation". Facts gathered 2026-10-30 in search/controller-comparisons.md: seven comparisons fit the controller definition; cost entered the basis in three and reversed none (it agreed with yield or broke a tie). The three are all in the automated review pipeline; the routine controllers a person still drives have no cost-per-yield record at all. Scope condition from the operator (2026-10-30): computational cost is the binding consideration only where a controller runs unattended, as the review and revision system does; elsewhere the scarce resource is human judgment, which the declared objective already counts. Concluded 2026-10-30 against the Russell–Wefald ingest. Their rule: a computation's net value is its benefit to the decision (zero unless it changes the choice) minus its cost, and controllers of unequal cost are compared on one scale with the exchange rate swept, reporting whether the preference flips within the range. Commonplace's recorded practice (yield first, cost as tiebreaker) is that rule's limit as the exchange rate goes to zero, which the paper itself notes for A*. So the economic layer changes nothing on the record, and it would matter exactly where the rate is not negligible: unattended pipelines, where the record already uses cost. What transfers is structural, not a computation; the paper needs numerical utilities, a standing default action and priors on how a step revises estimates, none of which connect routing or bundle ordering has. Three items worth keeping: (i) the scoring rule, credit only for a changed later decision and every step pays, which sharpens "what it brings to stronger evaluation" (the failure-explanation note already says this for failures); (ii) the exchange-rate sweep as the form an unequal-cost comparison should take, qualifying the controller note's matched-resources condition; (iii) the meta-greedy prediction that a myopic controller undervalues setup steps such as reading before routing, an experiment not run. Verdict, revised on the operator's reading (2026-10-30): the record shows the zero-rate regime because most controllers are still driven by a person, and that is the regime Commonplace is leaving as it adds unattended pipelines. So the net-value form is not a separate layer now and becomes the required comparison as automation grows. Two items follow. The controller note's matched-resources comparison stays valid in every regime, unattended included: yield under equal budgets needs no exchange rate. What unattended runs add is the frequent case where controllers' costs differ materially; there, the net-value form with the exchange rate swept is an option alongside it, not a replacement (Astra's correction, 2026-10-30). And the instrumentation gap is the actionable one: cost-per-yield records for routine controllers are cheap to add before the pipelines multiply and costly after, which is the machinery-investment argument applied to measurement. It belongs with review-on-workjig and recurring-task-runner, the workshops building those runs; this workshop records it as a proposal candidate and does not build it.
2. Relaxing trigger — withdrawn 2026-10-30. The classification does not discriminate: discovering a mistake changes costs, and discovering a cheaper alternative requires evidence (Astra's objection). The epistemic versus resource-rational rivalry is deferred as an open question; no extraction depends on it.
3. Routine-conversion account. Draft the account with its five terms and apply it retrospectively to three conversions and one refusal to convert. If the account would have predicted each decision, it has content; if every case needs a sixth term, it is a list, not a mechanism.
4. Not worthwhile. The generalization is not worth adopting if tests 1 and 3 both fail: then the economic vocabulary names nothing the structural vocabulary misses, and the vocabulary collisions above are pure cost. Test 3 ran on 2026-10-30 and passed narrowly: two considerations missing from existing guidance, extracted as a section of the codify-versus-LLM note rather than a new note.

Candidate extractions, in order

1. Theory note: when conversion into routine pays (direction 6; tested by test 3 first).
2. Proposal in kb/theory/proposals/ or kb/reference/proposals/: the architecture-level conjecture with defeaters (direction 6). Read the proposal types before choosing.
3. Paragraph in the explanatory/constructive note classifying effort allocation (direction 1), applied as a separate commit at extraction.
4. One question sent to theory-refinement-interface (direction 4).
5. Log entries: episode-handoff retention (direction 5), third-order lessons (direction 3), value decay (direction 5).
6. Proposal candidate in kb/reference/proposals/: cost-per-yield records for routine controllers, as the instrumentation unattended runs need before net-value comparison is possible (test 1). The cost measure is step-shaped, not a token count (operator, 2026-10-30): work that fits one session is cheap and roughly uniform; work that must cross the context boundary, through sub-agents, code-managed workflows or any multi-step process, pays for coordination machinery and handoff loss. The library holds the mechanism (context is scarce for feasibility before cost; two boundaries, three regimes; a scheduler inside a conversation degrades; crossing contexts needs coordination guarantees) and ADR 113/118 priced one crossing. A record should therefore state which regime a controller ran in and what crossing machinery it invoked, before any exchange rate is considered. The same step enters the routine-conversion terms: a routine's use cost and construction cost jump when the routine cannot stay inside one context, which is why frontloading is a conversion that pays by keeping work below the boundary.
7. Qualification of a search controller is tested by what it brings to stronger evaluation: the scoring rule (credit only for a changed later decision), and the unequal-cost comparison form added as an option when costs differ materially, keeping matched-budget comparison; linking the Russell–Wefald ingest (test 1). Closing condition 2's source check is met for metareasoning by that ingest and for bounded optimality by the existing one; resource-rational analysis is deferred with the rivalry.

Not worth doing

• A commitment ladder note (restatement).
• A retained-information value table (restatement).
• A three-level recursion note on its own (restatement unless test 1 shows cost changes a decision at the method level).
• Domain procedure examples (evidence for the proposal, not content for this workshop).