Normative Induction

Normative induction and Progress

Status: canonical. The Progress statistic, the edge-local practical certificate, and the finite three-term bound on the accounted export are Established lean-proved, with a nonvacuity witness in which all three terms are positive. The practical semantics and the evaluation protocol are typed inputs.

Legitimacy says what a trajectory owes and to whom. Normative induction is the quantitative half: given the obligations a legitimate process has incurred, does what it owes come to bear on what it does, at a rate, against an accounting nobody can rig after the fact?

The handoff

Legitimacy exports the accounted obligation state O_P: the exposed occurrences with their immutable anchors, the live docket, and the proof-relevant account of each occurrence. It exports no weights, importance, service intensities, probabilities, securities or market geometry. Which obligations matter more is decided downstream by an evaluation protocol the application declares, so that the process which incurs obligations cannot also set the terms on which it is scored.

Two views of the same state serve different consumers. The live docket is what remains owed now — what a scheduler consumes. Historical exposure is everything that ever entered the process's responsibility — what evaluation consumes. A docket cleared by disposal is not a clean record.

The evaluation

An evaluation of O_P is indexed by the state and reads two of its fields: the exposed occurrences and their anchors. It never inspects the account. It declares:

• an evaluation measure μ over exposure and a transport plan T from exposures to service occurrences, both committed before the responses they score are observed — a learner does not choose its test distribution after seeing its mistakes;
• for each service occurrence s, the one response Π_s actually realized there from the market state, and a public operative defect d_s;
• an anchored loss Λ_{r,s} of a response, as a function of the anchor r — two occurrences with one anchor get one loss and two transport rows, so multiplicity is in μ and T, never in the loss;
• a worst loss D and the certificate constants.

The statistic

Progress = Σ_{e,s} T(e,s) · Λ_{anchor(e), s}(Π_s) + D · (1 − Σ_{e,s} T(e,s))

The transport-weighted loss of every served edge against the response realized at its service, plus the unserved evaluation mass charged at the worst loss. This is the realization round's statistic. An exposure-level headline loss with one number per exposure is a different endpoint — an exact witness separates the two — and the theorem about it is an optional corollary, not the canonical bound.

The practical certificate

For each edge with positive transport,

PracticalCert(e, s, Π_s): Λ_{anchor(e), s}(Π_s) ≤ M_es · d_s + ε_es .

One service occurrence realizes one response, and every exposure transported to it certifies against that same response. This is where joint practical-response compatibility lives: two obligations can share a feasible operative region while no single response is adequate for both, and then one of them is not certified and its mass is residual. Value correspondences, approximate optimizers, finite policy menus and adequate-set semantics are sufficient ways to produce the certificate; none is public structure.

The bound

Under nonnegative transport and defect, PracticalCert on every served edge, the uptake certificate λ_s d_s² ≤ ρ_s for the intensity λ_s spent at each service, and the amplification bound Σ_e T(e,s) M_es ≤ Γ · λ_s / Σλ:

Progress ≤ Γ · √(Σ_s ρ_s / Σ_s λ_s) + Σ_{e,s} T(e,s) ε_es + D · r ,

with r = 1 − Σ T the residual mass, which lies in [0, 1]. Three terms, three failures: serviced constraints not taken up or amplified; decision or semantic error in transported responses; evaluation mass left unserved.

This is a sufficiency theorem. Any realization that supplies the named certificates gets the bound.

The decision interface

**Status: the static interface is Lean-proved and unregistered; the margin is a typed input with two routes; the dynamic case is open. One realization now derives the whole uptake package rather than assuming it: the rowwise warrant realization of the Normative Inductor (inductor.warrant-inductor <a class="ref claim" href="/claims/inductor.warrant-inductor/">[lean-proved, active]</a>) produces the defect as the displayed inadequate excess, the practical certificate from certified exclusion and margins at the realized market through the soft gate, and the work relation from the certificates' mass times the tolerance; its adequacy semantics — the adequate set and the anchored losses — stay typed inputs, as this page says they must.** What the bound above asks of decision theory is the practical certificate on each supported edge, and the adequate-set route factors that into a semantic half — adequate responses cost at most ε_ad, every response at most D — and one decision-theoretic half: the realized response distribution's mass off the adequate set is affine in the public defect. The static bill is therefore a continuity/coupling condition, not expected-utility maximization.

adequacy semantics → compiled region → market proximity → sound stable adapter → PracticalCert → Progress

• Adapter abstraction. Any map from displayed scores to response distributions that is sound at a region point (mass off the adequate set at most θ) and Lipschitz in the scores at rate κ per unit sup-distance satisfies the coupling, and then the certificate with M = D κ, ε = ε_ad + D θ. Every arrow from the adapter to Progress is a literal hypothesis match.
• Soft gate, one realization. Weight each response by a ramp of width δ on its adequacy score above a threshold τ, times a task preference, and normalize. With the region excluding inadequate responses at τ and a margin of adequate preference mass W above τ + 2δ, the mass off the adequate set is at most the inadequate preference mass over W, per unit of defect relative to δ. The natural hard gate — task-argmax over responses priced above τ — is sound and has no Lipschitz constant: at every positive defect it can execute an inadequate response with mass one. Continuity is the whole decision-theoretic content.
• Positive margin is semantic or epistemic, not decision theory. Region soundness excludes; it marks nothing. The margin comes by route A — the compiler positively marks an adequate response at the region point, a completeness condition — or by route B — the market displays one on its own, which a learning theorem about prices could supply asymptotically and without rate. Both routes are pointwise at the occasion and apply to adequacy that is derivable or externally certified. Adequacy known only through later settlement is outside the static theorem; the bridge from average calibration to average inadequate-action mass is not proved.
• Charged through amplification. The constant M = D κ enters Progress through the amplification hypothesis, so the exact end-to-end condition is the existing one; the local reading is that market error must shrink relative to the certified margin with the inadequate-to-adequate preference-mass ratio controlled.
• Not part of the definition. The soft gate is a realization of the adapter Pi, not a component of the Normative Inductor; the Inductor's substrate and the evaluation are unchanged by it. Task competence inside the adequate set is the application's; a bounded-inductive-rationality learner on the gated sequence supplies it for a supplied decision-problem sequence.
• Inquiry. When no response is confidently adequate the gate returns an inquiry response, charged at D through the residual and never a violation; in the Lean this is a wrapper silent under margin, and the cleaner semantics compiles "inquiry is adequate" from a certified conflict so that inquiry is an ordinary response.
• The competence layer is the lexical Continuation BRIA auction. A block's realized score is her later evaluation of the block that happened, gated by legitimacy, less ϖ per violation detected after the fact and attributed to the continuation; a continuation is evaluated as a hypothesis's bid on the residual less ϖ per structurally recognized violation less ϖ times the market's prices of the shortfall and taint events; the winner is settled on the realized residual with the prices added back (LEAN BRIACorrigibility). A declared violation loses at every block by ranges alone, no calibration overturns it, and the priced risk enters as an exchange rate with inquiry absorbing it.
• The permission layer is that term's advance-recognition face. In front of the adequacy gate sits a permission weight: one on inquiry, zero on a declared violation of the allocation — recognized structurally, not from prices — and otherwise the product of two ramps (one below a lower threshold, zero above an upper one) on the market's prices of the shortfall event and the taint event. It is slack for the lexical auction (filter_slack, forecast_slack) and kept as defence in depth. For any bounded preference in place of the auction the composed gate multiplies it with the adequacy ramp and normalizes with a mass floor on inquiry; it reads prices and adds no trader, so the market is untouched; a declared violation gets probability zero at every day whatever the preference; mass is zero above the upper threshold and bounded by the ramp between; soundness composes as the sum of the two layers' error terms and continuity as the product of the ramps with the inquiry floor keeping the normalization Lipschitz, so the practical certificate — and Progress — hold with the composed constants M = D·κ, ε = ε_ad + D·(θ_inq + θ_perm) (LEAN DecisionComponent) — the general theorem. Generability of the two priced events is a named hypothesis (items 101 <a class="ref item" href="/items/101/">[item 101, open]</a> and 90), not a component of the bound.
• The dynamic layer. Progress is agnostic about how responses are selected. When an admissible act changes the next admissible set, a myopic gated learner is safe and not competent against legitimate continuation policies; the learner that supplies temporally extended responses is Continuation BRIA, which controls learning error against accountable continuation claims on every non-dominant block schedule, and leaves promise recognizability and joinability as the external bridges (Corrigibility).

What is billed

The evaluation protocol — μ, T, D — is the application's. The truth of each PracticalCert is the practical-semantics contract: that Λ scores the response against the immutable anchor rather than a later substitute, that the map from the realized response to the anchored response space has the declared causal meaning, and the inequality itself. Logical Induction prices what settles; it does not determine counterfactual policy values, and the theory does not derive them from market conformance.

---

Evidence. The evaluation on the accounted state, PracticalCert, the uptake package and the bound are [NormativeInductionInterface.lean](https://github.com/A-M-Berns/alignment-workspace/blob/c24159764974232dfd5b47a10b671f12d6f9c244/lean/Workspace/Normativity/Contrib/NormativeInductionInterface.lean), registered as ni.progress-bound <a class="ref claim" href="/claims/ni.progress-bound/">[lean-proved, active]</a>; the finite algebra and the endpoint separation are [NormativeInductorComposition.lean](https://github.com/A-M-Berns/alignment-workspace/blob/c24159764974232dfd5b47a10b671f12d6f9c244/lean/Workspace/Normativity/Contrib/NormativeInductorComposition.lean), registered as progress.edge-bound <a class="ref claim" href="/claims/progress.edge-bound/">[lean-proved, active]</a>, progress.edge-bound-quadratic <a class="ref claim" href="/claims/progress.edge-bound-quadratic/">[lean-proved, active]</a> and progress.headline-separation <a class="ref claim" href="/claims/progress.headline-separation/">[lean-proved, active]</a>. The minimal certificate and the joint-compatibility analysis are the practical certificate round's [CERTIFICATE.md](https://github.com/A-M-Berns/alignment-workspace/blob/c24159764974232dfd5b47a10b671f12d6f9c244/projects/normativity/legitimacy/rounds/2026-09-05-practical-certificate/CERTIFICATE.md) <a class="ref round" href="/rounds/2026-09-05-practical-certificate/">[round]</a> and [JOINT_COMPATIBILITY.md](https://github.com/A-M-Berns/alignment-workspace/blob/c24159764974232dfd5b47a10b671f12d6f9c244/projects/normativity/legitimacy/rounds/2026-09-05-practical-certificate/JOINT_COMPATIBILITY.md) <a class="ref round" href="/rounds/2026-09-05-practical-certificate/">[round]</a>. The decision interface — the adapter abstraction, the soft gate with its sharp constant, the hard-gate witness, the two margin routes and the inquiry wrapper — is [GatedChoice.lean](https://github.com/A-M-Berns/alignment-workspace/blob/ab260c0eade2f39de7c06a5ac58649945d66c9ab/lean/Workspace/Normativity/Contrib/GatedChoice.lean), with the register in the decision-theory-bill round's [DECISION_THEORY_BILL.md](https://github.com/A-M-Berns/alignment-workspace/blob/ab260c0eade2f39de7c06a5ac58649945d66c9ab/projects/deference/rounds/2026-09-06-decision-theory-bill/DECISION_THEORY_BILL.md) <a class="ref round" href="/rounds/2026-09-06-decision-theory-bill/">[round]</a> and [NORMATIVE_CHOICE_THEOREM.md](https://github.com/A-M-Berns/alignment-workspace/blob/ab260c0eade2f39de7c06a5ac58649945d66c9ab/projects/deference/rounds/2026-09-06-decision-theory-bill/NORMATIVE_CHOICE_THEOREM.md) <a class="ref round" href="/rounds/2026-09-06-decision-theory-bill/">[round]</a>; its declarations are unregistered. The fixed-era instances — one era, one settled semantics — are on Progress, Serviceability and Liability and affordability, and are realization material rather than part of the generic theory.