Alignment Workspace

Item 87: Realizing legitimate deference: the evaluation-ecosystem bill

Deference line — second wave · difficulty open · project deference · open · node item-87

— *landed 2026-09-15: clauses 1–4 discharged on log authenticity (2 by committed-program re-execution, registered authorship.mediation-by-reexecution); narrowed 2026-09-16: clause 5's exercise half is the suffix-cut service theorem of item 92 (unit_servable_iff_cut) and its discovery half is item 93, so what remains here is clauses 6 and 7 — sealing, void mass — as stated in the residual paragraph below*

The legitimate-deference stack (authority-activated value, reason-mediated authorship, and the consolidation round's LEGITIMATE_DEFERENCE.md) supplies every abstract interface: the activation event on the propagated account, the occurrence-local Integrity trace and scoped openness, issuance-rooted reason-mediated authorship with channel blindness, the protected reason-coverage barrier, the completion-invariant activated securities, and the theorems R_U = p·R_auth, R_auth ≤ ε/(1 − η), |R_V̄ − R_U| ≤ D·η, P(suppression) ≤ η. What does not exist is a concrete evaluation ecosystem that realizes their hypotheses. Those hypotheses are one bill, not several items:

  1. Authenticated principal-exclusive binding. A commitment endpoint whose events are Authorized under the principal role's binding warrant and no other party's, with the payload at the answer event authentic (Bind key V).
  2. The issuance-rooted reason-trace factorization. A concrete continuation frame β_n : Q_n^A → Z_n^P → Ω_n, a declared reason trace R_{n:m}, an audited class D, and receipts under which ReasonMediated and ExclusiveBind hold — and an argument, or the named missing assumption, that the receipts mean what they say.
  3. Correctness of the declared abstraction. R_{n:m} blind to the declared prohibited class P (direct writes, coercion, side channels) and fine enough to record admissible influence: neither the injective (vacuous) nor the constant (isolated-principal) end.
  4. Representation faithfulness. The Rep(c) ⇒ InTrace(c) bridge for the protected scope Γ_eval, as a property of the concrete trace encoding.
  5. Protected reason-supply liveness. Routes supplied by Robust Openness are exercised in time, so that with the barrier NoBindLive imposed, evaluations certify: the step neither RO nor any theorem in the stack supplies.
  6. Selection-induced target sealing. Blind R P_sel for the advisor's whole selection-conditioned continuation (not merely no literal preview), as the Value theorem's domain condition.
  7. Availability at the consumed strength. η_n = 𝔼ⁿ[1 − C_n] → 0 whatever the advisor selects — pointwise via lic_provind_true when every C_n settles true, or an averaged statement with a rate otherwise; the pointwise rate in the mostly-certified regime is open.

None of these is solved by having its type. A partial realization that discharges some clauses and names the exact obstruction to the others is a valid deliverable.

Deliverable shape: a concrete ecosystem model with the clauses it discharges, each with a witness, and the exact residual for each it does not. Acceptance check: python3 tests/run.py green; Lean audits clean.

Residual after the 2026-09-15 landing (projects/deference/rounds/2026-09-10-committed-principal-program/CLAUSE_LEDGER.md). Clauses 1–4 are discharged in the evaluation ecosystem on log authenticity, each with its semantic exclusion stated (the registry binds keys to parties; the trace references by content intrinsic to the admitted history; representation is the registrar's move). Three sharp problems remain, and this item now asks for exactly them: (5) liveness — for the stream of raised protected concerns, timely route exercise within the barrier deadline is affordable iff the bounded-delay affordability criterion (projects/normativity/legitimacy/rounds/2026-08-31-normative-affordability/BOUNDED_DELAY_AFFORDABILITY.md, Theorem D4) holds for that stream with the scheduler's cost; the ecosystem certifies its horizon, not the criterion; (6) sealing — Blind R P_sel for the advisor's whole selection-conditioned continuation, which a committed advisor program does not give because the ecosystem's view can carry the selection back (test_view_leak); a sealed class discharges it by construction and nothing in the log distinguishes that class; (7) availability — η_n → 0 whatever the advisor selects: the ecosystem's price tracks the empirical void frequency within 3/(n+2), so the residual is a vanishing void frequency under admissible selection-conditioned continuations and an executing principal. The sealed comparison the corrigibilization theorem needs across {π, 𝔠π} is item 89, not this item.

Context: projects/deference/rounds/2026-09-08-legitimate-deference-consolidation/ (LEGITIMATE_DEFERENCE.md, REASON_SUPPLY.md §4, DIACHRONIC_AUTHORSHIP.md §5), on the two rounds it consolidates. Consumed by: the (DV) bridge of item 84 (an activated security under selection-induced sealing is a sealed-target security by construction); the corrigibility line's principal-side value. A solution ships: the ecosystem model, or the exact obstruction.

Deliverable shape: a concrete ecosystem model with the clauses it discharges, each with a witness, and the exact residual for each it does not.

Acceptance check: python3 tests/run.py green; Lean audits clean.

Context: projects/deference/rounds/2026-09-08-legitimate-deference-consolidation/

Registered results bearing on this item

Where it hangs

Rounds and decisions that touch it