ADR-029: Build the conditional, not procedural memory
- Status: Accepted and BUILT (2026-09-02, §9) — REOPENED 2026-09-03 (§10), BUILT the same day (§11), and released STANDALONE (§12): the Procedural vertical this ADR declined ships as 80 questions at headroom +0.81. The co-design hold in §11e is lifted; the counterparty is notified once the work is usable, not asked in advance. §9 and §10e stand; §6a and §11e's closing do not, and neither does the title — it is kept because retitling an ADR hides the reversal.
- How it got here, which is the useful part: an earlier draft proposed narrowing ADR-027's exclusion and was put to an adversarial review that refuted three of its four decisions (§8). The maintainer then set aside the one political objection, leaving the two technical ones — and re-examining those produced a smaller, better design that dissolves all of them (§7b).
- Date: 2026-08-31, rewritten 2026-09-01 after review.
- Relates to: ADR-026 §8, which is where the exclusion actually lives; ADR-027 §1; ADR-028, whose acceptance criterion any new vertical must clear.
1. THE EXCLUSION IS BILATERAL. THIS PROJECT CANNOT NARROW IT.
ADR-026 records it twice, and both times as an agreement rather than a preference:
No Procedural vertical — consumer-side by agreement (needs tools; agentic). — ADR-026 §8, line 792
Procedural memory stays consumer-side (it is agentic and needs tools); it is out of scope for AgentEval and this family. — ADR-026, line 373
Procedural memory has an owner: the consuming project, by agreement. The first draft of this ADR asserted the opposite — that if the exclusion stood, procedural memory would have "no owner anywhere in the product" — and used that as the cost that made narrowing look like the only responsible answer. That claim was false, and it was the load-bearing one.
The correct move is therefore not to amend ADR-027 §1 on one party's authority. It is to put the finding in §2 to the counterparty and ask whether the agreement should be reopened, because the premise it was struck on is the part that no longer holds. That question has been sent (see the 2026-09-01 correspondence); their answer is an input this repository cannot contain, which is precisely the condition under which a decision stays open rather than being guessed.
2. THE FINDING THAT IS WORTH SENDING
ADR-027 §1 excluded procedural memory because it "needs tools, a live agent loop and observed outcomes, so it cannot be a static corpus." The mechanism half of that reason does not hold, and it is worth the counterparty knowing:
IToolCapableAgentalready hands an agent tool schemas and returns a trace of emitted calls.CanaryTool.Executeis optional, andEvidenceFidelity.IntentToActalready distinguishes a call the model emitted from one that executed.
So an agent can be handed tools, asked to carry out a procedure it learned sessions earlier, and scored on the emitted call sequence — no sandbox, no side effects, no state threading. Observed outcomes are needed to test whether a system improved; they are not needed to test whether it remembered.
One structural gap is also real and unclaimed: nothing in the family has a conditional. Checked across all nine generators. Whether that deserves a shape, and where, is §5.
3. WHAT THE FAMILY ALREADY COVERS, WHICH THE FIRST DRAFT MISSED
The first draft listed five properties as evidence for a new vertical. Four of them have homes:
| property | existing home |
|---|---|
| ordered steps | episodic/list-order, temporal/occurrence-order — and pairwise-order credit already ships (PairwiseOrderAccuracy, TypedMemEvalRunner.cs) |
| retired steps | forgetting |
| partial revision | semantic/current-value, bitemporal/correction-depth |
| preconditions | unhomed |
| conditional branch | unhomed — the only one nothing covers |
ADR-027 §11.1 sets the standard this must be judged against: "A three-shape vertical whose shapes each have a plausible existing home is a weak vertical." By that standard the honest scope is at most two candidate shapes inside existing verticals, not a tenth vertical.
4. THE CONTRADICTION AT THE HEART OF THE FIRST DRAFT
It argued distinctness and buildability with premises that negate each other:
- Distinctness (§2): a procedure is different because its order is causally required — violating it is an error, not merely a wrong answer.
- Buildability (§3): it survives V2 by borrowing Temporal's technique — invented steps whose order is arbitrary and stated.
Arbitrary-and-stated is the negation of causally-required. ADR-026 §5.2 already settles which one the family permits: ordered gold must be "arbitrary (V2) and derivable only from session sequence, never from narrative logic (no 'starter before dessert' orderings)." Causal necessity is narrative logic. So the property that would make a procedural vertical distinct is the property the family's own construction rule forbids — and once removed, what remains is Episodic list-order.
This contradiction must be resolved in writing before any design proceeds. Either the order is arbitrary (and the vertical is not distinct) or it is causal (and it cannot pass V2). No third option has been articulated.
5. TWO MORE REASONS THE EVIDENCE PREDICTS REJECTION
V2 is documented blind in exactly this question class. tools/name-collision-audit.json,
measured 2026-08-30 — the day before the first draft — records: "V2 scores only a leak that AGREES
with gold… Harm therefore concentrates in ORDERING questions." A procedural corpus is ordering
questions end to end. And inventing names does not reach the leak: in a procedure the prior is
carried by the verbs (back up before migrating, test before deploying), not by the entity names,
so a non-referential noun bank buys nothing.
No discrimination argument was offered, and ADR-028 predicts a poor one. Acceptance now rests on
V1 − V9 ≥ 0.15. ADR-028 §7.4 names the exact structural cap a procedural question inherits —
near-closed-choice forms "where the question names the entity it asks about"
(temporal/occurrence-order 0.05, bitemporal/belief-at-instant 0.11,
episodic/participant-attribution 0.20). "What are the steps of the Verrin changeover?" names its
own procedure and is lexically self-retrieving, which drives V9 up and headroom toward zero. The
family's nearest existing conditional-shaped join, conjunction/order-then-value, shipped saturated
at V9 15/15, headroom 0.00, for its whole life.
6. WHAT IS ACTUALLY DECIDED HERE
6a. ADR-026 §8's exclusion stands unchanged. This project does not narrow it.
6b. The mechanism premise in ADR-027 §1 is recorded as refuted (§2). That is a finding about the reason, not a change to the decision.
6c. The conditional-branch gap is recorded as real and unhomed (§3). If anything proceeds, it proceeds as at most one or two shapes inside existing verticals, judged by ADR-028's discrimination floor like everything else.
6d. The §4 contradiction is open and blocking. No design work should start before it is resolved in writing.
7. What is asked of the counterparty
Whether the agreement recorded in ADR-026 §8 should be reopened, given that the mechanism it was struck on (§2) does not hold. Nothing else in this ADR needs their answer; nothing in this ADR proceeds without it.
7b. RESOLUTION — build the CONDITIONAL, not "procedural memory"
Added 2026-09-01 after the maintainer set aside the ownership objection ("we can decide this alone; the other party is simply not ready to use it, but if it works, it will"). That removes §1 as a blocker. It does not remove §4 or §8's boundary refutation, which are technical and decide whether the thing would WORK — so they were re-examined rather than waved through, and one of them turns out to point at a much better design.
The gap, verified rather than asserted
Of the five properties §3 claimed were distinct, four have existing homes. One does not, and it is now measured: across all 470 shipped questions, not one requires the model to resolve a conditional. Five questions contain the word "if" or "unless", and in every case the conditional is quoted PAYLOAD, never the thing being asked:
"…it sticks unless you lift before you push. Was that me or you?" — attribution "…the warranty lapses if a service is missed?" — context recall
Nothing in the family tests "if X, then what?"
Why this is the whole opportunity, and procedural memory is not
Scoping to the conditional dissolves every objection at once, which is the sign it is the right shape rather than a smaller version of the wrong one:
| objection | why it no longer applies |
|---|---|
| §1 ownership (ADR-026 §8) | A conditional-resolution shape does not claim to BE procedural memory, so the agreement is never touched — regardless of whose call it is |
| §8 naming | No invented vocabulary, no over-claim; it is a shape, not a memory type |
| §4 contradiction | Never arises. The distinctness claim is not about causally-required order at all |
| §8 boundary refuted by construction | The new rule is not an ordering rule, so it cannot collapse into Temporal's membership test |
And §4's contradiction turns out to have had a resolution neither the draft nor the review named, worth recording even though it is no longer needed: INVENTED CAUSALITY. "The Vreskade check must run before the Quorlory sync, because Quorlory consumes Vreskade's output" is causally required inside the fiction and unguessable from outside it, because no model holds a prior about Vreskade. ADR-026 §5.2 bans orderings a model can INFER, not causality as such. That would have rescued the ordering argument; it is simply no longer the argument being made.
The decision
One shape, conditional-branch, inside Conjunction. Not a tenth vertical.
Conjunction is already defined as questions no single memory type can answer, and resolving a branch
is find the procedure -> find the condition's state -> select the branch — structurally the same
join as alias-then-count and order-then-value. ADR-027 §11.1's standard ("a vertical whose shapes
each have a plausible existing home is a weak vertical") is therefore satisfied by NOT making it a
vertical.
The boundary rule is mechanically checkable and cannot collapse into an existing vertical:
The answer must CHANGE depending on a condition stated in the haystack. A question whose answer is the same under both branches is not a conditional question and is refused at generation.
That is enforceable the way check_temporal refuses a digit, and no existing vertical can express it
— which is exactly why the gap exists.
The one real cost
Conjunction ships 50 questions across three shapes (15/15/20). A fourth at parity puts every shape near 12, below the ~15 line at which this family's own guidance says a shape supports diagnosis rather than a claim. So Conjunction grows to ~65 — one declared size moves, and the consuming project is told before it does.
What is NOT decided
Procedural memory as a MEMORY TYPE remains out, and ADR-026 §8 stands untouched. If it is ever revisited, §8's naming finding applies: it should be called Procedural, scoped explicitly to the memory layer, because that is what the consuming engine already calls it and inventing a word to dodge an over-claim helps nobody.
8. What the review refuted, kept because it is the useful part
The first draft proposed four decisions. Three were refuted and the fourth declined to ratify:
| proposed | outcome |
|---|---|
| Narrow ADR-027's exclusion to the enacted half | Refuted — the exclusion is ADR-026's and bilateral; the "no owner" premise was false |
| Name the vertical Protocol, not Procedural, to avoid claiming the excluded category | Refuted — Prospective is already named for a construct it only partly delivers, so the standard being applied is not one this family holds |
| Boundary rule: ≥3 ordered elements and mention order ≠ execution order | Refuted by construction — a question satisfying both was built from Temporal's own machinery and check_temporal returned zero failures. Rule (b) is Temporal's membership test verbatim (narration_inversions < 1), not a boundary against it |
| Make it a tenth vertical | Not ratified — it answers §6.3 while the boundary rule is still open, and in this family a vertical is an isolation claim, enforced mechanically rather than in prose |
The lesson worth keeping: an amendment that moves scope should first establish whose scope it is. The first draft spent its length on whether the split was intellectually sound and never checked the one line that said who owns the decision — a line already present in this repository, and read earlier the same day.
9. BUILT, 2026-09-02 — conjunction/conditional-branch
Fifteen questions, tme-cnj-051 through -065. Conjunction 50 → 65, the one cost §7b declared.
9a. The construction
Two gold sessions, and which one is lexically reachable from the question is the whole design:
| names | reachable from the question? | |
|---|---|---|
| rule (semantic) | the attribute and the condition | yes — the question names the attribute |
| state (episodic) | the condition only | no — you must read the rule to learn what to look for |
"How the usual courier gets picked: the Kelvaryn access open means Nettleford Runners, closed for resurfacing means Marlow Carriage, anything else means Bardsey Logistics." "Logging it — the Kelvaryn access is open as of now." "Going by the standing rule, which usual courier is in force?"
That is a real second hop, and it is the property order-then-value had to be rebuilt to obtain —
that shape shipped saturated at V9 15/15 because one gold session named both the anchor and the
answer.
9b. Measured
| V1 / V8 / V9 | 15/15 · 15/15 · 0/15 |
| headroom (reachable) | 1.00 (1.00) — the widest in the family |
| V2 / V3 | 15/15 · 15/15 |
| coverage | 0.50 |
The structural prediction and the measurement agree exactly. Before probing, on the shipped corpus: the rule is in BM25's top-5 for 15/15 questions, the state for 0/15, both for 0/15 — and 0/15 is precisely what V9 returned. Coverage of 0.50 is 1-of-2 on every question by construction, so the echo knob has nothing to move here and 0.50 is a structural floor that happens to sit on the band's lower edge. A band failure on this shape means the rule stopped being retrievable; it is never a reason to widen the band.
9c. Two things checked because the numbers flatter us
V9 = 0/15 is an extreme value, and extreme values are wiring faults until proven otherwise. Read back, the reference model retrieves the rule, sees three named branches and declines: "the conversations don't say whether the Peskadd permit is granted, still pending, or something else." Calibrated abstention, not a broken arm.
The rule session names all three outcomes, so a reader holding the rule and missing the state is
choosing among three named candidates — a chance floor of 1/3 that closed_choice_k cannot see,
because it parses the question and the candidates live in the haystack. This family has been
caught by unpublished chance floors twice (ADR-028 §7.4; V2's 69 unearned passes), so it is now a
published column: chance_floor, chance_floor_reason, v9_above_chance. This shape reads
−0.3333 — the baseline is below chance because it does not guess.
The field is named chance_floor, not chance_floor_given_rule. A generic instrument keyed on a
shape-specific name is the applied-once defect; any future shape whose haystack hands the reader a
closed set publishes the same two fields and gets the same column.
9d. What building it found in a neighbour
alias-then-count's decoy count was hardcoded at 2 while gold is 2 + (index % 3), so on every
third question the decoy designation carried exactly gold's count — five of fifteen. On those
five, a system that resolves the alias to the wrong place still produces the right answer, which
is the precise failure the shape exists to detect. The join was not load-bearing on a third of it.
It surfaced as a V3 leak: with gold ablated the model answered "two deliveries were taken at the
Peverel building" — a different designation, the same count, and require_distinctive cannot
separate them because the distinctive token is the digit they share. V3 read 50/50 before this on
a corpus where five questions could leak. The pass was luck, not evidence.
Fixed, and guarded: the decoy's count now differs from gold's, and check_conjunction counts
delivery sentences per designation and refuses equality. V3 is 65/65; alias-then-count
coverage moved 0.3356 → 0.4344, toward the band.
The guard's own first version over-counted — it matched any non-gold session naming the decoy, including filler that states a designation link and records no delivery. It reported 4 against a real 2. Worth recording because an over-counting guard invites being loosened, and the loosening would have taken the real check with it.
9e. The name bank, and a trap closed before it fired
CONDITIONS is registered in tools/audit_name_collisions.py — the same exposure MILESTONES had,
where four of six names were real and the shape asks which came first. All six audited live: zero
real, zero unparsed, six invented. Confirmed positively rather than read off an absence.
Registering it also exposed the extractor: collect() did head(str(entry)), which on a nested
bank produces "('the Kelvaryn access', ('open', ...))" and sends that to the model as a name.
The model says REAL: no — correctly — and the real name inside the string is silently exonerated.
No shipped bank was nested when this was written; CONDITIONS is the first. The unpacking is
fixed and a punctuation check now raises rather than auditing a string no corpus contains.
10. REOPENED, 2026-09-03 — build Procedural. The technical objection was withdrawn in §7b.
§6a said this project does not narrow ADR-026 §8's exclusion. That is reversed here, on the maintainer's authority, and the reasoning that got here is short because most of it is already above.
10a. The blocker I kept quoting was solved in this same document
§6d called the §4 contradiction "open and blocking": a procedure is distinct because its order is causally required, and ADR-026 §5.2 forbids orderings a model can infer from narrative logic.
§7b resolves it. "The Vreskade check must run before the Quorlory sync, because Quorlory consumes Vreskade's output" is causally required inside the fiction and unguessable outside it, because no model holds a prior about Vreskade. §5.2 bans orderings a model can INFER, not causality as such. §7b then set the resolution aside with the words "it is simply no longer the argument being made" — because the scope had already narrowed to the conditional, not because it fails.
I quoted §6d as a live blocker after §7b had dissolved it. That was the error, and it survived into the quality board and a reply to the maintainer.
10b. Execution was never required, and this ADR already said so
§2: "Observed outcomes are needed to test whether a system IMPROVED; they are not needed to test whether it REMEMBERED." Asking for a procedure and grading the answer tests memory. Execution tests learning. Only the first is this family's subject, so the N4 (enacted) half stays out and N3 (declarative) is the whole build.
10c. The gap is measured, not argued
Across all 485 shipped questions — question text, gold answer and full haystack:
| property | occurrences |
|---|---|
precondition language (must … before, cannot … until, only after, prerequisite) |
0 |
causal-necessity language (because, so that, consumes, depends on) |
0 |
§3's table listed preconditions as unhomed and only the conditional was built. It is still
unhomed. episodic/list-order recalls an ordered list, but an arbitrary one — nothing in the
family tests an order that must hold, or a constraint that is neither a step nor a value.
10d. What is being built
Four shapes, 20 each, 80 questions. conditional-branch is NOT among them — it already shipped
inside Conjunction and duplicating it would be the weak-vertical failure ADR-027 §11.1 names.
| shape | what only it tests |
|---|---|
step-order |
an order that must hold; violating it is an error, not a wrong answer |
precondition |
a constraint that is neither a step nor a value, reached through a second hop |
amended-step |
one element of a sequence replaced, the rest intact |
retired-step |
a position removed from a sequence — not a fact invalidated |
The two known risks each have a technique demonstrated this week. §5's V2 blindness on ordering
questions is answered by invented causality (no prior can apply, and V2 tests exactly that). §5's
structural cap — "the question names the entity it asks about" — is answered the way
conditional-branch answered it: put a required hop in a session that names only the
subordinate entity, so the question cannot reach it. Measured there: rule retrievable 15/15, state
0/15.
10e. What is NOT decided, and is disclosed rather than assumed
ADR-026 §8's exclusion is bilateral. The maintainer holds this side of it and has exercised that call; the counterparty has not been asked and is not being asked to approve in advance. They are told what we are doing in the next disclosure, before it lands in a release they consume. If they object, this is one vertical and it can be withdrawn — which is the argument for telling them rather than for waiting.
Nothing in §9 changes. conjunction/conditional-branch stays where it is; the conditional is a
join and belongs in Conjunction, not in a procedure recital.
11. OUTCOME, 2026-09-03 — built and measured. Four deltas from §10d's plan.
Status: Accepted and implemented. Corpus e86a271e323f, 80 questions, four shapes at 20 each,
exactly as §10d planned. What follows is what §10d could not know.
11a. What it measured
SUPERSEDED BY §14. The
V6 77/80below was flattering and is corrected there, along with the corpus fix that replaced it. The V1/V2/V3/V8 figures stand; V9 is now 16/80 and headroom +0.80.
| shape | V1 | V8 | V9 | headroom | discriminates |
|---|---|---|---|---|---|
step-order |
20/20 | 20/20 | 1/20 | +0.95 | yes |
amended-step |
20/20 | 20/20 | 3/20 | +0.85 | yes |
precondition |
20/20 | 20/20 | 4/20 | +0.80 | yes |
retired-step |
20/20 | 20/20 | 7/20 | +0.65 | yes |
Vertical: V1 80/80, V2 80/80, V3 80/80, V6 77/80, V8 80/80, V9 15/80, headroom +0.8125 — the largest in the family, and fully reachable since V8 equals V1. Coverage 0.594; per shape 0.525 / 0.575 / 0.600 / 0.675, all in band, calibrated per shape from the first build.
§10d's two techniques both held. Invented causality: V2 80/80, so no leak agreed with gold. The second-hop construction: V9 15/80 against V8 80/80, which is the gap the asymmetry was built to produce.
11b. Delta 1 — precondition asked for one hop and graded two
§10d describes the shape as "reached through a second hop", and the first question text was "What has to be true before X can start?". Gold named both the gate and its prerequisite, so a reader that correctly named the gate alone was graded wrong: V1 15/20.
The corpus was not the defect; the question was. It now asks "…and what does that depend on in turn?" — V1 20/20. This is the shape-level form of a rule this family has now hit repeatedly: gold may not require what the question did not ask for.
11c. Delta 2 — V6 counted a decline as a success
amended-step came back V6 11/20 failing, and the failures were readers explicitly declining
the amendment link when a component was dropped — which is the correct behaviour, scored as though
they had reached gold. The instrument for this already existed (answer_must_name, built for
semantic/co-reference the day before) and simply had not been declared here. Declared: 77/80.
The three residual failures (tme-prc-008, 017, 018) are understood and unfixed.
11d. Delta 3 — the separability gate refused the corpus four times
84 leaking phrases → 61 → 1 → 0. Every failure was filler drawn from a bank or frame the gold
owned: a hardcoded membership sentence, then a shared step bank, then 'of the sequence' leaking
from the edit frames. Each fix gave filler the same bank; none weakened the check. Disjoint
filler banks are now asserted at import.
This is the same defect class as the 72 distractors that shipped in v0.32.0-beta. It was caught before generation completed rather than after release, which is the gate working.
11e. Delta 4 — §10e was overtaken. The counterparty asked to co-design.
§10e said the counterparty "is not being asked to approve in advance" and would be told in the next disclosure. They were told, and their reply raised no objection to the vertical but asked for something §10e did not anticipate: co-design the corpus intent against their AIP/skills track before the corpus is fixed, so that two procedural corpora which cannot share questions are not built in parallel.
That is a reasonable ask and it is honoured. A one-page shape list is written. The corpus is committed to a branch and is deliberately not released, so nothing they consume depends on it while their single reply is outstanding. The bilateral exclusion in ADR-026 §8 is not narrowed by this ADR alone; it is narrowed when they answer.
One open risk, stated rather than assumed: our step names are semantics-free invented artefacts. If their track needs steps that map to callable operations, that is a bank change — cheap now, expensive once the corpus is cited.
12. STANDALONE, 2026-09-03 (later) — the co-design hold is lifted, on the maintainer's decision.
§11e is superseded on one point and stands on every other. Its record of what the counterparty asked for is accurate and stays. What is withdrawn is the consequence it drew: that the corpus would be held unreleased until they replied.
The decision, stated without dressing it up. The maintainer has decided this project goes standalone: build, release, and notify the counterparty once the work is done and usable, rather than solicit their input first. This is not a claim that ADR-026 §8's exclusion has been renegotiated — it has not, and §1 is still right that this project cannot narrow a bilateral agreement by itself. It is a decision to act and disclose rather than to wait, accepting that this departs from the agreement's spirit, and it is recorded here so the departure is legible rather than discovered.
This reinstates §10e, which §11e had set aside. §10e's policy was already exactly the standalone one — "the counterparty … is not being asked to approve in advance … told what we are doing in the next disclosure … If they object, this is one vertical and it can be withdrawn, which is the argument for telling them rather than for waiting." §11e overtook that on their request to co-design; §12 puts it back. The withdrawability in §10e is the part that makes this defensible and it is unchanged: one vertical, removable, with no other corpus depending on it.
It is also consistent with what the maintainer said when §7b was written — "we can decide this alone; the other party is simply not ready to use it, but if it works, it will."
12a. What changes in practice
| before | after | |
|---|---|---|
| corpus | committed, held unreleased pending their reply | released on the normal cadence |
| the shape list | a request for input, with three gating questions | a notification, sent once the vertical is usable |
| their reply | a precondition | welcome, and not waited on |
| withdrawability | — | unchanged: still one vertical, still removable |
12b. What does NOT change
- The exclusion in ADR-026 §8 is not narrowed by this document. It is departed from, knowingly.
- §11a–§11d stand entirely — the measurement, and the four deltas from plan.
- The enacted half stays out (§10b). Nothing here reopens it.
- The corpus remains withdrawable. If the counterparty objects on seeing it, removing the vertical costs one generator, one corpus, one enum member and one row in each published table.
13. THE REPLY ARRIVED, 2026-09-03 — no conflict, and three substrate changes taken.
Sequence matters here, so it is stated first. §12's standalone decision was made without the counterparty's answer and did not depend on it. Their reply then arrived on its own. It is recorded here as what it is — a confirmation that cost us nothing to have waited for and that we did not wait for — not as retroactive permission.
13a. What they said
No conflict, merge. The four shapes map onto their AIP plan rather than competing with it:
step-order and precondition are depends_on and typed preconditions, amended-step is
SUPERSEDES, retired-step is retirement. They confirmed invented over recognisable — not as a
concession but as required by their side too, since recognisable procedures are the Yarrow class
waiting to happen and invented content also avoids training-data contamination of skill content. And
they accepted the enacted-half caveat verbatim: it is a harness, not a vertical.
They also observed something we had not: our invented causality is their typed-port dataflow stated in prose. "Vreskade consumes what Quorlory produces" is a produces/consumes edge.
13b. The three substrate changes, all taken
Each is carried in the substrate and asked by no shape, so none of them changes what this vertical measures, what it grades, or its 20-per-shape parity.
| change | cost | |
|---|---|---|
| a | applicability — the teaching sessions now state when the procedure applies and when it must not run (trigger_when / do_not_use_when) |
a haystack change, batched into a regeneration already required |
| b | per-step operation ids with 0–2 named argument slots, in the extension |
extension-only, free |
| c | the produces/consumes edge exposed in the extension as dataflow |
extension-only, free |
(b) and (c) cost nothing because question_key is derived from the question, answer, date and
haystack and excludes the extension — so extension-only changes do not invalidate a single probe
call. (a) does change haystack bytes, and was free in practice only because §11's V6 defect had
already forced a regeneration.
Operation ids are derived, not tabulated (the Quorlory sync → quorlory_sync, slots from a
deterministic hash of the step name), so a step cannot be added without one. They are
semantics-free on purpose: tool shape without real-world semantics, which is what keeps the
invented-causality property that gives V2 80/80.
13c. The one design consequence worth naming
An applicability session names the procedure, which is the thing check (a) exists to ration. It is permitted, and the reasoning is that check (a) guards a specific failure — a retriever reaching the answer in one hop — not the procedure name as such. An applicability session carries no step, no order, no gate and no revision, so it cannot collapse the second hop. Check (a) is scoped to gold and was already unaffected; check (b) gained a narrow exemption for the procedure name only, and still forbids these sessions from naming any step.
What it does do is compete for the retrieval budget on the question's own words, which makes V9
harder rather than easier. A new check (b2) pins exactly one applicability session per question,
non-gold, with its fields present — because no shape asks an applicability question, so nothing else
in the file would notice if the substrate silently rotted.
14. V6 REPORTED FOUR NUMBERS BEFORE IT REPORTED A TRUE ONE, 2026-09-04.
§11a published V6 77/80 and §11c explained one defect behind it. Both were premature. The arm was
wrong in both directions before it was right, and each step moved the number for a different
reason — only the last moved the corpus.
| V6 | what was actually happening | |
|---|---|---|
| 1 | 60/80 | readers declining the link scored as reaching gold. Fixed by declaring answer_must_name (§11c) |
| 2 | 77/80 | flattering, and published. The dependency sessions state the adjacent pairs of a 4-chain, so their transitive closure is the gold order, and filler stated only isolated pairs. The gold chain was the only complete order in the haystack, so the membership session — this shape's entire second hop — was redundant. 14 of 60 membership-drop samples reproduced the gold order verbatim; only 3 were condemned, the rest rescued by the resolution grader saying declined |
| 3 | 68/80 | truer, after a guaranteed rival chain — and now condemning on luck. No chance_floor was declared, so v6_needs collapsed to 1: one hit in three samples condemns, against a guesser firing at 1−(2/3)³ = 0.70 at k=3 |
| 4 | 51/52 | honest where it could decide, silent where it could not. k=2 leaves no hit threshold at three samples, so all twenty step-order questions returned undecidable |
| 5 | 70/72 | a second rival chain puts k at 3, where the threshold is 3-of-3. V6 scores 72 of 80 questions instead of 52, and step-order is measured rather than blind |
14a. Three rules this cost enough to be worth stating
A chance floor absent is not a floor of zero. It collapses the required-hit threshold to 1 and condemns components that are load-bearing. That is the understating direction, which is exactly why it survived a probe and a review: a corpus that looks worse than it is raises no alarm. A number wrong in the safe direction is still a number nobody can cite.
An undecidable result is not a pass, and a smaller denominator is not a better score. 51/52
reads better than 68/80 and says nothing whatever about a quarter of the corpus. Any V6 figure
must be read with its scope: this vertical's is now 72 of 80.
A structural claim needs competition, not declaration. gold_components_load_bearing: true was
true of the amendment shapes and false of step-order, in the same corpus, under the same flag.
What separates them is that the amendment shapes had _rival_edits and step-order had nothing to
compete with its chain. The generator now refuses a corpus whose rival chains are missing, short, or
overlapping — check (d2).
14b. What it cost, stated because the trade should be visible
step-order realises 0.400 coverage against a 0.50 floor, down from 0.4625 with one rival chain
and 0.525 with none. It is the furthest out-of-band shape in the family, and it moved away from
the band on purpose — the one direction the ratchet in TypedMemEvalCoverageBandTests exists to
refuse, so the reason is recorded there rather than the number quietly raised.
That is a proxy traded for a measurement, which is the trade ADR-028 licenses: accept a shape on measured discrimination, calibrate on the proxy. Removing the rivals would restore the coverage number and delete the construct.
14c. Still open
Eight precondition questions remain undecidable — their haystacks happened to carry only two
distinct conditions, and the floor is counted per question rather than assumed. Guaranteeing a third
rival condition would close it, at the cost of a regeneration and a re-probe. Recorded rather than
fixed, because 72 of 80 is a scope worth citing and the fix is not urgent.