2026-08-12-legitimacy-reorganization

Agent Task: Reorganize the “Leverage” Program Around Legitimacy, Normative Constraint, and Normative Learning

Maintainer: A. M. Berns Model provenance: GPT-5.6 Sol Date: 2026-08-12

Mission

Execute a conceptual and repository-level reorganization of the research program currently organized under the umbrella name “leverage.”

The project has outgrown that name and, increasingly, the finite constructions through which its central ideas were discovered. The purpose of this task is not merely to rename files or rewrite exposition. It is to determine whether a more abstract organization around normative constraint theory, legitimacy theory, and normative learning theory actually yields a cleaner and more mathematically productive theorem architecture.

The task therefore has two inseparable parts:

1. Execute the conceptual reorganization throughout the relevant research artifacts.
2. Stress-test the reorganization mathematically and conceptually, recording where it succeeds, where it fails, and which theorem obligations it exposes.

The task must end in a draft PR containing the reorganization, audit results, theorem map, counterexamples or failure cases, and an explicit assessment of whether the new organization should be retained.

Do not assume in advance that the proposed factorization works. Treat it as a serious conjecture to prosecute.

⸻

I. Current Research Context

The research program began from the leverage problem: how can empirical facts, reasons, warrants, commitments, defeaters, and related structures exert genuine numerical or decision-relevant constraint on normative propositions that do not themselves straightforwardly settle against ground truth?

That investigation produced substantial machinery, including:

• warrants/reasons and applicability conditions;
• commitments and provenance;
• rebutters and undercutters;
• graded/credal constraints;
• constraint propagation / the former “leverage” decomposition;
• LP-dual and underwriting interpretations;
• constrained-marketmaker approaches to Logical Induction compatibility;
• historical answerability;
• migration across conceptual/ontological revision;
• a docket/inquiry direction preventing obligations or questions from silently disappearing;
• reasons-responsiveness;
• attempts to obtain an online-learning theorem, most recently through \Phi-regret-style comparators.

The finite normative-learner constructions have been valuable proof laboratories, but the program now appears to have discovered a more general object.

The emerging central question is:

What makes a trajectory of a bounded reasoner legitimate, and when does movement through the space of legitimate trajectories constitute learning?

This suggests that “leverage” is no longer an appropriate umbrella term.

⸻

II. Proposed Conceptual Architecture

The working proposal to test is the following.

A. Normative Constraint Theory

This replaces “leverage” as the name for the relevant statics.

Normative constraint theory asks:

Given a reasoner in some state and some structure of reasons, evidence, commitments, warrants, objections, defeaters, etc., what constraints do those considerations place on admissible responses or successor states?

At an abstract level, investigate whether this can be represented by something like

\Gamma(x,r)\subseteq X,

where x is a reasoner state, r represents available reasons/normative information, and \Gamma(x,r) is the set or structure of normatively admissible responses.

Do not prematurely commit to this exact representation if the existing mathematics suggests a better abstraction.

The former “leverage” results should be reinterpreted wherever appropriate as results about:

• normative constraint;
• constraint propagation;
• strength of normative constraint;
• feasibility under normative constraints;
• underwriting/enforcement of normative constraints.

The historical term leverage may remain when discussing the origin of the research program or where it names a genuinely specific technical quantity that cannot be replaced without loss. It should otherwise be deprecated as the umbrella vocabulary.

Do not perform blind search-and-replace.

⸻

B. Reasons-Responsiveness (RR)

The proposed local legitimacy condition is reasons-responsiveness.

Very roughly:

x_{t+1}\in\Gamma(x_t,r_t).

Interpretation:

A transition is locally legitimate only insofar as it respects the normative constraints generated by the reasons available at that stage.

The exact definition must be extracted from the existing architecture rather than stipulated from prose.

Determine:

• what information RR must quantify over;
• whether RR applies to belief changes, commitment changes, rule changes, ontology changes, or all of these;
• whether RR is binary, graded, or parameterized;
• whether the existing certification machinery already witnesses RR;
• whether arbitrary changes to the normative standards themselves can satisfy a naïve RR condition.

RR should connect the normative-constraint statics to the dynamics.

⸻

C. Diachronic Answerability (DA)

RR alone appears vulnerable to normative-history laundering: a reasoner might change its standards, language, ontology, applicability rules, or commitments and thereby evade constraints that were live under an earlier state.

The proposed diachronic condition is answerability.

Its intended content is approximately:

Normative liabilities generated earlier in a trajectory cannot disappear merely because the reasoner changes its representation, standards, commitments, or ontology. They must remain identifiable as live, be explicitly and legitimately discharged, or migrate in an accountability-preserving way.

Extract the strongest clean abstract formulation supported by the existing answerability/migration machinery.

Investigate whether DA behaves like a conservation law over normative history.

⸻

D. Legitimacy Theory

The central working hypothesis is:

\boxed{\text{Legitimate trajectory} \approx \text{RR}+\text{DA}.}

More precisely, legitimacy theory should characterize trajectories

\tau=(x_0,x_1,\ldots)

whose individual transitions are reasons-responsive and whose history remains diachronically answerable.

Do not simply define legitimacy as “RR and DA” and declare victory.

Determine what this conjunction actually entails.

The most important mathematical question is:

What useful end-to-end invariant follows from a sequence of reasons-responsive transitions under diachronic answerability?

Search for composition, preservation, conservation, non-laundering, or trust-like results.

⸻

E. Normative Learning Theory

Normative learning theory should become a strict extension of legitimacy theory, not a synonym for it.

The intended distinction is:

Legitimacy determines which changes are permissible. Learning determines whether the reasoner improves through that permissible space.

The current working factorization is approximately

\text{RR + DA} + \text{coverage/inquiry} + \text{online performance criterion} \Longrightarrow \text{normative learning}.

Coverage/inquiry prevents the learner from obtaining trivial success by refusing to encounter, docket, or answer relevant normative demands.

The online-learning component should compare the learner against an appropriate class of legitimacy-preserving counterfactual transformations, rather than arbitrary edits.

The recent \Phi-regret attempt is important context. The conceptual comparator shape was promising, but the current repository representation produced serious technical obstacles, including a growing action set due to ledger IDs and insufficient guarantees against comparator closures capturing prohibited information.

Do not claim a regret theorem that has not been proved.

Instead determine whether the new abstraction makes the desired theorem cleaner to state, what comparator class it naturally suggests, and which representation theorem or decoder would be required to make existing online-learning results apply.

⸻

III. Relationship to Logical Induction

The reorganization must explicitly clarify the relationship between this research program and Logical Induction.

The current hypothesis is that Logical Induction and legitimacy theory constrain different aspects of rational trajectories.

Logical Induction gives a powerful global/economic rationality criterion over a sequence of beliefs/prices: roughly, resistance to systematic exploitation by bounded traders.

The legitimacy program is concerned with something LI does not by itself obviously represent:

whether particular transitions are licensed by reasons and whether the reasoner remains answerable across changes to its own normative/conceptual structure.

Investigate rather than merely assert this distinction.

Clarify at least:

1. Which existing LI properties already imply anything resembling RR or DA?
2. Which do not?
3. Whether legitimacy should be implemented as a constraint on an LI-like process, as an additional layer around LI, or abstractly enough that LI is merely one possible bounded-reasoning substrate.
4. Which parts of the existing constrained-marketmaker / underwriting work remain relevant under the new organization.
5. Whether the normative-learning theorem should itself be LI-derived, ordinary online-learning-derived, or substrate-independent at the abstract level.

Do not force an answer if the mathematics does not yet determine one. Mark unresolved alternatives clearly.

⸻

IV. Deference / Corrigibility Interface

A major motivation for extracting legitimacy theory is that it may supply a missing ingredient in the separate deference program.

The deference problem involves a less capable reasoner H, a more capable advisor/reasoner A, and trajectories in which interaction with A changes H.

The manipulation concern is:

It is not enough that A causes H to improve according to some endpoint criterion. A might alter H’s standards, commitments, ontology, or evaluative machinery so that the resulting state retrospectively endorses a transformation that would not have been legitimate from the prior trajectory.

Endpoint endorsement is therefore potentially self-certifying.

The proposed interface is that the deference theory should be able to consume RR + DA as a legitimacy condition and then add its own trust/deference hypotheses.

Schematically:

\underbrace{\text{RR + DA}}{\text{legitimate trajectory}} + \underbrace{\text{deference/trust conditions}}{\text{inter-agent}} \Longrightarrow \text{legitimate corrigibility}.

Do not attempt to solve the entire deference project in this PR unless a small theorem follows naturally.

Instead, produce a precise interface specification:

• What objects must the legitimacy layer expose?
• What does it certify about an H-trajectory?
• What properties compose from H_0 to H_T?
• What additional trust condition would deference need?
• What manipulation concerns are genuinely ruled out by RR + DA?
• Which manipulation concerns survive and require additional assumptions?

A particularly important possibility to examine is whether the right endpoint property is process trust rather than conclusion endorsement: H_0 need not endorse all beliefs of H_T, but may have grounds to trust that H_T arose through a legitimate, answerability-preserving continuation.

This should be related carefully to the existing deference/trust machinery, including any DDB-like trust condition already present in the workspace.

⸻

V. Theorem Extraction

The reorganization succeeds only if it improves the theorem architecture.

Attempt to extract a compact theorem/conjecture arc approximately of the following form. Change it if prosecution shows a better structure.

1. Constraint representation / propagation

Show how the concrete warrant/commitment/credal machinery induces an abstract normative-constraint structure.

Possible target:

Existing finite normative-constraint models instantiate the abstract constraint interface.

Identify which former “leverage” results become constraint-propagation results.

1. RR realization

Show that the existing certified update machinery satisfies the proposed abstract RR condition, or identify exactly why it does not.

1. DA / conservation theorem

Seek a theorem roughly of the form:

Every outstanding normative liability has at later times either an identifiable descendant, a legitimate discharge, or an explicitly unresolved status.

Formalize this cleanly.

1. Non-laundering theorem

Seek a corollary or stronger result:

A normative liability cannot disappear solely through changes of representation, standards, ontology, bookkeeping, or applicability machinery.

Determine the hypotheses required.

1. Composition / end-to-end legitimacy

This is a priority.

Determine what follows from

x_0\rightsquigarrow x_1\rightsquigarrow\cdots\rightsquigarrow x_T

when every transition satisfies RR and the trajectory satisfies DA.

Seek a useful relation between x_0 and x_T that is stronger than “every local check passed.”

This is likely the key export to deference.

1. Normative-learning theorem shape

State the strongest honest online-learning target supported by the abstraction.

Ideally identify a class

\Phi_{\mathrm{leg}}

of legitimacy-preserving counterfactual transformations and formulate a regret criterion against it.

Do not paper over the known \Phi-regret representation problems.

1. Deference theorem shape

State a precise candidate legitimate corrigibility theorem showing how RR + DA could combine with an appropriate trust/deference hypothesis.

Distinguish proved results from conjectures.

⸻

VI. Adversarial Prosecution

Actively try to falsify the proposed architecture.

Construct minimal examples or countermodels for at least the following attacks.

A. Standard laundering

The reasoner faces an objection under standard C_t, revises C_t\to C_{t+1}, and thereby makes the objection disappear.

Does RR catch this? Does DA?

B. Ontology laundering

The reasoner migrates to a new representation in which an old obligation has no obvious syntactic counterpart.

Can DA preserve the liability without requiring literal vocabulary preservation?

C. Defeater laundering

The reasoner generates or accepts a formally admissible defeater for every inconvenient reason.

Can a trajectory satisfy RR while obviously gaming the reason structure?

If so, is this a failure of RR, of the normative-constraint statics, or something that requires a separate learning/inquiry condition?

D. Agenda manipulation

Advisor A controls which questions, evidence, or cases ever become salient to H, while every actual update is RR and DA-compliant.

If this passes RR + DA, record that fact rather than expanding “legitimacy” ad hoc. Determine whether coverage/inquiry is the correct additional protection.

E. Self-certifying transformation

A changes H’s standards so that later H endorses both A and the transformation itself.

Determine exactly what RR + DA rule out and what they do not.

F. Excessive conservatism

Construct cases of clearly desirable transformative learning.

Make sure RR + DA do not require:

H_0 \text{ already endorses } H_T.

The framework must permit genuine conceptual, evaluative, and epistemic improvement.

⸻

VII. Independence Tests

Determine whether the proposed components are genuinely distinct.

Seek small counterexamples showing, if possible:

• RR without DA;
• DA without RR;
• RR + DA without coverage;
• RR + DA + coverage without learning;
• good online performance without RR;
• good online performance without DA.

These are valuable even if trivial.

They should tell us whether the decomposition

\text{legitimacy} + \text{inquiry} + \text{performance}

has real mathematical content rather than merely verbal modularity.

⸻

VIII. Repository Reorganization

After the conceptual prosecution, reorganize the relevant repository material to reflect the strongest surviving architecture.

The exact file layout should follow the repository’s existing conventions, but the conceptual top level should distinguish:

Normative Constraint Theory

Statics of reasons, warrants, commitments, defeaters, credal constraints, constraint propagation, enforcement/underwriting.

Legitimacy Theory

Reasons-responsiveness, diachronic answerability, migration, conservation, non-laundering, composition.

Normative Learning Theory

Coverage/inquiry, counterfactual evaluation, legitimate comparator classes, regret/online-learning targets.

Interfaces

Logical Induction integration and deference/corrigibility integration.

Deprecate “leverage” as the umbrella name in active roadmap and explanatory material.

Preserve historical provenance. Do not rewrite old audit logs or historical artifacts so that they falsely appear to have used the new terminology at the time.

Where useful, mark terminology explicitly:

Formerly developed under the working name “leverage”; now treated as normative constraint theory.

Avoid gratuitous churn in formal identifiers if renaming them creates implementation risk without conceptual benefit. Record any intentionally retained legacy names as technical debt.

⸻

IX. Required Consolidation Artifact

Create or substantially revise a central research document that allows a new researcher to understand the program without reading the historical sequence.

It should contain:

1. Problem statement
2. Core objects
3. Normative constraint theory
4. Reasons-responsiveness
5. Diachronic answerability
6. Definition or candidate definition of legitimate trajectory
7. Theorem/conjecture map
8. Normative learning extension
9. Logical Induction relationship
10. Deference/corrigibility interface
11. Counterexamples and known limitations
12. Open mathematical obligations
13. Mapping from old “leverage” terminology/artifacts to the new architecture

Keep the abstract theory clearly separated from finite implementations.

⸻

X. Success / Failure Criteria

At the end, explicitly judge whether the reorganization worked.

Strong success

The new architecture yields:

• abstract definitions of normative constraint, RR, and DA;
• at least one nontrivial preservation/conservation/non-laundering result;
• a meaningful composition or end-to-end legitimacy invariant;
• a clean separation between legitimacy and normative learning;
• a cleaner formulation of the online-learning target;
• a precise RR+DA interface consumable by deference;
• concrete counterexamples delineating what legitimacy does not guarantee;
• the finite models reinterpretable as implementations/witnesses rather than the theory itself.

Partial success

The conceptual decomposition is cleaner, but one or both major frontier questions remain unresolved:

1. the end-to-end invariant generated by RR + DA;
2. the online-learning theorem over legitimate comparators.

This is an acceptable outcome if the PR isolates the exact mathematical obstruction.

Failure

Treat the reorganization as unsuccessful or premature if:

• RR cannot be stated independently of implementation details;
• DA does not add a meaningful invariant beyond RR;
• “legitimate trajectory” amounts only to relabeling existing machinery;
• the abstraction makes the learning theorem harder or less precise to formulate;
• the deference project cannot consume RR + DA without importing essentially the entire normative learner;
• adversarial examples show that the intended interpretation systematically outruns the formal properties.

Do not hide a failure verdict.

⸻

XI. Execution Discipline

Proceed in this order:

1. Inspect the current repository/workspace and identify the live canonical artifacts.
2. Produce a map from existing machinery to the proposed architecture.
3. Extract candidate abstract definitions.
4. State theorem candidates before restructuring everything around them.
5. Prosecute those theorem candidates with proofs, proof sketches, model checks, or counterexamples as appropriate.
6. Run the adversarial and independence tests.
7. Revise the conceptual architecture in response to findings.
8. Reorganize the repository around the architecture that survives.
9. Update roadmap/status documents so they accurately distinguish proved, conjectured, blocked, and exploratory claims.
10. Run all relevant tests/checks.
11. Review the diff for historical distortion, terminology churn, broken references, and unjustified mathematical claims.
12. Commit the changes on an appropriate branch, push it, and open a draft PR.

Do not spend the task polishing prose before the mathematical architecture has survived prosecution.

Do not claim a theorem merely because it is plausible.

Do not delete useful finite constructions simply because they are no longer foundational. Reclassify them as implementations, examples, countermodels, or proof laboratories.

⸻

XII. PR Requirements

The draft PR description should include:

• Why: why “leverage” is being deprecated as the umbrella concept.
• New architecture: normative constraint → RR + DA / legitimacy → normative learning, with the deference branch.
• What changed: files/modules/docs affected.
• What was actually proved or mechanically verified.
• What failed under prosecution.
• Counterexamples discovered.
• Status of the RR + DA composition/end-to-end theorem.
• Status of the legitimate-comparator online-learning theorem.
• Exact proposed interface to deference/corrigibility.
• Relationship to Logical Induction after the reorganization.
• Remaining blockers and recommended next theorem target.

The PR should make it possible to answer, from the diff alone:

Did this reorganization merely improve the names, or did it reveal a better mathematical theory?

That is the central question of this task.

⸻

Target conceptual picture

Treat this only as the hypothesis to be tested:

\boxed{\text{Normative Constraint Theory}}

\downarrow

\boxed{ \text{Reasons-Responsiveness} + \text{Diachronic Answerability} = \text{Legitimacy} }

followed by two extensions:

\text{Legitimacy} + \text{Coverage/Inquiry} + \text{Online Performance} \longrightarrow \boxed{\text{Normative Learning Theory}}

and

\text{Legitimacy} + \text{Inter-agent Trust/Deference} \longrightarrow \boxed{\text{Legitimate Corrigibility}}.

The job is to find out how much of this diagram can actually be made mathematically true.

Model provenance: GPT-5.6 Sol

The legitimacy architecture round

Verdict: partial success. The decomposition is real and one part of it is a proved conservation law that already existed under another description. The conjunction under test is not sufficient, and the round exhibits what is missing. The online-learning target is worse off than it looked: the comparator class the abstraction suggests is provably trivial on any trajectory whose constraint responds to the record.

• THEOREM_MAP.md — the verification register: statements, derivations, scopes, and what is not established.
• PROSECUTION.md — the six attacks and the independence matrix, with the test that decides each.
• src/ — the abstract conditions, the fixtures, the conservation sweep, the comparator analysis.
• Legitimacy wiki — interpretation and philosophical gloss.

python3 tests/run.py

Nothing here is registered in CLAIMS.md and nothing is in Lean. The four candidate port statements are named in THEOREM_MAP.md.