Steering Exposure Statement
How far can steering move a contributor's expected credit, and so their payout, while every individual receipt stays honest? A bound per contributor, tested under a protocol frozen before evaluation, with an independent verifier and a simulated settlement.
Decision
PASS
frozen protocol, one evaluation run
Image credit
95.5%
of 3,072 pairs favor χ²
Single-decision credit
95.3%
of 2,688 pairs favor χ²
Integrity
0 violations
across 6,480 pairs; audit 64/64
Tamper battery
12/12
verdicts as declared
Settlement
$266.89
expected to move per $10,000; 8,000 receipts verified
1. The primary decision
An earlier result in this research proves an outcome-only steering tilt leaves every receipt unchanged while expected credit moves by a covariance. The statement bounds that movement by sd_base(credit) × √χ²(steered‖base). The protocol asked whether that analytic bound beats the strongest total-variation statement (the exact worst case at the same TV distance) by at least 10% in half of all contributor–steering pairs, in both model classes.
Each curve is the cumulative distribution of the analytic χ² bound divided by the exact TV worst case, over 8 evaluation models per class and all declared steering families. The vertical rule marks the 0.9 decision threshold.
2. The qualification: no single budget is enough
The χ² bound clears the TV test, but the density-ratio cap is often tighter. For every indicator filter the ratio-cap class sits inside the χ² ball of the same size (MATH.md §5), so the ratio worst case can never exceed the χ² worst case there. For smooth tilts on image credit, χ² is usually tightest. A statement therefore reports the minimum of the three exact worst cases. For filters, all three budgets follow from the acceptance rate the sampler already logs.
Share of pairs where each budget yields the smallest exact worst-case drift. Ties are assigned to χ², then to the ratio cap.
3. Settlement rehearsal: honest receipts, moved money
Evaluation model A0 with near-copy rejection (a memorization-guard style filter, acceptance 81.5%). Each of two runs (unsteered, then steered) splits its own simulated $10,000 budget over 4,000 outputs by receipt credit. The verifier recomputed all 8,000 receipts independently and every one matched. Nothing about any receipt is false, yet the filter is expected to move $266.89. The largest expected loss is dark-13 at $69.99, inside its stated exposure of ±$201.84.
Contributors sorted by expected change. Realized changes include sampling noise (standard errors $3.83 to $19.99 across components); the statement bounds the expected change. Pools not shown in the chart:
Most of the moved money goes to the model pool (the learned baseline), because rejecting near copies removes exactly the outcomes where retrieval credit is concentrated.
4. Steering that sees the component
Receipts stay valid only when the steering depends on the outcome alone. The auditor classifies per-step processes from their recorded tables, not their names. Where receipts change, a statement asserts no exposure and flags the receipts.
5. Independent verifier and tamper battery
The verifier recomputes every probability and credit through a separate per-outcome code path, minimizes the χ² dual by golden-section search, and computes TV and ratio worst cases by threshold bisection. It returns VERIFIED, REFUSED or ABSTAIN, following the Clearance Certificate convention.
6. What this does not establish
Credit is component-posterior accounting in a declared mixture, not causal influence, deletion effect, authorship or entitlement. Both model classes are small and exactly enumerable; trained generators need estimated budgets and credit variance, which remain open. The inequalities are classical; the contribution is their application, exact computation with certificates, and the comparison. Statements are hash-bound and unsigned. The full mathematics, protocol and results are kept in the project archive.