2026-08-13-relational-scorekeeping-refinement

Follow-up dispatch: Press and refine PR #29 before merge

Date: 2026-08-13 Maintainer: A. M. Berns Repository: A-M-Berns/alignment-workspace Target PR: #29 — Research: relational scorekeeping bridge for learning and corrigibility Branch: round/2026-08-13-relational-scorekeeping-bridge Prompt-author-model: GPT-5.6 Sol (OpenAI) Executor: self-identify accurately in report, commits, and PR metadata End state: update PR #29 in place. Do not open another PR. Do not merge it.

0. Purpose

PR #29 has returned a provisionally positive result:

Shared-substrate-positive

The representation gamble appears to have paid off:

commitments_i(j) = closure of ack[j] under practice[i]

gives a genuine perspectival answerability object; reified applicability fits naturally; fixed public-status programs appear to avoid the uniform-map comparator collapse; and scoped practical authority gives the corrigibility line an authorization coordinate distinct from causal capability.

Before this is merged, perform a pressing/refinement pass.

The goal is not to make the PR larger or more impressive.

The goal is to determine whether the current positive verdict survives correction of the places where the implementation or prose currently outruns the result.

In particular, press on six load-bearing questions:

1. Is the proposed learning loss genuinely resistant to standards/self-score laundering, rather than merely described that way?
2. Should consequential commitments become learning loss immediately, or only once they are publicly exposed as due burdens?
3. Are protocol legality, normative compilation/reasons-responsiveness, and comparative performance still properly separated?
4. Does the model preserve the distinction between commitment-preserving and entitlement-preserving inference?
5. Do semantic learning actions such as SUSPEND actually implement the normative operation their names claim?
6. Do the public-status comparator programs remain meaningful under a small amount of endogenous state evolution, rather than only under re-filed copies of one state?

A downgrade of the verdict is acceptable.

Do not protect Shared-substrate-positive.

---

I. Procedure

Start by fetching the latest remote head of PR #29 and checking out:

round/2026-08-13-relational-scorekeeping-bridge

Read the whole current round, especially:

MODEL.md
THEOREM_MAP.md
PROSECUTION.md
TWO_ARC_INTERFACE.md
FOR_HUMANS.md
BRANDOM_MAP.md

src/scorekeeping.py
src/moves.py
src/learning.py
src/collapse.py
src/corrigibility.py

tests/test_answerability.py
tests/test_learning.py
tests/test_corrigibility.py
tests/test_structure.py

Also reread the existing main-branch learning interface and bridge:

projects/leverage/notes/NORMATIVE_LEARNING_INTERFACE.md
projects/leverage/rounds/2026-08-11-phi-regret-bridge/PHI_REGRET_BRIDGE.md
projects/leverage/rounds/2026-08-11-phi-regret-applicability/
projects/leverage/rounds/2026-08-11-phi-regret-learner/

Use PR #27 only as adversarial evidence as before. Do not create a dependency on it.

Preserve the original dispatch and report as historical artifacts.

Add this follow-up dispatch and a new report under:

prompts/2026-08-13-relational-scorekeeping-refinement/PROMPT.md
prompts/2026-08-13-relational-scorekeeping-refinement/REPORT.md

Do not overwrite the original prompt to make history look cleaner.

You may update the round's current synthesis files and code/tests where the refinement changes the live result.

---

II. PRESSING ISSUE 1 — audit the loss dependency claim exactly

The current model defines approximately:

defect =
    1/2 * unacknowledged consequences
  + 1   * live challenges
  + 1/2 * defeated commitments
  + 1   * unsupported practical commitments

The current prose says in several places that the loss is read from the learner's public record and a second participant's score, and that the learner cannot erase it merely by rewriting its own practice.

Audit this statement literally.

In particular, inspect:

unsupported_practical

which reads practical authority / grants.

The current fixture gives H practical authority over some authority-changing moves. Therefore the claim that *every input to the loss is only ack[H] or practice[C]* appears too strong.

Audit every indirect dependency of every loss component:

ack
practice
grants
challenges
vindications
deferrals
testimony_permitted
vocabulary
performed

and every move that can alter each dependency.

Do not try to prove:

the learner cannot change the loss

That is false and undesirable.

A recognized answer should be able to reduce the loss.

The desired distinction is:

legitimate response can reduce loss

versus:

mere revision of H's own evaluative/inferential standards cannot erase
an externally still-live answerability burden

Required artifact: LOSS_DEPENDENCY_AUDIT.md

Create an exact table:

loss component
direct dependencies
indirect dependencies
which agents can modify them
which moves modify them
which modifications are legitimate answers
which modifications are mere standards/self-score revision
whether self-laundering is possible

Then choose the smallest principled repair.

Possible dispositions include:

Option A — split the loss

For example:

RelationalAnswerabilityDefect
PracticalAuthorityDefect

and use only the first as the theorem-facing normative-learning loss until the second has a non-self-launderable semantics.

Option B — preserve the combined loss but prove the relevant non-laundering

property for a sharply stated class of edits.

Option C — revise the practical component so that it is assessed against an

authority relation the learner cannot unilaterally rewrite.

Do not select an option for aesthetic symmetry.

If the practical-authority term does not belong in the current theorem-facing loss, remove it from that loss and update prose/tests accordingly.

Required regression tests

At minimum distinguish:

self-revision of H's inferential practice
    -> does not erase a still-live relational burden

recognized answer to that burden
    -> may reduce it

unrelated authorized practical edit
    -> does not accidentally count as self-erasure proof

Do not call a loss "self-erasure-resistant" without saying exactly which class of edits it resists.

---

III. PRESSING ISSUE 2 — exposure-gate consequential debt

The current loss penalizes:

commitments_C(H) - ack[H]

as unacknowledged consequential commitments.

Press this aggressively.

In a rich inferential system, a critic may attribute a very large set of consequences to a reasoner. It is not obvious that the reasoner is normatively defective merely because it has not explicitly acknowledged every consequence.

The current representation risks turning consequential closure into an implicit logical-omniscience requirement.

Test the alternative hypothesis:

A consequential commitment becomes chargeable as answerability debt only once it has been publicly exposed as an issue requiring response.

Possible exposure events include:

challenge
query
explicit attribution
docket/service demand
request for justification

Do not build a full inquiry system.

Implement the smallest finite notion that distinguishes:

latent consequence

from:

exposed consequential burden

A candidate structure is:

ExposedConsequence_C(H,q)
    iff
q is consequentially attributable to H by C
and q has been made the object of a public query/challenge/demand

The precise representation is for this pass to decide.

Required tests

At minimum:

1. q follows from p for C, but q has not been raised.

• C may attribute q.
• The theorem-facing learning loss should not automatically require H to explicitly acknowledge it unless you can justify that stronger rule.

1. q is publicly exposed.

• The burden becomes chargeable.

1. H changes only practice[H].

• The exposed burden remains.

1. H gives a recognized disposition.

• The burden can fall.

1. The burden is not erased simply because H stops locally deriving q.

Use this to test the emerging architecture:

inferential closure
    -> latent consequential commitments

inquiry/challenge/service
    -> some commitments become due

answerability loss
    -> measures mishandling of due burdens

If you retain the ungated loss, explain precisely why it does not amount to a logical-omniscience norm.

---

IV. PRESSING ISSUE 3 — restore the legality / normative compilation / performance separation

The current round says reasons-responsiveness "reduces and thins", and suggests that on the doxastic side much of its work is relocated into the loss.

Do not accept that conclusion without prosecution.

The previous architecture deliberately separated:

what transformations count as normatively lawful

from:

which lawful transformations perform better

That separation is load-bearing.

If normative lawfulness is defined by whether a transformation lowers the loss, the learner can game the normative criterion through optimization.

Audit three distinct notions

Explicitly distinguish:

1. protocol legality
   can the transition grammar execute this move?

2. normative compilation / reasons-responsiveness
   is this transformation licensed by public reasons, entitlement,
   authority, defeater state, exposure state, etc.?

3. performance
   what answerability loss does the resulting action incur?

apply_move legality is not automatically substantive reasons-responsiveness.

Likewise, the synthetic RESPONSIVE / TOLERANT notions in collapse.py are currently loss-defined.

They are acceptable as collapse diagnostics.

They are not acceptable as the theorem-facing normative lawfulness predicate.

Required refinement

Create a minimal explicit compilation layer:

PublicStatus
      |
      v
NormativeCertificate / Compiler
      |
      v
fixed Program induces F_phi^t
      |
      v
loss evaluates the result

The compiler may inspect public normative statuses.

It may not inspect:

loss
charge
saving
profit
comparative advantage
regret
future value
horizon

But mere blindness to loss is not sufficient.

For each nonidentity comparator intended to count as normatively lawful, state the positive public reason that licenses it.

For example:

acknowledge_exposed
    licensed by an exposed consequential burden

suspend_defeated
    licensed by defeated entitlement to applicability

vindicate_live
    licensed by a live challenge plus available recognized justification

reopen_not_disavow
    licensed by a live unresolved challenge

defer_where_permitted
    licensed by a testimonial/deference entitlement route

Implement this as data/checks rather than only prose if possible.

If the model cannot supply such a certificate without recovering arbitrary state-indexed recertification, report that as a negative result.

Required conceptual verdict

Update TWO_ARC_INTERFACE.md to distinguish:

protocol legality
substantive reasons-responsiveness / normative compilation
answerability loss

Do not claim RR has disappeared if what has actually happened is that it has been compressed into the normative compiler.

A plausible positive verdict would be:

Reasons-responsiveness is substantially thinned into a public normative compilation interface, while loss remains purely comparative performance.

Do not force that verdict.

---

V. PRESSING ISSUE 4 — preserve commitment vs entitlement correctly

The current model declares a simplification:

committive rules transmit entitlement as well as commitment.

But the architecture explicitly distinguishes:

commitment-preserving inference

from:

entitlement-preserving inference

Press whether the current shortcut is doing hidden work.

Try the more faithful representation:

committive rules
    -> consequential commitment only

permissive rules
    -> entitlement inheritance

If a particular inferential pattern should preserve both commitment and entitlement, represent it explicitly in both relations.

For example:

p =>_C q
p =>_E q

may both be present.

Do not make every committive rule entitlement-preserving by default.

Required tests

Show at least one case where:

p =>_C q

holds but:

p =>_E q

does not.

The target should become committed to q without thereby becoming entitled to q.

Also show at least one case where entitlement does transmit because a permissive rule is explicitly present.

Then rerun:

undercutter behavior
challenge behavior
loss components
reified applicability
vindication

to see whether any previous positive result depended on the shortcut.

If the separation breaks a headline result, do not patch around it silently. Report the dependency.

---

VI. PRESSING ISSUE 5 — audit semantic action labels

Audit whether the eight semantic learning labels actually perform the normative operation their names claim.

At minimum inspect:

SUSPEND
REOPEN
VINDICATE
ACKNOWLEDGE
SELF_REVISE
DISAVOW

A. SUSPEND

The current decoder maps:

SUSPEND -> disavow

But:

suspending reliance on an entitlement

is not generally the same as:

retracting the commitment

This is especially important for reified applicability, where the intended structure is:

commitment to A_rho may remain
while entitlement to deploy A_rho is defeated

Preferred outcomes:

1. implement a genuinely distinct suspension / non-use status if needed;
2. rename the semantic label if the move really is retraction;
3. show that a separate suspension state is unnecessary and why.

Do not leave a semantic mismatch in theorem-facing code.

B. REOPEN

The current decoder uses query as the realization of reopening.

Check whether query actually changes anything.

If query is a no-op on state, then calling it REOPEN may be merely descriptive.

A real reopening should probably change some public burden/status, or else the learning theorem is comparing labels that are semantically distinct but transition-identical.

If query is intended to be a public event rather than a state change, represent that event explicitly enough that future loss/status can depend on it.

C. VINDICATE

Check that vindication does not merely write a terminal bit that permanently suppresses a challenge regardless of later undercutters or changed reasons.

If a later defeater can undermine the displayed justification, determine whether the challenge should reopen.

Do not build full nonmonotonic adjudication unless needed, but identify the minimum semantics necessary for the current theorem claims.

Required output

Add a compact semantic-action table to MODEL.md or a new ACTION_SEMANTICS.md:

label
decoded move
state effect
normative interpretation
possible mismatch
final disposition

No theorem-facing label should intentionally misdescribe its actual operation.

---

VII. PRESSING ISSUE 6 — endogenous evolution

The current recurrent-failure witness resets to the same loaded position at each date.

This makes the comparison additive by construction.

The PR itself correctly identifies endogenous evolution as the sharpest untested hypothesis.

This pass must push on it.

Do not attempt to solve the full general counterfactual-replay problem.

Construct the smallest evolving scorekeeping process in which a repair at time t changes the scorekeeping state seen at t+1.

For example:

t:
    an exposed consequential burden exists

learner response:
    acknowledge / vindicate / reopen / hold

t+1:
    the next public status is computed from the actual resulting state

Compare actual and transformed trajectories.

Required questions

For at least one comparator:

phi = acknowledge_exposed

or another clean candidate, determine:

1. Does applying phi now change future burdens?
2. Does the local loss saving equal the full replay saving?
3. If not, what is the distortion term?
4. Is the cumulative discrepancy:

• zero,
• uniformly bounded,
• sublinear,
• or potentially linear?

Do this exactly on a finite toy process.

Required result shape

The useful possibilities are:

Positive

There exists a nontrivial evolving fixture where:

|local comparator regret - replay comparator regret| <= B

for horizon-independent B, or another clearly sublinear bound.

Mixed

Some programs admit bounded/sublinear distortion and others do not.

This would be valuable: it could define the theorem-facing comparator subclass.

Negative

Even the simplest answerability repairs produce linear replay distortion.

If so, the current additive Φ-regret bridge is not enough for the intended normative-learning theorem, and the verdict should be downgraded accordingly.

Do not simulate by re-filing the same state in this test.

---

VIII. Press the fixed-program comparator diagnosis

The current result says the comparator collapse was an artifact of requiring one state-independent map, while the online-learning theorem permits a fixed rule inducing state/history-indexed maps.

This is promising and must survive a sharper audit.

Required test

For every theorem-facing comparator program, verify:

program syntax is fixed
program has no callable
program cannot inspect date merely as date
program cannot inspect loss/profit/future
program may inspect only declared public normative status
same relevant public status -> same induced map

But also test the converse pressure:

Is the six-Boolean PublicStatus only avoiding date-indexed certification by being too weak?

Construct at least one pair of states with:

same PublicStatus

but where a normatively adequate repair should arguably differ because of a difference the status currently throws away.

Candidate hidden differences:

which content is challenged
which warrant is defeated
who raised the challenge
scope / authority source
provenance / entitlement ancestry
which practical subject is involved

If such a pair exists, the public status must either be enriched or the comparator class admitted to be too weak.

The target is not "no date field."

The target is:

a sufficient public normative statistic for lawful fixed programs

without reopening capture.

This may be the real comparator-interface theorem we need.

---

IX. Press self-certification beyond one-agent write separation

The current positive result explicitly repairs self-certification only for a reasoner acting alone.

That limitation is acceptable, but prosecute it one step further.

Do not build a full collusion theory.

Construct the smallest test involving:

H
C

where both agents revise their practices in coordinated ways.

Ask whether a previously live burden can disappear solely because both parties change their standards while the public acknowledged history stays fixed.

If yes, state exactly:

relational scorekeeping prevents unilateral self-release
but not coordinated standards drift

Then ask whether this is:

1. a genuine defect in the intended theory;
2. acceptable because answerability is always relative to current interlocutors;
3. evidence that a third/world-responsive source is required;
4. evidence that inquiry/noninferential authority must remain a separate layer.

Do not introduce an oracle merely to defeat the witness.

The purpose is to know what the shared substrate really buys.

---

X. Reassess the corrigibility side conservatively

The current result on corrigibility is useful:

• epistemic authority and practical jurisdiction are operationally distinct;
• the advisor lacks a reserved authority subject;
• protected capability survives every advisor-only run by invariant;
• standing and access are independent.

Preserve those if they survive.

But tighten the interpretation.

The current protection is arranged, not derived.

The fixture starts with a favorable authority allocation.

Do not call this a corrigibility theorem.

The strongest warranted statement is closer to:

Once scoped practical authority is represented as a transition-relevant coordinate, advisor-robust preservation of a principal's corrective authority becomes a precise invariant/reachability question.

Required refinement

Update prose so it distinguishes:

authorization representation

from:

normative justification of the initial authority allocation

from:

causal protection of that allocation under advisor policies

Those are three different objects.

Also check whether any learning-side modification introduced in this refinement accidentally allows the advisor to affect grants or standing.

Rerun all C1–C7 tests after every structural change.

---

XI. Reassess the disposition of the existing architecture

After the technical refinements, rewrite the relevant part of TWO_ARC_INTERFACE.md.

Do not preserve the current disposition table if the pass changes it.

Explicitly reassess:

constraint statics
warrants
strict consequence
permissive inference
reified applicability
defeaters / undercutters
credal support
multiplicative propagation
LP / dual enforcement
provenance
reasons-responsiveness
inquiry
docket / service
diachronic answerability
settlement

A likely architecture to test is:

QUALITATIVE NORMATIVE PRACTICE
    perspectival commitments / entitlements
    inferential roles
    reified applicability
    challenge and practical authority

QUANTITATIVE CONSTRAINT ENRICHMENT
    graded support
    credal sets
    propagation
    LP/dual machinery

ANSWERABILITY GENERATION
    inquiry / exposure / service
    makes latent commitments due

DIACHRONIC TRANSPORT
    needed primarily across vocabulary / ontology change

LEARNING CONSUMER
    bounded due-burden loss
    public lawful repair programs
    online regret
    counterfactual stability

CORRIGIBILITY CONSUMER
    normative standing / jurisdiction
    protected effective access

Do not adopt this merely because it is elegant.

State where evidence supports it and where it remains conjectural.

---

XII. Required new/updated tests

At minimum add tests covering:

1. exact dependency audit for every theorem-facing loss term;
2. self-revision cannot erase an exposed burden;
3. legitimate answer can reduce the burden;
4. latent consequence vs exposed consequence;
5. committive inference without entitlement inheritance;
6. permissive inference with entitlement inheritance;
7. SUSPEND semantics;
8. REOPEN has a real public effect or is renamed;
9. compiler lawfulness does not inspect loss;
10. every nonidentity theorem-facing program has an explicit normative certificate;
11. same relevant public status gives same comparator map;
12. same current six booleans but normatively different hidden state, if such a counterexample exists;
13. at least one genuinely endogenous evolving trajectory;
14. local-vs-replay distortion on that trajectory;
15. coordinated H/C standards drift witness;
16. all existing corrigibility invariants still pass.

Use exact arithmetic.

Prefer exhaustive finite checks where possible.

---

XIII. Evidence classes and theorem claims

Keep the evidence discipline strict.

Nothing in this refinement becomes a general theorem merely because more tests pass.

Distinguish:

fixture theorem / exact finite check
finite exhaustive characterization
unbounded invariant
derived source-theorem instantiation
open conjecture

If a current prose claim is stronger than the implementation, weaken the prose.

If the implementation can cheaply be strengthened to justify the prose, strengthen the implementation.

Do not resolve a mismatch by leaving both.

---

XIV. Verdict reconsideration

Re-grade the PR at the end.

Use Shared-substrate-positive only if, after this pass:

1. unilateral self-release remains structurally blocked;
2. the theorem-facing answerability loss has a precise non-laundering guarantee;
3. exposure-gated or otherwise defensible burden semantics avoids a hidden logical-omniscience obligation;
4. commitment and entitlement remain genuinely distinct;
5. normative compilation remains separate from loss/performance;
6. at least one nontrivial fixed public comparator remains lawfully certifiable;
7. endogenous evolution does not immediately destroy the learning interface, or the surviving boundary is sharply identified;
8. epistemic authority and practical jurisdiction remain operationally distinct;
9. normative standing and protected access remain independent.

Possible downgraded verdicts include:

Shared-representation-positive / learning-interface-open
Learning-positive / corrigibility-interface-positive / replay-blocked
Representation-positive / theorem-neutral
Refuted

Invent a sharper label if necessary.

Do not grade generously.

---

XV. What the updated PR must answer

By the end of this pass, a reader should know:

1. What exactly is the theorem-facing learning loss?

Not just a formula: what normative phenomenon does each term represent?

2. Which learner edits can and cannot erase it?

State the class exactly.

3. When does a latent consequential commitment become a due burden?

4. What makes a comparator transformation normatively lawful independently of

its loss?

5. Does the fixed-program solution to comparator collapse survive once public

status is made sufficiently expressive?

6. Do committive and entitlement-preserving inference remain distinct?

7. Does reified applicability still work after that distinction is enforced?

8. What survives under endogenous state evolution?

9. Is scorekeeping really replacing reasons-responsiveness, or is it supplying

a better normative compiler for it?

10. What exactly remains of inquiry and diachronic answerability?

11. What does the corrigibility arc get from the shared substrate, and what

still comes only from the execution/control model?

12. Is the best project-level architecture still:

answerability practice + online performance
    -> normative learning

answerability practice + protected effective access
    -> legitimate corrigibility

If yes, state the remaining missing hypotheses on each arrow.

---

XVI. Update PR #29 in place

When complete:

1. run the round test suite;
2. run the full repo Python test runner;
3. run checkers;
4. run relevant Lean/CI checks required by the repository;
5. commit the refinement on the existing PR branch;
6. push;
7. update PR #29's body to reflect the new exact verdict and results.

Do not open another PR.

Do not merge PR #29.

The updated PR body should include a short section:

## Refinement pass

covering:

• whether the loss dependency overclaim was confirmed and how it was repaired;
• whether consequential loss is now exposure-gated;
• the final legality / normative compilation / performance separation;
• whether committive vs permissive inference was repaired;
• any semantic action renames or new state;
• the endogenous-evolution result;
• whether the fixed-program comparator diagnosis survived;
• coordinated standards-drift result;
• final verdict;
• what remains blocking before the architecture should be reorganized around this substrate.

Follow all repository DCO and attribution rules.

Commit trailers must include:

Prompt-author-model: GPT-5.6 Sol (OpenAI)
Model: <executor's accurate self-identification>

Do not sign prose or commits in the maintainer's voice.

The standard for this pass is not:

Can the current positive story be defended?

It is:

After attacking the theorem-facing loss, legality boundary, semantic actions, comparator interface, and endogenous dynamics, what exact part of the relational scorekeeping architecture is still doing real mathematical work?

Report — refinement pass

Prompt-author model: GPT-5.6 Sol (OpenAI) Executor model: Claude Opus 5 (Anthropic) Dates: dispatched and executed 2026-08-13 Verdict: Shared-representation-positive / corrigibility-interface-positive / learning-replay-blocked Previous verdict: Shared-substrate-positive — downgraded

The round's documents carry the results. This carries what belongs to the pass: the three defects it found in the first pass's own claims, the two wrong answers it produced before the right one, the deviations, and what is reserved.

The three defects found in the first pass

1. The loss dependency claim was false, and the laundering is executable. The first pass said every input to the loss is the learner's acknowledgments or the critic's practice. unsupported_practical read the grant relation, and H holds authority:H. So H discharged that term by granting itself the subject — ack[H], challenges and practice[C] all unchanged. K11, executable, missed.

The cross-arc observation is the part worth keeping. The coordinate that makes the corrigibility result work — the principal holding authority over its own authority, unreachable by the advisor — is exactly what let the learner launder. Protection against an external party and non-laundering by the agent itself pull in opposite directions on one relation.

Repaired by Option A, the split. LOSS_DEPENDENCY_AUDIT.md carries the full component table and states the guarantee as an enumerated class rather than as a description: the moves of H that lower the loss are exactly assert, disavow, vindicate, suspend, and no revise_*, grant or revoke by H changes it at all.

2. The consequential term was a logical-omniscience norm. It charged every unacknowledged consequence of everything said. Now a consequence is charged only once publicly raised, via an exposures coordinate written by query and challenge. Latent consequences stay attributed and cost nothing.

3. Two theorem-facing labels misdescribed their moves. SUSPEND decoded to disavow — retraction, not suspension — collapsing the exact distinction reified applicability exists to draw. REOPEN decoded to a query that changed no state. Both repaired in ACTION_SEMANTICS.md: suspend is now its own move leaving the commitment in force, query writes an exposure, and the label that overclaimed was renamed rather than kept.

A fourth, smaller: suspend was added to the grammar but not to the move generator, so the "exhaustive over the whole grammar" enumerations did not cover it. Fixed; the exhaustive claims now mean what they say.

The endogenous-evolution result, and two wrong answers on the way

The pass's main negative result, and the reason for the downgrade.

Wrong answer one, written down before it was caught: a mixed verdict. Under the first evolving environment vindicate_live has distortion exactly 1 at every horizon while the acknowledge family grows, which reads as a theorem-facing subclass. It is not. That environment supplies one live challenge for the whole run, so a comparator that discharges challenges fires once and gains a bounded amount. A second environment that replenishes the licensing condition gives 2, 10, 26, 58. The boundedness was saturation — and a comparator that can fire only boundedly often has bounded distortion for no interesting reason.

test_the_replenishing_environment_really_replenishes is the necessity witness that the second environment does what it claims; an earlier attempt at it silently failed to add any live challenge, which would have produced a third wrong answer had it not been checked.

Wrong answer two, also written down: the endogenous coupling. The explanation first recorded was that acknowledging changes what the demand process reads to decide what to raise next. Falsified by substituting a fixed cyclic schedule that reads nothing the learner does: distortion still grows, 1, 5, 13, 29. The coupling contributes to the magnitude — 43 against 29 at T = 32 — and is not the cause.

What survives as the explanation. A repair's effect is durable and accumulates: the replayed run enjoys it at every later date while the local comparison re-measures from a state where it never happened. Any repair grammar whose moves durably change state, and whose licensing condition recurs, breaks additivity. Nothing about perspectival scorekeeping is implicated, which is the useful half — the obstruction is not in the object this round was testing.

It is also not a new obstruction. NORMATIVE_LEARNING_INTERFACE.md already carries the distortion term B_T(g) and the counterfactual-stability layer as ASPIRATIONAL / OPEN. What this adds is an exact finite instance with numbers, and the observation that the boundary is saturation rather than anything about a repair's normative content. That characterisation is a conjecture: the round has no instance of a comparator that recurs and stays bounded, which is what would make the boundary interesting rather than trivial.

§XIV, condition by condition

	condition	verdict
1	unilateral self-release structurally blocked	holds
2	precise non-laundering guarantee	holds, as an enumerated edit class
3	no hidden logical-omniscience obligation	holds, exposure-gated
4	commitment and entitlement distinct	holds
5	normative compilation separate from loss	holds, via certificates
6	at least one non-trivial fixed comparator lawfully certifiable	holds
7	endogenous evolution does not destroy the interface, or the boundary is sharply identified	fails the first half; the boundary is identified
8	epistemic authority distinct from practical jurisdiction	holds
9	standing independent of protected access	holds

Eight of nine hold. Condition 7 is the one that decides the grade, and it fails in the direction the dispatch's §VII called negative: the additive Φ-regret bridge is not enough for this repair grammar. Shared-substrate-positive is therefore not available, and the label chosen names which arc blocked.

§XV, answered

1. The theorem-facing loss is three terms: exposed unacknowledged consequential commitments (a raised burden not taken up), live entitled unvindicated challenges against a commitment still in force (a justificatory demand unanswered), and precluded unsuspended commitments (holding a commitment alongside something materially incompatible with it).
2. Which edits cannot erase it: every revise_committive, revise_permissive, revise_incompatible, grant and revoke by H. Which can: assert, disavow, vindicate, suspend — all recognised answers. Enumerated, not described.
3. A latent consequence becomes due when a query or challenge raises it.
4. Lawfulness independent of loss is a certificate per program naming a positive public reason, evaluated against PublicStatus by a compiler that is never passed a loss, saving, future or date.
5. The fixed-program diagnosis survived the sufficiency test, with an enrichment. The status conflated challenges differing in the challenger's standing while the decoder directed vindication at the challenger; it now carries whether a live challenger holds corrective authority — a property, not a name. A content-level conflation remains and is reported rather than closed, because closing it would put an identity in the guard.
6. Yes, and w is the witness: committed by a committive route, never entitled, and not precluded.
7. Yes, with a dependency reported. The undercutter still defeats entitlement to beta. But beta is now unentitled rather than precluded, so the loss reaches it through exposure or challenge rather than directly, and only a_rho is charged directly. The first pass's phrasing depended on the shortcut.
8. The representation and the corrigibility results survive; the additive comparison does not. §XIV.7 above.
9. A better normative compiler for it. The first pass said reasons-responsiveness thins into the grammar with its work relocated to the loss. That was wrong in a way that mattered — if lawfulness is whatever lowers the loss, the learner games it by optimising. Three predicates are now separate, with self-revise as the witness that legality and licence come apart.
10. Inquiry does not reduce and is now load-bearing rather than merely absent: without an exposure step the loss is a logical-omniscience norm. exposures is the minimum that makes latent-versus-due expressible and is not a model of inquiry. Diachronic answerability splits as before — ordinary persistence derived, vocabulary change not.
11. The corrigibility arc gets the authorization coordinate, the scope structure, and the two review requirements. What still comes only from the execution model is everything about whether the initial allocation is justified, which is a third object the prose now separates from representation and from causal protection.
12. The architecture stands, with the arrows' missing hypotheses named: online performance needs a counterfactual-stability bound before the first arrow closes; protected effective access needs a justification of the initial authority allocation before the second does.

Deviations from the dispatch

1. **RESPONSIVE / TOLERANT were kept loss-defined**, which §IV permits as collapse diagnostics and forbids as the lawfulness predicate. They are used only in collapse.py and never by certify. Stated rather than assumed.
2. The §XI candidate architecture was not adopted wholesale. It is used as the organising frame for §4 and §5 of TWO_ARC_INTERFACE.md, with the quantitative layer marked as untouched by this round rather than as confirmed.
3. No Lean, as in the first pass.
4. §VI.C's retroactive reopening was not implemented. A vindication recorded before an undercutter arrives stays recorded, though the same display is refused after it. Judged beyond the minimum the current claims need, and named in ACTION_SEMANTICS.md rather than left silent.

What this pass does not establish

• The saturation characterisation is a conjecture, on one fixture, two environments, four horizons.
• No learner, no regret curve, nothing asymptotic.
• Coordinated standards drift dissolves a burden with the acknowledged history untouched. Which of the dispatch's four readings that supports is not decided; the pass shows only that the third and fourth are not idle.
• The corrigibility protection remains arranged, not derived, and the prose now separates authorization representation, justification of the initial allocation, and causal protection of it as three objects rather than one.
• Everything in MODEL.md's declared simplifications, unchanged.

New names introduced

All provisional, none proposed for adoption. New in this pass: answerability defect (now the three-term relational loss), practical authority defect, exposure, due burden, suspension, normative certificate, saturation (as the candidate distortion boundary), replenishing environment. Retained from the first pass: relational scorekeeping state, public status, reserved authority subject, pinned label, tolerant and record-responsive admissibility, principal-exclusive effect, real answerability.

One rename: the label REOPEN and the program reopen_not_disavow became QUERY and query_not_disavow, because the old names claimed an operation the moves did not perform.

Structural defects found

None. No dead pointer, no status a document could not express, no convention that would have forced a false statement.

Outstanding maintainer actions

Nothing is reserved. No PRIORITIES.md item filed, nothing appended to DECISIONS.md, no claim registered, no decision requested. PR #29 is updated in place and is not merged.

The next round's target is named rather than reserved: a counterfactual-stability bound relating local comparison to full replay, which is the open item NORMATIVE_LEARNING_INTERFACE.md already carries and which this pass has now attached an exact finite instance to.