Review budget enforcement is a separable decision

Type: ../../types/design-proposal.md

Archived (see archive README). Retired by ADR 082: there is no review budget left to exceed, and the artifact-side bound's fail-or-warn severity is decided in the ADR (WARN, with FAIL deferred). What remains here is the 2026-08-25 current-state anchor and the four enforcement options with the forces weighed against them — design texture only.

Review link budget prices reviewer attention, now archived alongside this one, proposes measuring what a reviewer opens. It does not settle what happens when a reviewer would exceed the budget, and that question is independent: the same measurement supports either answer.

Current state (as of 2026-08-25)

ADR 079 selected option A for semantic/grounding-alignment. The gate now budgets sixteen distinct linked artifacts, makes reaching the limit non-verdict-bearing, and requires any unchecked material route to be named while the verdict is scoped to inspected material. Available and consumed cost remain telemetry rather than review identity.

The paired assay behind that decision found four mechanism-aligned findings that the five-link instruction missed, all within sixteen artifacts. Failing on exceedance would not have recovered them; raising the reading ceiling did.

Disposition (2026-08-25)

Resolved for the current grounding criterion. Advisory disclosure is selected; enforced failure, per-run selection, and a second hard ceiling are rejected for now. The bounded evidence records the outcome and the trigger for revisiting it. This proposal is ready for archival after its inbound references are retired.

The question

Does exceeding the budget make a pair FAIL, or does it remain a disclosure obligation?

The two answers differ in what a verdict means. Under enforcement, a PASS asserts the reviewer covered the artifact within budget. Under disclosure, a PASS asserts the reviewer covered what it judged load-bearing and said what it did not. Both are defensible; they are not the same claim, and a corpus of verdicts means different things depending on which is in force.

Forces

Disclosure is what makes a sampled verdict honest, and it already works unprompted. Job 8051's reviewer named its four unopened links without being asked to. Enforcement adds nothing there; it changes only what happens next.

Enforcement can fail an artifact for being link-dense rather than for being wrong. A well-connected note that cites eight sources responsibly would fail a budget an under-cited note passes, which inverts the incentive the KB wants.

Advisory budgets erode. A number nobody enforces drifts toward decoration, and a reviewer under time pressure has no reason to respect it.

A fixed choice may be wrong for different sweeps. A broad --all-gates pass over hundreds of notes and a targeted single-note review have different tolerances for cost. That argues for selectability rather than a constant.

Options

A. Keep advisory, state it. Write today's convention into the gate: exceed the budget only with disclosure. Cheapest, changes no verdict, makes the existing practice legible.

B. Enforce: exceeding fails. Strongest guarantee about what a PASS means. Risks failing artifacts for density.

C. Selectable per run. A flag or partition setting choosing advisory or enforcing. Fits the differing tolerances of sweeps versus targeted reviews, at the cost of a verdict whose meaning depends on run configuration — which the review store would then need to record, or the corpus becomes uninterpretable.

D. Enforce a hard ceiling above an advisory budget. Disclosure up to the budget, failure past a larger limit. Two numbers instead of one, both needing derivation.

Earlier supersession candidate

Exceeding a review budget splits the task, also archived, proposes a third answer that dissolves this question: exceeding the budget triggers decomposition into covering passes, so a verdict is neither sampled nor failed. What survives here is the disclosure requirement — a pass should still record what it covered — but as provenance rather than as the mechanism keeping a sampled verdict honest.

The cap-lift assay rejected splitting for now, so the options became live and option A was selected.

Adoption criteria

Reopen only if sampled reviews at the sixteen-artifact ceiling produce material tail divergences or disclosure proves insufficient to interpret their verdicts. Do not adopt per-run selection without recording the setting in review identity.