Examples: the unit-of-training article, 2026-09-05
Every edit made to kb/articles/the-software-house-as-the-unit-of-training.md
by the operator-led method, with the operator's words where the operator gave
them, the agent's diagnosis, and the text before and after. Run started from
the procedure in README.md; the whole-article read (step 2) was
done first and reported before any prose work.
1. Training vs. learning: state the relation once
- Commit: see log (
Workshop:trailer) - Kind: Unintroduced term (undrawn distinction between two alternating terms)
- Operator's verdict: "ok - fix both", accepting the whole-article read finding: the article uses training (title, regime name, "What counts as training") and learning ("theory-mediated learning", the unit-of-learning note) without stating their relation, and "What is trained" cites the unit-of-learning note as if the terms were interchangeable.
- Diagnosis: The duality survives from the article's origin as the
"training doctrine" half of the split (
292b0dd0); it is deliberate but undrawn — training is the regime imposed, learning the capacity change it produces. Fixed by stating the relation once, at the point where the clash first bites: the section that puts the unit-of-learning note under the unit-of-training title.
Before:
What is trained: the whole house
The deployed system, not the model alone, is the unit of learning because several components jointly determine its behaviour.
After:
What is trained: the whole house
In this article training names the regime — what production is arranged to do to the house — and learning names the retained change in capacity that results. Both have the same unit. The deployed system, not the model alone, is the unit of learning because several components jointly determine its behaviour.
2. TL;DR: restore the sample-efficiency claim
- Commit: see log (
Workshop:trailer) - Kind: Lost claim after restructure (partial: one of three hypotheses dropped from the TL;DR)
- Operator's verdict: "ok - fix both", accepting the whole-article read finding: the TL;DR's last sentence names only "diagnosis and transfer", while the hypotheses section asks three questions (influence, transfer/recovery, sample efficiency) and the frontmatter description names all three ("causal influence, transfer, and learning cost").
- Diagnosis: The TL;DR silently narrowed the empirical hypothesis to two of its three questions. Fixed by naming the third; sample efficiency is the article's own term for it (the sample-efficiency hypothesis).
Before:
dependencies change. Whether this improves diagnosis and transfer compared with other uses of the same evidence is an empirical hypothesis.
After:
dependencies change. Whether this improves diagnosis, transfer, and sample efficiency compared with other uses of the same evidence is an empirical hypothesis.
3. TL;DR: introduce the mediator, label the example
- Commit: see log (
Workshop:trailer) - Kind: Unintroduced term; unlabelled example
- Operator's verdict: "1 - apply" (phase-2 sweep candidate 1)
- Diagnosis: Mediator connects to the subtitle's theory-mediated learning but the TL;DR never says what mediates what; the checking sentence is an example without being announced as one.
Before:
The house's program theory is its understanding of the software's purpose, organization, and how to handle new requests. The proposed mediator is an explicit project theory: one possible written carrier of that understanding, stating design commitments, causal assumptions, and invariants. A house that explains why some files need product checks may adapt its checking policy when dependencies change.
After:
The house's program theory is its understanding of the software's purpose, organization, and how to handle new requests. The regime's proposed mediator — what experience revises and later decisions consult — is an explicit project theory: one possible written carrier of that understanding, stating design commitments, causal assumptions, and invariants. For example, a house whose theory explains why some files need product checks can adapt its checking policy when dependencies change.
Skipped: the premise boundary paragraphs (sweep candidate 2)
- Operator's verdict: "2 skip - building the seed is from a different cycle"
- Lesson: The agent's proposed why for "This boundary applies during the run" (that a human-built seed does not violate it) read the seed into the wrong cycle; the seed belongs to a different cycle than the run the boundary governs. A verdict-without-its-why fix that supplies the wrong why is worse than the compression; when the why is inferred rather than stated nearby, flag it for the operator instead of writing it in.
4. Derived indexes are an exception; the scope sentence gets its own move
- Commit: see log (
Workshop:trailer) - Kind: Verdict without its why; unsignposted role
- Operator's verdict: "3 indexes are kind of on the boundary between
symbolic and distributed parametric - I think we can state that they are an
exception" (phase-2 sweep candidate 3); follow-up after the first commit:
"actually algorithms don't need to be pinned" — the pin covers
distributed-parametric models only, so the sentence keeps only the embedding
models pinned and names the building algorithms as ordinary revisable
machinery. This corrects a claim the pre-edit article text also made
("under pinned algorithms and embedding models"). Second follow-up: "maybe
double check about the pinned algorithms - the main idea is that the house
can revise both natural language and symbolic forms of its own definition -
this needs to be clear" — the paragraph now states that rule by
representational form before placing indexes as the boundary case. Third
follow-up: "the support for the indexes exception is that it is a
traditional machinery used since the first computers and it is fast - in
general it is closer to symbolic than distributed parametric in its
operation (but is still partially global) - check this with the definitions
we have in this KB". Checked against
kb/notes/definitions/representational-form.md: the operational-closeness claim maps onto the assigned-consequences axis, but the definition lists dense-vector indexes as distributed-parametric outright and handles the vector-store case as a mixed artifact split by operative part, so "sit on the boundary" was replaced by the split — symbolic in operation, distributed-parametric values from a pinned model over revisable records, hence no new learned parameters enter during regeneration. - Diagnosis: The index sentence stated a permission with no reason. The agent proposed "looks like an exception but is not one"; the operator corrected the frame: indexes sit on the boundary between symbolic and distributed-parametric forms, so state them as a genuine exception to the pin, justified because only the input records change. The article's reading instruction (experiments run before a complete house exists) was also buried as the paragraph's last sentence; it now stands alone.
- Frame correction: The agent's rewrite denied the exception to defend the premise; the operator accepted the exception instead. Lesson: when a rule has a borderline case, saying "this is an exception, and here is why it is safe" can be more honest than arguing the case back inside the rule.
Before:
Training changes its surrounding state and machinery while keeping distributed-parametric models fixed. Derived indexes may be regenerated from changing records under pinned algorithms and embedding models. The experiments below test learning mechanisms in bounded components before a complete automated house exists.
After:
Training changes its surrounding state and machinery while keeping distributed-parametric models fixed. The rule is by representational form: the house may revise both the natural-language and the symbolic forms of its own definition; only its distributed-parametric models are pinned. Derived indexes are an exception, and a safe one. An index is traditional machinery, in use since the first computers, and symbolic in its operation: construction and lookup are defined, fast, and testable, even though its content is derived from the whole record set. Where the values are embedding vectors they are distributed-parametric, but they come from a pinned embedding model applied to revisable records, so regenerating the index during the run introduces no new learned parameters.
The experiments below do not wait for a complete automated house. They test its learning mechanisms in bounded components.
5. The mechanism paragraph: unpack "separately testable from the outcome"
- Commit: see log (
Workshop:trailer) - Kind: Verdict without its why
- Operator's verdict: "apply the rest" (phase-2 sweep candidate 4)
- Diagnosis: A three-coordinate-clause mechanism sentence, and a closing verdict ("separately testable from the outcome of coherent modification") whose reason the reader could not reconstruct: mechanism claims are about how an update came to be, so they can differ between houses whose outcomes match.
Before:
The proposed mechanism is that the explanation directs diagnosis toward the consumer path, the failure challenges the assumption that the input list is exhaustive, and a revised account identifies other files that need checks. Testing a second, untouched snippet asks whether that revision guides a later decision. These are claims about how updates are produced, separately testable from the outcome of coherent modification.
After:
The proposed mechanism has three steps: the explanation directs diagnosis toward the consumer path; the failure challenges the assumption that the input list is exhaustive; and a revised account identifies other files that need checks. Testing a second, untouched snippet asks whether that revision guides a later decision. These are claims about how the house produces its updates, and they can be tested separately from the outcome: two houses can both end up modifying coherently while only one got there by revising an explanation.
6. "These episodes" and "persistence alone"
- Commit: see log (
Workshop:trailer) - Kind: Unintroduced term; verdict without its why
- Operator's verdict: "apply the rest" (phase-2 sweep candidate 5)
- Diagnosis: Episodes appeared as an anaphor with no antecedent; "persistence alone is insufficient" trailed the requirement it qualifies. Reordered: introduce episodes as episodes of production, state the insufficiency with its reason, then the requirement.
Before:
Production implements current requirements and supplies evidence through operating consequences and retained history. These episodes may also be replayed, simulated, or augmented. Evidence of training requires experience to cause a retained change in the house's capacity on later requirements. Persistence alone is insufficient: every product edit changes the starting conditions of later work.
After:
Production implements current requirements, and each episode of it supplies evidence: operating consequences and retained history. Episodes may also be replayed, simulated, or augmented. Evidence of training requires more than a change that persists — every product edit changes the starting conditions of later work. It requires experience to cause a retained change in the house's capacity on later requirements.
7. Introduce continuations
- Commit: see log (
Workshop:trailer) - Kind: Unintroduced term
- Operator's verdict: "apply the rest" (phase-2 sweep candidate 6)
- Diagnosis: Continuation first appeared here carrying the whole experimental design, and recurs in the hypotheses section ("paired continuations"). Introduced at first use: copies of the house that resume work from identical product snapshots.
Before:
After correcting the failing edit in the example, give two continuations identical product snapshots. Retain the revised explanation and checking policy in one; restore their earlier versions in the other.
After:
After correcting the failing edit in the example, run two continuations: copies of the house that resume work from identical product snapshots. Retain the revised explanation and checking policy in one; restore their earlier versions in the other.
8. The embodied-in-code comparison, said in ordinary words
- Commit: see log (
Workshop:trailer) - Kind: Compressed description
- Operator's verdict: "apply the rest" (phase-2 sweep candidate 7; the agent flagged "embody different lessons" as its interpretation of "equivalent current functionality" and the operator accepted)
- Diagnosis: "Compare implementations with equivalent current functionality" did not say what is compared with what or why equivalence matters. Reordered so the patch-is-both-production-and-training point leads, and the comparison is spelled out: implementations that currently behave the same but embody different lessons.
Before:
When learning is embodied in product code, compare implementations with equivalent current functionality on later maintenance tasks. A patch can be both production and training, and the acquired capability can remain specific to one product. The comparison identifies its contribution to later work.
After:
Learning can also be embodied in product code: a patch is then both production and training. To measure what it taught, compare implementations that currently behave the same but embody different lessons, on later maintenance tasks. The acquired capability can remain specific to one product; the comparison identifies what it contributes to later work.
9. Define the training lineage at first use
- Commit: see log (
Workshop:trailer) - Kind: Unintroduced term (two sites)
- Operator's verdict: "apply the rest" (phase-2 sweep candidate 8; the agent flagged its definition for checking and the operator accepted)
- Diagnosis: Training lineage appeared first in the continuity bullet and again in the Bitter Lesson section as "that training lineage" with no local antecedent, never defined. Defined at first use as the history of retained changes made while the models stay pinned; the second site now says "the training lineage".
Before (continuity bullet):
outside a model checkpoint. Their effective use may still depend on the model, so replacing it requires revalidation and lies outside the fixed-model training lineage.
After (continuity bullet):
outside a model checkpoint. Their effective use may still depend on the model, so replacing the model requires revalidation and falls outside the training lineage — the history of retained changes made while the models stay pinned.
Before (Bitter Lesson): "that training lineage; the seed may be human-built"
After (Bitter Lesson): "the training lineage; the seed may be human-built"
10. Introduce source observations
- Commit: see log (
Workshop:trailer) - Kind: Unintroduced term
- Operator's verdict: "apply the rest" (phase-2 sweep candidate 9)
- Diagnosis: Source observations separates the facts supplied to every treatment from the observations each continuation makes itself — load-bearing for the whole treatment table, unexplained. Introduced in place: the facts observed before the continuations start, which each treatment retains in its own form.
Before:
Use paired continuations with the same starting product, fixed models, tools, source observations, request sequence, and resource ceilings. Vary how observations are retained, then let each continuation learn from its own actions and observations. Compare four treatments:
After:
Use paired continuations with the same starting product, fixed models, tools, request sequence, and resource ceilings, and the same source observations: the facts observed before the continuations start, which each treatment retains in its own form. Vary that retained form, then let each continuation learn from its own actions and observations. Compare four treatments:
11. Unpack "the retained learning pathway"
- Commit: see log (
Workshop:trailer) - Kind: Verdict without its why
- Operator's verdict: "apply the rest" (phase-2 sweep candidate 10)
- Diagnosis: Why the reconstruction sentence was there, and what "the retained learning pathway" meant, were both compressed: the treatments differ in what survives to the next decision, not in what the model can think in the moment.
Before:
their explanations. All may revise executable machinery and its tests. A model can reconstruct an explanation while reasoning in any treatment. The comparison concerns the retained learning pathway.
After:
their explanations. All may revise executable machinery and its tests. In any treatment the model can still build an explanation while reasoning; no treatment forbids thinking in theories. What differs is what survives to the next decision: the comparison concerns what each treatment retains, not what the model can construct in the moment.
12. Map the two kinds of change onto the two histories
- Commit: see log (
Workshop:trailer) - Kind: Unintroduced term
- Operator's verdict: "apply the rest" (phase-2 sweep candidate 11)
- Diagnosis: History arrived mid-paragraph ("both histories") and the mapping from the two kinds of change to the two histories was implicit. Now each kind of change is introduced as its own request history.
Before:
Test two kinds of change separately. Adding another configured exporter input preserves the initial account of direct inputs. Adding indirect includes breaks its assumption that the list is exhaustive.
After:
Test two kinds of change separately, each as its own request history. In the first history, adding another configured exporter input preserves the initial account of direct inputs. In the second, adding indirect includes breaks its assumption that the list is exhaustive.
13. Influence versus explanatory guidance
- Commit: see log (
Workshop:trailer) - Kind: Verdict without its why
- Operator's verdict: "apply the rest" (phase-2 sweep candidate 12)
- Diagnosis: "That establishes influence; following a misleading instruction can also produce errors" compressed the contrast: influence is not explanatory guidance, because a model that merely obeyed the text would show the same error pattern. The reason is now stated.
Before:
That establishes influence; following a misleading instruction can also produce errors. Evidence of useful explanatory guidance additionally requires correct handling of unstated consequences and appropriate revision under contradiction.
After:
That establishes influence, but influence alone is not explanatory guidance: a model that merely followed the text as an instruction would produce the same errors. Evidence of useful explanatory guidance additionally requires correct handling of consequences the text does not state, and appropriate revision when evidence contradicts it.
14. Name the misdirection before measuring it
- Commit: see log (
Workshop:trailer) - Kind: Unintroduced referent
- Operator's verdict: "apply the rest" (phase-2 sweep candidate 13)
- Diagnosis: "Measure that misdirection" pointed at something the previous sentence had not named. The reversal is now called what it is — the theory misdirecting the house — before it is measured.
Before:
Its initial advantage may disappear or reverse when assumptions break. Measure that misdirection separately from recovery after revision: rapid recovery can erase the initial loss in the whole-run score.
After:
Its initial advantage may disappear or reverse when assumptions break: the theory then misdirects the house. Measure that misdirection separately from recovery after revision, because rapid recovery can erase the initial loss in a whole-run score.
15. The null-result limit, said plainly
- Commit: see log (
Workshop:trailer) - Kind: Compressed description
- Operator's verdict: "apply the rest" (phase-2 sweep candidate 14)
- Diagnosis: "Leaves open reconstruction of equivalent understanding from other state" compressed the equivalent-reconstruction point the conjecture article states plainly. Unpacked in the same plain form.
Before:
A null intervention result leaves open reconstruction of equivalent understanding from other state.
After:
A null intervention result is inconclusive: removing the designated theory text may change nothing because the house reconstructs the same understanding from other retained state.
16. Signpost the exporter example
- Commit: see log (
Workshop:trailer) - Kind: Unsignposted role (transition)
- Operator's verdict: "apply the rest" (phase-2 sweep candidate T1)
- Diagnosis: The section titled "Why this is theory-mediated learning" opened "Consider a hypothetical release exporter" with nothing saying the example carries the argument.
Before:
Why this is theory-mediated learning
Consider a hypothetical release exporter. It builds a deployment manifest
After:
Why this is theory-mediated learning
The claim is easiest to show on a concrete case. Consider a hypothetical release exporter. It builds a deployment manifest
17. Signpost the generality paragraph's four jobs
- Commit: see log (
Workshop:trailer) - Kind: Unsignposted roles
- Operator's verdict: "apply the rest" (phase-2 sweep candidate T2)
- Diagnosis: Four sentences doing four jobs — concession, positive evidence, reporting rule, negative evidence — with nothing marking the first as a concession. The concession is now announced and the reporting rule attached to the evidence sentence it qualifies.
Before:
Some objectives, authority boundaries, hard dependencies, runtimes, and trusted kernels may remain fixed. Evidence for generality comes from adapting as the house takes on new kinds of work, including changes to its initial responsibilities. Evaluations must report the work and period examined. Repeated dependence on people to supply a missing ontology, decomposition, or evaluator counts against the proposed learning capacity.
After:
Generality does not mean everything is trainable: some objectives, authority boundaries, hard dependencies, runtimes, and trusted kernels may remain fixed. Evidence for generality comes from adapting as the house takes on new kinds of work, including changes to its initial responsibilities; evaluations must report the work and period examined. Repeated dependence on people to supply a missing ontology, decomposition, or evaluator counts against the proposed learning capacity.
18. Lead into the two governance functions
- Commit: see log (
Workshop:trailer) - Kind: Unsignposted role (transition)
- Operator's verdict: "apply the rest" (phase-2 sweep candidate T3)
- Diagnosis: The bold Admission and Credit assignment definitions arrived with no sentence saying these are the functions that do the governing the previous paragraph described.
Before:
Admission decides which changes take effect. Credit assignment decides
After:
Two functions do that governing. Admission decides which changes take effect. Credit assignment decides
19. Flow pass: deduplicate the pin and split the index paragraph
- Commit: see log (
Workshop:trailer) - Kind: Redundancy after revision; paragraph overload (flow-pass findings F1 and F4)
- Operator's verdict: "ok - go with your recommendation" on the flow-and- redundancy check run after the sweep
- Diagnosis: The run's edits left the pin stated twice in adjacent sentences ("while keeping distributed-parametric models fixed" next to "only its distributed-parametric models are pinned"), and the index exception had grown to take over a paragraph whose topic sentence is the seed. The redundant clause is dropped and the exception gets its own paragraph.
Before:
Training changes its surrounding state and machinery while keeping distributed-parametric models fixed. The rule is by representational form: the house may revise both the natural-language and the symbolic forms of its own definition; only its distributed-parametric models are pinned. Derived indexes are an exception, and a safe one. An index is traditional machinery,
After:
Training changes its surrounding state and machinery. The rule is by representational form: the house may revise both the natural-language and the symbolic forms of its own definition; only its distributed-parametric models are pinned.
Derived indexes are an exception, and a safe one. An index is traditional machinery,
20. Flow pass: restore the antecedent of "this boundary"
- Commit: see log (
Workshop:trailer) - Kind: Lost antecedent after revision (flow-pass finding F5)
- Operator's verdict: "ok - go with your recommendation"
- Diagnosis: The training/learning edit (entry 1) changed the section's lead to talk about the unit, leaving "this boundary" in the next paragraph pointing at nothing.
Before: "For coherent product evolution, this boundary must include the current product,"
After: "For coherent product evolution, the unit's boundary must include the current product,"
21. Flow pass: name the production boundary and the distinction
- Commit: see log (
Workshop:trailer) - Kind: Unintroduced referents (flow-pass finding F6; a narrower fix of skipped candidate 2 that leaves the "applies during the run" sentence untouched)
- Operator's verdict: "ok - go with your recommendation"
- Diagnosis: Paragraph 1 ended "from outside that boundary" with no boundary named; paragraph 2 opened "The distinction turns on" with no distinction named. The boundary is now named at first mention and the second paragraph opens by saying what it decides.
Before:
supply requirements, facts, feedback, and acceptance judgments about visible behaviour from outside that boundary.
The distinction turns on the decision supplied: users may say what the product should do or report what it did.
After:
supply requirements, facts, feedback, and acceptance judgments about visible behaviour from outside the production boundary.
Which side of the boundary a contribution falls on turns on the decision supplied: users may say what the product should do or report what it did.