Alignment Workspace

2026-08-12-corpus-reconciliation

2026-08-12 · project deference · path projects/deference/rounds/2026-08-12-corpus-reconciliation

Prompt prompts/2026-08-12-corpus-reconciliation/PROMPT.md

Deference corpus reconciliation — August 2026

Maintainer: A. M. Berns Model provenance: GPT-5.6 Sol Repository: A-M-Berns/alignment-workspace

Purpose

The repository has just ingested Abram Demski’s new deference source corpus at:

projects/deference/note-dump-2026-08-11/

This corpus supersedes projects/deference/note-dump-2026-06-27/ as the current source material, while preserving the June tree as historical provenance.

However, the workspace’s canonical deference/corrigibility surfaces were developed largely against the June corpus and subsequent independent Stage I–V work. The August corpus contains corrections to earlier claims, substantial new results, a revised internal wiki of record, new work on Total Trust / Value / Tower relationships, varying-question deference, faithful acceleration, and early legitimacy/corrigibility ideas.

Run a reconciliation pass answering:

Given the newly ingested August 11 Demski corpus and the workspace’s independently developed deference/corrigibility work, what should the workspace now regard as the current research state, and what changes—if any—should be made to its canonical roadmap, ledger, priorities, and decision surfaces?

This is not a request to blindly promote the new corpus into canonical status.

The corpus is source material. Adjudicate it.

Current workspace architecture

The active deference/corrigibility line lives primarily under:

  • projects/deference/README.md
  • projects/deference/notes/CORRIGIBILITY_PAPER_LEDGER.md
  • projects/deference/notes/CORRIGIBILITY_ROADMAP.md
  • projects/deference/notes/LI_NATIVE_DEFERENCE.md
  • projects/deference/notes/LI_NATIVE_DEFERENCE_FOR_HUMANS.md
  • projects/deference/notes/FINITE_MODEL_SKELETON.md
  • projects/deference/notes/DISPATCH_QUEUE.md
  • projects/deference/notes/CARTESIAN_FRAMES_DEFERENCE_BRIDGE.md
  • PRIORITIES.md
  • DECISIONS.md
  • RESEARCH_STATE.md

The previous inherited source corpus remains at:

projects/deference/note-dump-2026-06-27/

The new source corpus is:

projects/deference/note-dump-2026-08-11/

Read its ORIGIN.md, top-level README/index materials, wiki-of-record surfaces, audit/status material, and the substantive files necessary to establish what changed.

Do not assume every file in the 2026-08-11 tree has equal authority. Respect the corpus's own hierarchy, status labels, provenance warnings, and supersession rules.

Research-program context

The deference line asks when and how a bounded agent can defer to another agent without simply surrendering final authority.

The motivating thesis is roughly:

Epistemic superiority need not entail final authority.

The longer-term target is a meaningful corrigibility theorem: a theorem explaining when deference to a stronger reasoner preserves some relevant capacity of the principal to remain authoritative, correct, revise, recover, or avoid being foreclosed upon.

The workspace has learned several important negative lessons.

1. Logical-Induction deference

Earlier work studied relations among a slower/principal inductor, a faster/advisory inductor, and updated or augmented agents.

The line includes notions such as:

  • Total Trust;
  • Value;
  • Tower-like relations;
  • faithful acceleration;
  • delayed prediction / later computation;
  • cross-agent calibration and forcing;
  • deference to another inductor;
  • successor/future-agent constructions.

A standing issue has been separating algebraic implications from genuine consequences of the Logical Induction Criterion.

The June corpus’s own audit found that several forcing results imported their crucial forcing inequalities as hypotheses rather than deriving them from an explicit market/trader construction.

The workspace therefore opened PRIORITIES.md items 7–9 around:

  • modelling the actual market and traders;
  • the doubly-soft weight class;
  • forcing headlines that were squeezes over hypotheses too close to the conclusion.

Subsequent workspace work partially closed item 7 by building an actual FAF trader/criterion-to-signed-forcing chain, while leaving cross-process residue.

2. Corrigibility / successor work

The workspace then tried to move from epistemic deference to a richer corrigibility question.

Several comparator constructions failed for a structural reason:

Two authorization regimes that induce the same realization map are indistinguishable in a type whose only semantically relevant output is that realization map.

This led to a type-level diagnosis: corrigibility requires representing something like authorization, jurisdiction, corrective capacity, or future control explicitly rather than merely decorating an ordinary choice function.

The current workspace therefore distinguishes:

  • epistemic/value deference;
  • execution;
  • authority/jurisdiction;
  • future corrective capability;
  • foreclosure.

PRIORITIES.md Q3 currently asks how foreclosure should even be expressed.

3. Cartesian Frames bridge

A recent round explored Cartesian Frames as a candidate language for representing what is lost when future corrective agency is restricted or transferred.

That round produced useful candidate objects but did not solve Q3:

  • a frame has no intrinsic time coordinate;
  • the relevant authorization relation is still missing;
  • “future” remains supplied externally by the modeller;
  • transfer/restriction structures represent counterfactual control better than entitlement.

So this is promising interface work, not yet the corrigibility theorem.

4. Current epistemic discipline

The workspace distinguishes sharply among:

  • inherited/source-corpus claims;
  • model-generated prose;
  • kernel-verified declarations;
  • registered workspace claims;
  • genuinely workspace-established results.

Do not collapse those distinctions.

A theorem being present in Abram's corpus does not automatically make it a workspace-established result.

A theorem being kernel-checked in the imported corpus does not automatically make it registered in this repository.

What changed in the August 11 source corpus

At minimum, investigate the following apparent changes. Do not trust this list as complete; verify against the new corpus.

Faithful acceleration

The new corpus contains a substantial adjudication of earlier faithful-acceleration claims.

It reports errors in earlier documents, distinguishes surviving and refuted claims, and records corrected positive and negative results.

Determine exactly what this does to:

  • the workspace’s inherited-status ledger;
  • the description of the June corpus;
  • existing roadmap assumptions;
  • PRIORITIES.md items 7–9;
  • any claims currently described using outdated faithful-acceleration summaries.
Total Trust / Value / Tower

The corpus appears to contain a considerably more developed structural picture connecting:

  • Total Trust;
  • Value;
  • Tower;
  • Mart or martingale-like conditions;
  • centered-bet squeeze arguments;
  • conditional-stability requirements;
  • counterexamples to stronger implications.

Determine which parts:

  1. merely sharpen results already represented in the workspace;
  2. supersede old descriptions;
  3. create new reusable lemmas;
  4. alter what the major open theorem actually is.
Varying-question deference

The corpus includes positive and negative results for varying-question settings, including a theorem labelled roughly SS and an adversarial/reversal construction.

Determine whether this opens a distinct research track, closes an existing implicit question, or should simply become context for another priority.

Attribution correction

The August 11 corpus records a same-day correction concerning what had been called an “Eisenstat conjecture.”

Treat this carefully.

The corpus reportedly establishes that its previous formulation was its own construal and that Sam Eisenstat disputed that construal as representing his intended claim.

Ensure no workspace document perpetuates the old attribution.

Do not assert Sam Eisenstat's actual position beyond what the source supports.

Legitimacy / corrigibility

The corpus contains early material relating trust, deference, corrigibility, legitimacy, and endpoint preservation.

This is particularly relevant to the workspace because the workspace independently arrived at a richer authorization / foreclosure / corrective-capacity program.

Compare these carefully.

Ask:

  • Are they addressing the same object?
  • Is Abram’s “legitimacy” notion an input to the current corrigibility architecture?
  • Does it supply a possible theorem statement?
  • Does the workspace’s type-level authorization diagnosis reveal a missing piece in the source proposal?
  • Does the new corpus make Q3 more precise?
  • Is there now enough structure to graduate an ingenuity question into an executable PRIORITIES.md item?

Do not force a synthesis if the notions are genuinely different.

Required work

A. Build a source-of-truth map

Produce a compact mapping of the major relevant claims/topics across:

  1. June source corpus;
  2. August source corpus;
  3. current workspace ledger/roadmap;
  4. current workspace formal work.

For each major topic classify the relationship as one of:

  • unchanged;
  • sharpened;
  • superseded;
  • corrected;
  • refuted;
  • newly added;
  • unresolved conflict;
  • source-only / not yet adopted.

The classification must be supported by exact paths.

Do not rely on remembered theorem names.

B. Audit the current canonical surfaces

Inspect:

  • projects/deference/README.md
  • CORRIGIBILITY_PAPER_LEDGER.md
  • CORRIGIBILITY_ROADMAP.md
  • LI_NATIVE_DEFERENCE.md
  • DISPATCH_QUEUE.md
  • PRIORITIES.md
  • DECISIONS.md
  • RESEARCH_STATE.md

Identify every place where the workspace currently:

  • points to the June corpus as if it were still the current source;
  • describes a result now corrected or superseded by August;
  • omits a major new result which materially changes the research state;
  • uses an attribution corrected in August;
  • carries an open problem whose premise has changed;
  • carries a priority whose target should be reformulated;
  • has already independently rediscovered or surpassed something in the new corpus.

Do not change prose merely for freshness. Change only materially stale research state or pointers whose meaning has changed.

C. Reassess PRIORITIES.md

Pay special attention to deference items 7–9 and the ingenuity questions.

For every existing deference priority, answer:

  • Does it remain correctly posed?
  • Has August work partially or fully closed it?
  • Has its best formulation changed?
  • Is its acceptance condition still appropriate?
  • Has some newly important question become more urgent?
  • Has any apparent new question already been answered by existing workspace work?

Do not inflate the priority list.

A good reconciliation may leave most priorities unchanged.

If a new priority deserves filing, it must meet the repository's normal standard:

  • precise target;
  • self-contained context;
  • deliverable shape;
  • executable acceptance check where possible;
  • honest difficulty classification.

If the missing ingredient is still an idea rather than executable work, put it under Where ingenuity is the bottleneck, not as a numbered work order.

D. Reassess the corrigibility target

Give special attention to the relationship between:

  1. Abram’s August legitimacy/corrigibility material;
  2. the workspace's authorization/jurisdiction diagnosis;
  3. the Cartesian Frames bridge;
  4. Q3 on foreclosure;
  5. the desired eventual corrigibility theorem.

Try to state the strongest currently defensible picture of the target theorem.

It is acceptable—and useful—to conclude that no theorem statement is ready.

If one is now ready enough to dispatch, state:

  • primitives required;
  • hypothesis package;
  • conclusion;
  • why it is not trivial or circular;
  • what formal substrate it should live on;
  • what would count as a counterexample;
  • what first executable round should attempt.

Do not call something a “corrigibility theorem” merely because it contains a future principal and a deference relation.

E. Update the canonical surfaces where warranted

You have write scope for this reconciliation round.

Make minimal, principled edits to the canonical surfaces where the audit justifies them.

Likely candidates include:

  • projects/deference/README.md
  • projects/deference/notes/CORRIGIBILITY_PAPER_LEDGER.md
  • projects/deference/notes/CORRIGIBILITY_ROADMAP.md
  • projects/deference/notes/DISPATCH_QUEUE.md
  • PRIORITIES.md
  • RESEARCH_STATE.md

Modify DECISIONS.md only if:

  1. an actual maintainer-level research/governance choice has already been resolved by the round's evidence and falls within delegated scope; or
  2. a genuinely non-mechanical choice must be reserved to the maintainers.

Do not use DECISIONS.md as an activity log.

Where a maintainer decision remains necessary, add it to the single Awaiting the author queue with:

  • the exact choice;
  • evidence needed;
  • cost of deciding now;
  • cost of waiting.

Remember that Abram Demski and A. M. Berns are co-equal maintainers under repository governance.

Epistemic/status discipline

This is the most important constraint.

Do not turn:

“Abram’s current source corpus says X”

into:

“the workspace has established X.”

Use the repository's actual status vocabulary.

Distinguish at minimum:

  • source-corpus result;
  • inherited-established;
  • architected;
  • kernel-verified but unregistered;
  • conjectured/open;
  • workspace-established.

If the new corpus's own status is uncertain or internally disputed, preserve that uncertainty.

Same-day August 11 material should receive especially careful provenance treatment.

A claim corrected by the corpus should be corrected in the workspace even if the replacement is not yet adopted as a workspace theorem.

Do not do these things

Do not:

  • rewrite the August corpus;
  • mutate the frozen June or August source trees;
  • “clean up” imported source material;
  • blindly make the August wiki canonical for the workspace;
  • promote imported Lean declarations into workspace claims without the repository's registration requirements;
  • duplicate results already established by Stage I–V;
  • silently erase historical decisions;
  • dispatch a new proof project during this round;
  • overfit the workspace to theorem names or vocabulary unique to the imported corpus;
  • treat Abram's informal use of “corrigibility” or “legitimacy” as automatically identical to the workspace's current concept.

This is an adjudication pass.

Deliverables

Create a round directory under an appropriate dated path such as:

projects/deference/rounds/2026-08-12-corpus-reconciliation/

and preserve the dispatched prompt under the repository's standard prompt/provenance machinery as required by AGENTS.md.

The round should ship:

1. RECONCILIATION.md

The detailed source map.

For every material divergence, record:

  • old workspace position;
  • August source position;
  • evidence paths;
  • adjudication;
  • resulting workspace change or reason for no change.
2. FOR_HUMANS.md

A concise explanation of:

What did Abram's new corpus actually change about our understanding of the deference/corrigibility program?

Include a final section:

What should we work on next?

Rank only the top 1–3 research moves.

3. Canonical-surface edits

Only those justified by the reconciliation.

4. REPORT.md

Include:

  • files read;
  • files changed;
  • major adjudications;
  • old claims corrected;
  • new source results not adopted and why;
  • priorities added/removed/reworded;
  • decisions taken;
  • outstanding maintainer actions;
  • what the round does not establish.

Acceptance criteria

The round succeeds if, after it:

  1. A new researcher can read the canonical deference surfaces without needing to know that a newer Abram corpus was merged yesterday.
  2. Nothing in the canonical status story is known to rely on a result the August corpus retracts or materially corrects.
  3. The workspace does not overclaim imported results.
  4. The relationship between Abram's latest work and the workspace's Stage I–V work is explicit.
  5. PRIORITIES.md reflects the actual current bottlenecks rather than historical June bottlenecks.
  6. The corrigibility/legitimacy/foreclosure target is stated as sharply as the evidence currently permits.
  7. Historical provenance remains intact.
  8. All repository gates pass.

Final research question

End the report with your best answer to this question:

After incorporating the August 11 Demski corpus as evidence—but not automatically as canon—what is the strongest coherent picture we now have of the deference/corrigibility research program, and what single next piece of research would most increase our ability to prove a genuinely nontrivial corrigibility theorem?

Do not answer this aspirationally.

Answer from the reconciled evidence.

Model provenance: GPT-5.6 Sol

README projects/deference/rounds/2026-08-12-corpus-reconciliation/README.md

Corpus reconciliation

The verification register is RECONCILIATION.md. The Deference wiki is the human register.

Claims it registered

None.

Items it filed or answered

None recorded in the index.

Nodes pointing at it