Reusable commitment support as the route from theory building to other knowledge work

Type: reference/types/design-proposal.md · Tags: self-improving-systems, learning-theory

Workshop draft (closing condition 4 of open-ended-work-economics). Declares the design-proposal type so the contract checks it; it moves to kb/reference/proposals/ at extraction and does not adopt anything.

Current state (as of 2026-10-30)

  • Commonplace's declared objective is more useful and better-warranted knowledge work per unit of human judgment (Commonplace as an instrument). Its strategy is stated in pieces: build a theory builder and test whether it learns; start from a running system because use supplies selection where no fixed oracle exists; invest in machinery because an improvement is paid once and reused. No artifact states the step from that program to other kinds of knowledge work, or what would show the step is uneconomical.
  • The machinery that exists serves theory building in a KB: types and collection contracts, validators, review gates, freshness baselines, connect and ingest skills, the search-control and commitment notes, and the draft refinement interface (derive, compare, locate, revise, evaluate, apply) in the theory-refinement-interface workshop, which is not promoted.
  • Transfer has not been demonstrated: "Both instances were authored inside the system that states the claim and assessed by nobody outside it" (instrument note). No second kind of knowledge work runs on the machinery. The first-downstream-run workshop holds the template for an outside KB-building trial, not yet handed over.
  • The library's generality rule for the KB framework is to demote closed taxonomies to guarded defaults and gate universality on heterogeneous worked cases (universal knowledge framework); ADR 042 is its implemented instance (text contract plus profiles).
  • The only recorded shared engine built ahead of its second consumer was retired on its carrying cost (ADR 113 to ADR 118); the codify-versus-LLM note's "Stake and odds" section now carries that lesson.
  • The direction of development is toward unattended pipelines (review-on-workjig, recurring-task-runner), where computation cost becomes the binding comparison for controllers.

Problem

The work-level claim is held: open-ended work allocates search before decisive evidence, keeps "worth investigating" apart from "warranted to adopt", retains commitments by their ground-truth status, and converts stable patterns into routines. What is not held is the architecture-level claim the program rests on: that reusable support for these operations, built for theory building, is an economical route to supporting other kinds of open-ended knowledge work (scientific investigation, software architecture, legal drafting, exploration). The claim is implicit in the lead article's bet and in the software-house endpoint, but nowhere stated with its universality axis declared or its defeaters named. Without that statement, every success inside the KB reads as support and no failure reads as refutation.

The brief that posed this names the endpoint "a general computational environment". The library does not use that term and this proposal does not introduce it (see the workshop's vocabulary decisions); the conjecture is stated as generalization from theory building to other kinds of knowledge work, with the software house as the software-specific case of the endpoint.

The conjecture

Stated in the form the library already uses for factories, F_P = configure(G, K_P) (agentic substrate):

There is general machinery G — support for making, recording, criticizing, routing and retaining search judgments and commitments — such that each kind of open-ended knowledge work P is served by configure(G, K_P), where K_P is that kind's domain knowledge (its types, oracles, routines and vocabulary). Building and improving G once, and acquiring K_P per kind, costs less over the horizon than building per-kind tools, and the cost of the shared abstractions stays below the duplication they remove.

Three held results give it its economic form. A machinery improvement is paid once and reused by every later episode, so G's return grows with the number of kinds and episodes that use it (machinery investment). The expensive part per kind is acquiring K_P, not constructing machinery (factory construction does not establish knowledge acquisition), so a shared G does not remove the hard work; it removes rebuilding the easy part. And conversion of a kind's stable patterns into routines is the mechanism by which that kind's work gets cheaper, and the conversion machinery (observe, commit when stable, relax on failure signals, price the stake) is the same for every kind.

Universality axis, declared (per the universality-axis note):

  • Axis: constructional expressivity of G across kinds, plus reuse economy. Not production-knowledge acquisition reach: whether computation acquires K_P from evidence is the program's separate bet, and this proposal allows K_P to be human-supplied.
  • Covered class: kinds of knowledge work in which consequential branches are generated during the work and results accumulate across bounded episodes. Routine work, where no search judgment is uncertain, is out.
  • Supplied inputs: K_P by people initially; problems and acceptance judgments by users throughout.
  • Adequacy: the declared objective, measured as the ablation proposal measures it, per kind.
  • Human boundary: internal production roles may be human in K_P acquisition; the conjecture is about G's reuse, not about autonomy.
  • Horizon and resources: stated per test; the claim is comparative over a horizon long enough for G's reuse to amortize.

Design space

Three options for how G comes to exist. The proposal does not choose.

  1. Interface first. Promote the refinement interface and the search-control and commitment machinery as the declared G, then configure a second kind against it. Cheapest to state, dearest if the interface was shaped by theory building alone: a description that contains its target is weak expressivity (universality-axis note).
  2. Second kind first, extract later. Serve a second kind with the existing machinery as it stands, allow bespoke additions, and extract G from what two kinds in fact shared. This is progressive constraining applied to the architecture: commit the shared layer only after the pattern stabilizes across kinds. It pays observation time and risks the second kind's additions entrenching before extraction.
  3. No shared layer. Commonplace stays a KB framework; other kinds use generic harness features and their own tools. This is the rival the conjecture must beat on total cost, and the option the bootstrap note's last failure condition names ("a more direct learning method achieves better results at comparable total cost").

Option 2 is consistent with every rule the library already applies to itself (demote taxonomies to defaults, promote on stable repetition, size the stake to the stated odds) and with the ADR 113 lesson. That is a reason to expect it to be chosen, not a decision.

Forces

  • Entrenchment and coordination value. A shared G earns its cost only under an enduring constraint, scoped transfer evidence, or coordination value (entrenchment, coordination value). Adoption by more kinds raises migration cost without demonstrating transfer.
  • Carrying cost of shared abstractions. The one recorded case (ADR 113) had consistency machinery exceed the domain logic it served. The natural-language analogue is the checked-inline-blocks problem: one authoritative definition against literal, sufficient text in each bounded context.
  • Expressiveness is not economy. "An operation that is possible but too costly is different from one the permitted machinery cannot express" (fixed-model house). A G whose operations are available by interpretation for every kind establishes expressivity, not the conjecture.
  • Preferential codification. A shared layer takes the predictable parts of each kind; the kind keeps the hard parts (preferential codification). Success of G therefore does not show the kind's work got easier overall; the measure must be per unit of human judgment, not per task.
  • Cost is step-shaped at the context boundary. Work that fits one session is cheap and roughly uniform; work that must cross the boundary, through sub-agents, code-managed workflows or multi-step processes, pays for coordination machinery and handoff loss (context is scarce for feasibility before cost, coordination guarantees; ADR 113/118 priced one crossing). So the operations of G that stay inside one context are cheap for every kind, and the ones that must cross are where G's carrying cost and the per-kind cost both concentrate. The conjecture's economy therefore depends on how much of each kind's work G keeps below the boundary, which is a measurable property of a second-kind run.
  • Unattended runs make unequal-cost comparison common. Matched-budget comparison of controllers stays valid; where costs differ materially, net value with the exchange rate swept is the added option (the Russell–Wefald ingest). G's controllers need cost-per-yield records that state the regime they ran in before either comparison is possible.
  • The brief's wider claim is not carried. That economic reasoning is a more general organizing principle than epistemology, and the rivalry with resource-rational accounts of routinization, are deferred in the workshop and are not premises here.

Operativity and warrant

No consumer yet. What would consume G is a second kind of knowledge work run on Commonplace by an operator who is not its author, through the installed package and skills (ADR 086's delivery path), with K_P as that project's collections, types and routines. The first-downstream-run template is the nearest prepared instance, though its task is still KB building. No oracle is added by this proposal; the ablation proposal's measures are reused, and their warrant stops where that proposal says.

Adoption criteria

The proposal reduces to one question: does shared support deliver comparable outcomes at lower total cost in a second kind of work? The comparison is made within the second kind, against a competent baseline on the same work, so that architecture is not confounded with task difficulty.

Evidence that would support the conjecture:

  • A second kind of knowledge work runs on the machinery with only K_P added, and the record counts how much of G was reused unmodified, how much K_P was needed, what maintenance of G the second kind caused, and what share of the kind's work G kept inside one context.
  • On the same work, over a fixed horizon, Commonplace support reaches outcomes comparable to a competent domain-specific tool or to generic harness features alone (option 3), at lower total cost counted as setup, operation and maintenance. The ablation proposal's measures supply the outcome side; the cost side is the counts above.

Conversion of G's operations from interpreted to mechanical is one possible source of the savings, not a requirement: shared interpreted procedures can save authoring, discovery and review effort while staying interpreted, and making one operation mechanical does not by itself show net savings.

Evidence that would defeat it, drawn from the failure conditions in the bootstrap note and the instrument note, restated as cost comparisons:

  • The second kind requires a bespoke ontology and evaluator that share nothing operative with G, so nothing is reused.
  • Maintenance of G grows with the number of kinds by more than the savings the added kinds yield. Growth alone is not failure; growth that outweighs savings is.
  • The competent baseline reaches comparable outcomes at lower total cost over the same horizon.

The trigger for an adoption decision is the first completed second-kind run with the comparison recorded. One run supports a bounded next step; it does not establish the claim for every kind. Until then the proposal holds the conjecture and its defeaters, and the program's successes inside the KB remain evidence about theory building only.

Relevant Notes: