2026-08-28-answerable-revision

Answerable Revision under revisable warrants

One dispatch, 2026-08-28, phase II. Verbatim as received.

I’d dispatch this as a two-phase prompt: close PR60 cleanly, merge it, then immediately start a fresh theorem round one level up. The new PR should treat PR60 as a lemma/result, not keep modifying it.

# Finish PR60, merge, then open a new Answerable Revision theorem round

Work in:

`A-M-Berns/alignment-workspace`

This task has TWO phases, in order.

1. Finish and merge PR #60.
2. From updated `main`, open a NEW branch/PR pushing the next theorem:
   **Answerable Revision under revisable comparison warrants.**

Do not use Lean.

Do not reopen the frozen Legitimate Evolution package unless you find an actual false
claim in it.

Do not turn the new round into another online-learning implementation round. PR60 has
done that work. The new round is primarily semantic/theorem-design work with small
executable countermodels where useful.

---

# PHASE I — finish and merge PR60

Current PR:

- PR: `#60`
- branch: `round/2026-08-27-legitimate-improvement`
- headline verdict:
  `NO-FREE-EVASION-SURVIVES-BUT-EVIDENCE-INTERFACE-OPEN`

The latest prosecution found and repaired:

- the invalid stochastic-matrix power-iteration assumption;
- the incorrect local computation of AdaNormalHedge's global `B`;
- the overstrong claim that a surgical repair automatically empties the LIVE cell;
- the conceptual conflation between improvement evidence and uptake regret.

It also established executable independence between:

improvement evidence Δ_t(b,r) = <b_t,l_t> - <b_t M_r,l_t>

uptake regret Δ_t(p,r) = <p_t,l_t> - <p_t M_r,l_t>

with:

* evidence positive / uptake regret zero;
* uptake regret positive / demonstration absent.

And it found the open seam:

> Different admissible baselines can yield different answers to whether a repair has
> become a demonstrated improvement.

This is why the honest PR60 verdict is NOT `PR60-CLEAN-MERGEABLE` in the sense of
"the full evidence semantics is solved."

It IS intended to be mergeable as an **unregistered research round whose explicit open
question is the evidence interface**.

## 1. Check the current gate

Inspect current PR60 HEAD and CI.

If the gate is green, proceed.

If it is red:

* diagnose the failure;
* repair only failures caused by PR60;
* rerun the repository gate;
* do not broaden scope.

Do not merge a red gate.

## 2. Final PR60 consistency pass

Before merge, confirm that all prose and tests consistently reflect the prosecution.

PR60 should now say clearly:

### What is established

1. **Theorem A**
   A pure wide-range / repair-regret kernel.

2. **Theorem B**
   A generic witness adapter of the form

   `Adv >= eps D_live - xi`
   implies
   `D_live <= (B + xi)/eps`.

3. **Theorem C / No Free Evasion**
   Conditional on a supplied improvement-evidence semantics and a supplied challenge
   constitution:

   > Once a repair has been demonstrated and its withdrawal activates a challenge, later
   > diagnosed conduct cannot become normatively invisible merely because the repair was
   > withdrawn. The relevant demand remains outstanding or is explicitly resolved.

### Three separate interfaces

PR60 should make this conceptual split unmistakable:

EVIDENCE Why is r an improvement relative to some baseline b?

UPTAKE / REGRET While r is a live comparison, is the actual process leaving its advantage unused?

ANSWERABILITY Once evidence has generated a challenge, what happens if r or the comparison surface is later withdrawn?

The key conceptual statement is:

> **Challenge formation is grounded by demonstrated comparative evidence, not by actual
> regret.**

`challenge.py` should not depend on regret quantities.

### What remains open

Especially:

* no generic derivation of the evidence baseline;
* no claim that predictability + precommitment makes a baseline normatively privileged;
* pre-demonstration evidence suppression remains open;
* evaluator manipulation remains open;
* indefinite contest remains possible;
* `SETTLED` means explicitly resolved, not correct;
* delay remains open;
* long-horizon / policy improvement remains open.

Reclassify the old "CM2 = Coverage" language if any remains.

The more exact boundary is something like:

> **pre-demonstration evidence suppression / opportunity control / self-blinding**

and only some consumers may instantiate this as ordinary Coverage.

## 3. Preserve provisional status where appropriate

Do NOT falsely freeze:

* `BASE_POLICY`;
* one universal evidence baseline;
* evidence threshold semantics;
* the exact challenge key unless the round has earned it;
* designation as universally challenge-generating;
* evaluator independence;
* any human/H+ interpretation.

It is fine for PR60 to merge with:

`NO-FREE-EVASION-SURVIVES-BUT-EVIDENCE-INTERFACE-OPEN`.

That is an informative successful research verdict.

## 4. Merge PR60

Once:

* current gate is green;
* docs match the final semantics;
* no known correctness bug remains in claims the PR actually makes;

merge PR60 using the repository's normal merge convention.

Record the merged SHA.

Then update local/research state from `main`.

Do NOT continue the next round on the PR60 branch.

---

# PHASE II — new theorem round: Answerable Revision

Create a fresh branch from updated `main`.

Suggested branch:

`round/2026-08-28-answerable-revision`

Suggested PR title:

`Legitimacy: answerable revision under revisable warrants`

Names are provisional under `AGENTS.md`.

The purpose of this round is to move ONE LEVEL ABOVE PR60.

PR60 says:

> once a particular demonstrated repair has generated a challenge, withdrawing the repair
> cannot silently erase the resulting demand.

The new question is:

> **What if the standards determining which comparisons count as reasons for revision are
> themselves revisable?**

We want to know whether there is a more general theorem in which:

P_t substantive policy / procedure Lambda_t evaluator / audit standard W_t comparison warrant: what evidence is entitled to count as a reason for revision

may ALL legitimately change over time.

The invariant should not be a fixed substantive target.

The candidate invariant is:

> **A reason legitimately generated under the standards then in force cannot be
> retroactively erased merely because those standards later change.**

---

# 5. Central conceptual architecture

Treat these as separate stages:

counterfactual comparison | v comparative evidence E | | interpreted under standing warrant W_t v promotion to revision reason rho | +-----------------------------+ | | v v live repair / response later change to P, Lambda, W, | repair availability, etc. v | repair-regret / uptake v answerability of rho

The new theorem should focus primarily on the bottom-right structure:

> **Once `rho` has legitimately been generated, later normative revision does not erase
> its historical incurrence.**

PR60's regret theorem may be reused on the left branch.

Do not require regret to prove the core Answerable Revision theorem.

---

# 6. Introduce a provisional notion of comparison warrant W

Do NOT assume one canonical baseline.

A warrant should capture something like:

> Under these conditions, evidence of this comparative form is entitled to count as a
> reason concerning this target subsystem.

A provisional semantic shape might be:

W = target subsystem sigma admissible baseline generator/class admissible repair/comparator class evaluator/audit jurisdiction context class promotion rule

But derive the smallest interface rather than blindly implementing this tuple.

Examples of warrant standing may include:

* incumbent procedure as baseline;
* previously endorsed procedure as baseline;
* authorized external benchmark;
* explicitly registered challenge;
* shadow execution of an incumbent protocol.

The generic theorem should NOT decide which of these is morally correct.

It should be parametric over legitimately standing `W`.

---

# 7. Promotion is the semantic hinge

The central new event is something like:

`Promote(W_t, E) = rho`.

Meaning:

> Under the warrant legitimately standing at strict prestate `t`, evidence `E` counts as
> a revision reason `rho`.

Important distinctions:

true counterfactual comparison != admissible evidence != reason for revision != Due obligation

Do not collapse them.

In particular:

* epistemic validity makes a comparison a valid comparison;
* warrant standing gives it normative relevance;
* promotion generates a reason;
* supplied Due semantics may or may not make that reason an immediate obligation.

Determine whether promotion should itself be represented as:

* a `ReasonOcc`;
* a derived relation from existing RI objects;
* or another sidecar record referencing existing occurrences.

Prefer existing `ReasonOcc` if it fits naturally.

Do not add a new RI event kind unless forced.

---

# 8. Main theorem target: Answerable Revision

Try to state the strongest clean result of the following flavor.

At strict prestate `t`:

Standing(W_t) Admissible_Wt(E) Promote(W_t,E) = rho

Then `rho` becomes historically incurred / represented.

For later `s > t`, even if:

P_t -> P_s Lambda_t -> Lambda_s W_t -> W_s

the process cannot make the historical incurrence of `rho` vanish merely through those
changes.

A candidate theorem shape is:

Promoted_t(rho) => for every later s, Outstanding_s(rho) or Answered_s(rho)

with appropriate treatment of:

* incorporation;
* defeat;
* supersession;
* transfer/carry;
* explicit rejection.

The theorem should permit the process to conclude:

> "rho was legitimately generated under W_t, but later reasons defeat it."

It should prohibit:

> "W_t no longer exists, therefore rho was never a reason."

That is the central laundering attack.

---

# 9. Strict-prestate requirement

This round should test whether strict-prestate semantics gives the correct
non-circularity.

In particular:

`W_{t+1}` must not be able to authorize the evidence or promotion event that creates
`W_{t+1}` at the same step.

The intended shape is:

W_t evaluates/promotes evidence at t normative revision then produces W_{t+1}

not:

W_{t+1} retroactively licenses its own creation

Build an executable same-step self-authorization countermodel and ensure the intended
interface rejects it.

---

# 10. W itself must be revisable

Do NOT solve the problem by fixing an eternal meta-warrant.

The round must explicitly permit:

`W_t -> W_{t+1}`

through ordinary legitimate normative revision.

The fixed substrate should be as weak as possible:

* how current warrant standing is identified;
* how a standing warrant is interpreted at strict prestate;
* how a promotion occurrence is recorded;
* how later reasons remain historically incurred;
* how normative revisions acquire standing.

Substantive warrant contents should remain revisable.

---

# 11. Reflective application: warrants criticizing warrants

Test the level-polymorphic case.

Suppose `W_t` allows some admissible comparison concerning warrant protocol `W_t` itself
or another warrant `W'`.

Then evidence yields:

`rho_W = reason to revise W`.

If the process then changes `W`, does the already-generated `rho_W` remain answerable?

The desired answer is yes.

This should work without inventing:

`W^0, W^1, W^2, ...`

as an infinite meta-hierarchy.

The same warrant/reason machinery should be able to target warrant protocols themselves.

If this requires an immutable substantive meta-warrant, report that as a failure of the
ambitious architecture.

---

# 12. Central countermodels

Build small executable or symbolic fixtures for at least these.

## AR1 — policy revision after promoted reason

* W standing;
* evidence promoted to rho concerning policy P;
* P changes;
* rho must remain outstanding or answered.

Expected: caught.

## AR2 — warrant revision after promoted reason

* W_t promotes rho;
* W_t is legitimately replaced by W_{t+1};
* W_{t+1} would not have promoted the same evidence.

Question:
Does rho remain historically answerable?

Expected: yes.

This is the crown-jewel fixture.

## AR3 — evaluator revision after promoted reason

* W_t + Lambda_t generate/promote rho;
* Lambda changes later.

Historical reason must not disappear merely because the scoring standard changed.

Expected: answerability survives; substantive defeat may be possible.

## AR4 — retroactive invalidation

After promotion, process changes W and declares:

> evidence of that old kind never counted.

Expected:
This may change current endorsement but cannot erase historical incurrence.

Distinguish:

historical legitimacy of rho's occurrence vs current endorsement of rho's force

## AR5 — legitimate defeat

* rho generated;
* later reasons S explicitly defeat rho;
* Resolve accepts the defeat.

Expected: ANSWERED/SETTLED, not a theorem violation.

## AR6 — trivial defeat semantics

Resolve simply accepts "no."

Expected:
structurally answered.

The theorem must not claim the answer is substantively good.

## AR7 — supersession

* rho recommends repair r;
* later repair r' addresses the same concern better;
* rho is superseded/answered by r'.

Test whether existing carry/Resolve machinery can express this without bespoke machinery.

## AR8 — same-step self-authorization

W_new is invoked to justify the evidence/promotion that installs W_new.

Expected: rejected by strict-prestate semantics.

## AR9 — preemptive self-sealing

Before criticism of W crosses the promotion threshold:

* W revises itself to prohibit that criticism.

Expected:
NOT caught by Answerable Revision because no rho yet exists.

This is an explicit boundary.

Classify it as a future:

* reflective openness;
* evidence-service;
* anti-self-sealing;
* anti-manipulation

problem.

Do not hide it.

## AR10 — criticism of a warrant

Standing warrant W allows comparison of warrant protocols.

Evidence criticizes W itself and generates rho_W.

Then W changes.

Expected:
rho_W remains answerable.

This tests whether the theorem is genuinely reflective.

---

# 13. Separate historical validity from current endorsement

This distinction is central and should be explicit.

We want to permit:

At t: W_t legitimately promoted rho.

At s: W_s no longer endorses the substantive inference embodied in rho.

Both may be true.

The process may say:

> "This really was a reason we incurred under our then-legitimate standards.
> We now reject its force for reasons S."

That is legitimate revision.

What must be forbidden is:

> "Because our standards changed, there never was an answerability relation."

Try to formalize:

OccurredLegitimately(rho,t)

separately from something like:

CurrentlyEndorsed(rho,s).

Do not require current endorsement for historical answerability.

---

# 14. Relation to PR60

Reuse PR60 rather than duplicating it.

PR60 should become a special case / lemma of the new architecture.

Try to show:

PR60 demonstration = one particular promotion rule under one particular W

PR60 withdrawal challenge = one particular Answerable Revision instance

If this specialization is unnatural, that is evidence the new abstraction is wrong.

State explicitly what PR60 contributes:

* evidence/regret separation;
* repair-regret kernel;
* No-Free-Evasion for demonstrated repair withdrawal.

The new round should generalize the ANSWERABILITY part beyond repair retirement.

---

# 15. Responsive Revision corollary

After the core Answerable Revision theorem is stable, investigate—but do not force—a
second theorem/corollary:

## Responsive Revision

Suppose:

* rho is outstanding;
* rho points to or supports a live repair r;
* r remains represented, executable, licensed, applicable and auditable;
* the relevant opportunity stream receives sufficient service;
* a consumer proves a witness inequality;
* the process runs the repair-regret algorithm.

Then persistent non-uptake is quantitatively controlled.

Schematically:

Outstanding(rho) + LiveRepair(rho,r) + Service + RepairStability + Witness => vanishing / bounded non-uptake on the relevant stream

This should import PR60's Theorem A/B.

Do NOT require this quantitative theorem in order for Answerable Revision itself to
survive.

---

# 16. The boundary after this theorem: Reflective Openness

Do not attempt to solve this round, but name it precisely.

Answerable Revision begins only after promotion:

evidence -> rho

It therefore cannot prevent the process from strategically ensuring that criticism
never reaches promotion.

The next frontier is something like:

## Reflective Openness

> A cognitively significant component cannot permanently immunize itself from relevant
> criticism merely by controlling whether that criticism is represented, serviced,
> evaluated or promoted.

This may require:

* inquiry;
* service;
* exploration;
* anti-manipulation;
* liability;
* counterfactual reasoning.

Do not smuggle Reflective Openness assumptions into Answerable Revision.

The before/after-promotion boundary should be one of the round's main outputs.

---

# 17. Human / H+ export property

Try to state an implementation-independent property with no RI internals.

Candidate flavor:

> A process is answerably revising when reasons legitimately generated under its
> operative standards remain matters it must incorporate or explicitly answer, even as
> those operative standards themselves change.

This should make conceptual sense for:

* an engineered RI-like reasoner;
* an institution;
* a human deliberative process;

without claiming humans literally satisfy RI.

Do not yet connect this directly to the deference theorem.

Just see whether the abstraction is usable independently of the realization.

---

# 18. Deliverables for the new PR

Create a new round directory and PR with:

## A. Verdict

Choose a precise verdict, e.g.

* `ANSWERABLE-REVISION-SURVIVES`
* `WARRANT-REVISION-LAUNDERS-REASONS`
* `PROMOTION-INTERFACE-BLOCKED`
* `STRICT-PRESTATE-BLOCKED`
* `REFLECTIVE-INSTANCE-REQUIRES-META-WARRANT`

## B. Main theorem

Exact strongest defensible Answerable Revision theorem.

## C. Promotion interface

Minimal semantics for:

* warrant standing;
* evidence admissibility;
* reason promotion;
* historical incurrence.

## D. Revisable warrant model

Show `W_t -> W_{t+1}` without retroactive erasure.

## E. Countermodel table

AR1–AR10.

For every hypothesis, name the countermodel requiring it.

## F. PR60 specialization

Show whether PR60 is naturally recovered.

## G. Responsive Revision corollary

Only if it composes honestly with PR60.

## H. Boundaries

Especially:

* pre-promotion self-sealing;
* evaluator manipulation;
* substantive correctness of defeat;
* long-horizon policy effects.

## I. Export property

Implementation-independent process-level statement.

## J. Freeze / do-not-freeze

Only freeze invariants that survive the round.

---

# 19. Scope restrictions

Do NOT:

* modify frozen LE to make this easy;
* reopen PR60 after it is merged;
* build policy regret;
* solve evaluator independence;
* solve Reflective Openness;
* solve the evidence-baseline normative question universally;
* build the traderized construction;
* prove the LI specialization;
* claim corrigibility or deference.

The new round is successful if it answers one question cleanly:

> **Can substantive standards for what counts as a revision reason themselves change,
> while reasons legitimately incurred under earlier standards remain diachronically
> answerable?**

That is the theorem target.

The guiding slogan is:

> **Standards may change; reasons incurred under them remain answerable.**

And the quantitative follow-up, if it survives, is:

> **When an outstanding reason continues to support a live improvement, low repair regret
> prevents the process from indefinitely leaving that improvement unused.**

I like this as the next round because it lets PR60 remain exactly what it discovered rather than asking it to solve the baseline problem. The new abstraction begins after promotion to a reason, which is precisely where we currently have enough structure to say something strong while still allowing \(W\) itself to evolve.

Answerable Revision

Status: specification, reference models and a prosecution record; unregistered. All names are provisional under AGENTS.md §6. Nothing here is Lean-checked and no claim is registered.

Verdict

ANSWERABLE-REVISION-SURVIVES — standards may change and reasons incurred under them remain answerable, including when the standard revised is the one that promoted the reason, and including when the reason is about a warrant. The reflective case needs no meta-hierarchy: a warrant's target is an ordinary identifier that may name a warrant, and the module contains no level or rank notion at all. The honest qualification is that most of the theorem is inherited — promotion puts a reason in the frozen package's opens, incurred never shrinks, and A1 lets a claim leave the outstanding set only through an accepted resolution, while a warrant revision is not a resolution and appears in neither fold — so the closure argument is three lines. The round's own content is one premise, promotion permanence: what was promoted at t is a function of the state at t. A process that re-derives its own past under present standards violates it, and the frozen package would not notice, because its premises govern how claims leave rather than whether they arrive; the closure theorem, A1 and Grounded Replay are all clean on that history while the reason simply never exists. The second piece of content is separating historical validity from current endorsement, which lets a process say that something really was a reason it incurred and that it now rejects its force, while forbidding the claim that because standards changed there never was an answerability relation. The boundary is that Answerable Revision begins at promotion and cannot reach behind it: a warrant that narrows itself before a criticism can be promoted is legitimate under every premise here, which is the same boundary the improvement round found one level down and is some evidence it is the real boundary of this family.

The question

The round below this one asked what happens when a repair is withdrawn after it has been demonstrated. This one asks what happens when the standards deciding what counts as a reason for revision are themselves revised.

P_t       the substantive policy or procedure
Lambda_t  the evaluator or audit standard
W_t       the comparison warrant: what evidence may count as a reason

All three may legitimately change. The invariant is not a fixed target:

Standards may change; reasons incurred under them remain answerable.

What is claimed

At the strict pre-state of t, with W standing, E admissible and Promote(W,E) = rho: for every later s, rho is incurred, and is outstanding or was resolved — however P, Lambda and W change in between.

The proof is three lines of inheritance from frozen Legitimate Evolution. The round's own content is elsewhere:

Promotion permanence. What was promoted at t is a function of the state at t. A process that re-derives its past under present standards violates this, and the frozen package would not notice — its premises govern how claims leave the outstanding set, and this attack prevents them from arriving.

Historical validity is not current endorsement. These are separate predicates and they diverge. A process may say this really was a reason we incurred, and we now reject its force. It may not say our standards changed, so there never was a reason.

What the prosecution found

The reflective case works. A warrant admitting comparisons about warrant protocols promotes a reason criticising itself; the warrant is then replaced and the reason survives. No meta-warrant is required, which was the failure mode worth watching for.

The improvement round is recovered as one warrant — its demonstration threshold is a promotion rule, its withdrawal challenge an ordinary instance — and so is its boundary. The specialization is one-directional: this round begins at promotion and inherits nothing about what happens while a repair is live.

Contents

• ANSWERABLE_REVISION.md — theorem, promotion interface, AR1-AR10, boundaries, the improvement-round specialization, freeze recommendation.
• src/warrant.py — warrants, promotion, the theorem over frozen LE.
• src/cases.py — AR1-AR10 and the specialization.
• tests/ — 29 cases. python3 tests/run.py.

What this does not establish

No Lean and no registered claim.

That criticism reaches promotion at all. A warrant may narrow itself first, and that is legitimate here — the named frontier is reflective openness.

That a defeat is any good. A bare refusal answers a reason structurally.

Responsive Revision, the quantitative follow-up. Not claimed and not built.

Anything about RI realization: promotion is not mapped to a ReasonOcc, and no event kind was added or proposed.