Open-ended knowledge work shares one epistemic layer across roles

Type: types/note.md · Tags: constraining, goal-prioritization, learning-theory

The epistemic layer of a kind of knowledge work is the machinery that keeps three statuses apart and criticizes each by the standard its status admits:

  • derivable: what follows from the task and its sources, recoverable without new choices;
  • decided: a commitment, a retained resolution the task does not entail, which becomes ground truth when committed and is repaired by supersession (commitment, not derivation);
  • guessed: a search judgment, an unretained decision about where effort goes, whose authority stops at allocating further search (lightweight search control), or a tentative theory not yet decided by test.

The claim: in open-ended, bounded, text-mediated knowledge work, six roles consume this one layer, and each role names a failure that a tool without the layer shows. The claim is universal over those conditions. A counterexample is a tool that lacks the layer and performs a role below without the failure that role names.

Theory building (propose, derive, criticize, blame, revise, select) remains the registered name for the operations of roles 1, 5 and 6. Roles 2, 3 and 4 consume a different part of the layer: which status a resolution has, and what a check can establish. They need status and warrant tracking, not conjecture and refutation. The layer is therefore wider than theory building, and this note names it.

Premises

  1. Open-endedness generates branches. Consequential branches appear during the work, so the candidate set is not fixed in advance.
  2. Boundedness forces choosing before evidence. An actor cannot develop every branch and must choose which to develop before decisive evidence about them exists (search before evaluation).
  3. The context bound makes the work episodic. A text-mediated actor works in a bounded context, so what one episode settles reaches the next only through a retained artifact. Accumulation needs explicit retention, and retention needs to know which retained items are ground truth.

Premise 1 creates the guesses, premise 2 makes acting on them unavoidable, and premise 3 makes the statuses travel in artifacts, where a later reader cannot recover them unless they were recorded.

Roles

Role Consumer Takes from the layer Fails without it
1. Prioritisation effort allocation among branches return guesses and their criticism allocates by habit or cost alone and cannot improve its guesses
2. Constraining routine conversion the split between derivable and decided freezes a commitment as a fact; the relaxing trigger cannot be named in advance
3. Retention storage policy the same split stores derivations as truth or treats commitments as regenerable
4. Evaluation reach acceptance gate what a check can establish the accepted population drifts toward what the oracle measures
5. Method improvement reflective investment criticism of the method itself improvement stays inside the method's mistakes
6. Application theory building as a kind of work all three statuses for theories the work has no support beyond the actor's context

1. Prioritisation. Effort allocation consumes guesses about what an investment will return, and criticism of those guesses. The layer marks them as search judgments, so evidence that routes a candidate to costly assessment is not read as evidence that accepts it (candidacy evidence licenses escalation). A search judgment is then criticized by what it brings to stronger evaluation (search-controller test). Without the layer, no guess is recorded as a guess, nothing can be compared against its outcome, and allocation follows habit or cost.

2. Constraining. Constraining commits choices that narrow how a consumer reads an artifact; work is constrained when it runs under such choices in natural language and codified when it runs under codification. Conversion consumes the split between what the task fixes and what the converter decided. Only a decided resolution is a bet that may need relaxing, and only a recorded bet has stated odds against which a stake can be sized (stake and odds). Progressive constraining is the same split applied over time: commit what recurs, keep the rest open. Without the layer, the routine records a commitment as a fact, and the condition that would retire it is invisible until it happens.

3. Retention. Storage consumes the same split for a different decision. Commitments are ground truth and must be kept; derivations can be regenerated and belong in a cache whose persistence must pay for itself (recompute cost). Design rationale keeps the premises its interpreter cannot regenerate (unregenerable premises), and a superseded choice keeps a historical witness while a refuted belief loses standing (supersession). Each rule needs the status recorded. Without it, a store either keeps everything as truth or discards a commitment as if it could be recomputed.

4. Evaluation reach. An acceptance gate consumes what its check can establish. Checks differ in how cheaply they verify correctness (oracle strength), and automation extends as far as verification does (automation boundary). When the layer records which qualities a check discriminates, unchecked qualities stay marked as guessed. Without it, a pass reads as acceptance of everything, and under the conditions in weakly discriminated qualities the accepted population is enriched in what the oracle measures and not in what it does not.

5. Method improvement. Reflective investment consumes criticism of the method, not of its outputs. A machinery improvement can outrank immediate learning because every later episode reuses it (investing in machinery), but machinery persists only by its warrant (warrant, not position). Without the layer the method's choices are not marked as decided, so nothing invites criticism of them, and learning stays inside the method's fixed decomposition (inherits its mistakes).

6. Application. Theory building is itself a kind of knowledge work the layer supports. A theory builder states theories in localized units, acts on them, criticizes what they say and lets the result shape the next round. Commonplace is the first worked case: it builds such a system and tests whether it learns (research companion). Without the layer, a theory's status lives in the actor's context and ends with the episode.

Testable form

General prediction. A knowledge tool without the layer fails in the ways the role table names. It is refuted by a tool that lacks the layer and does not show a role's failure in that role.

Record-level prediction (role 2). A routine whose resolutions were classified as derivable or decided at conversion can name its relaxing trigger in advance. A routine adopted on cost alone cannot. A resource-rational rival, in which routines are adopted because they are cheap and good enough, predicts no difference in advance naming, because cost does not record which resolutions are bets.

The prediction was abstracted from one pair in Commonplace's record; the pair is its origin, not a test. The proposal behind the artifact-run engine (ADR 113) called its premise, that harnesses would bring their own code scheduling, "the operator's expectation, not an observed fact", and marked the machinery that existed only because code could not launch sub-agents as "the part expected to be discarded" (the archived proposal "Code-scheduled workflows", which ADR 113 adopted). ADR 118 retired the engine when headless modes let code launch workers. The covered_by tag mark (ADR 026) deferred adoption and named no condition for removal; the one head that declared it dropped it twice, and ADR 090 retired it because nothing consumed the relation. A test needs conversions recorded after this note, with their classification made at conversion.

Weak point

Prioritisation needs preferences as well as beliefs: what a branch is worth depends on what the actor wants, not only on what the branch will produce. The layer covers beliefs, including return guesses relative to a given objective. It does not choose the objective. A change of objective counts as improvement only against a level outside both objectives (objective revision), so goal change stays licensed from outside the work. Role 1 is therefore partial: the layer improves allocation under a goal, not the goal.

Scope

  • The conditions are conjunctive: open-ended, bounded, text-mediated. Work with a fixed candidate set needs no prioritisation role; an actor whose context holds the whole work needs no retention role.
  • The note asserts that the roles share one layer. It does not assert that Commonplace's machinery is the right implementation of the layer, or that support built for theory building generalizes to other kinds of knowledge work; that is the separate architecture-level conjecture in reusable commitment support.
  • The six roles may not be exhaustive, and roles 2 and 3 consume the same split. They stay separate because their consumers and failures differ: conversion decides what to freeze, storage decides what to keep.

Relevant Notes: