Exceeding a review budget splits the task

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

Archived (see archive README). Retired by ADR 082: the artifact-side bound keeps a claim's joint support on one page, so nothing partitions and the splitting question does not arise. What remains here is the 2026-08-25 current-state anchor, the atomicity of a review pair it recorded, and the partition options weighed against the joint-support hazard — design texture only.

A budget on how much linked material one review pass may open has so far implied one of two bad answers when the budget is exceeded: the reviewer samples, or the pair fails. Sampling silently weakens what a PASS means. Failing punishes an artifact for being well-cited.

There is a third answer, and it dissolves the question rather than answering it: when the evidence exceeds one pass, split the review into passes that together cover it, and combine their outcomes. Coverage is never traded away; only the number of passes varies.

Current state (as of 2026-08-25)

• A review pair is atomic. review_pairs records one outcome per (note_path, criterion_path) per review, and nothing in finalization.py or batch.py expresses partial coverage, subsets, or combination.
• semantic/grounding-alignment now budgets sixteen distinct linked artifacts globally and treats reaching the limit as disclosure rather than failure.
• Available-cost measurement shipped (adbc1cc0). Over 337 artifacts: p50 is 7 distinct artifacts and 67 KB, p90 is 16 and 148 KB, max is 35 and 355 KB. The selected ceiling covers the measured p90 offer.
• Consumed-cost capture ships as soft per-pair telemetry. In a twelve-note capped-versus-uncapped assay, four mechanism-aligned findings appeared only after fuller reading reached 6–16 artifacts. The three uncapped tail cases offering 21–23 artifacts all passed.

Disposition (2026-08-25)

Rejected for now by ADR 079. Raising the ceiling covers every divergent case in the assay without adding partial-coverage, combination, or freshness semantics. The assay supplies no divergent tail fixture above sixteen. It also cannot validate or refute the severed-support hazard below because every capped baseline was PASS. The bounded evidence records the result. This proposal is ready for archival after its inbound references are retired.

What splitting requires

Decomposable coverage. One logical judgment covered by N passes, combined into one outcome: PASS only if every pass passes, FAIL if any fails, WARN aggregating. That is a new concept in the review model, not a gate rule.

A partition rule. Which evidence each pass examines, chosen so the passes together cover everything exactly once.

Combination and freshness. A pair's baseline must pin every pass that contributed to it, or a stale subset silently validates a fresh verdict.

The central hazard: joint support

The obvious partition is by link — pass one takes links 1–5, pass two takes 6–10. It is wrong, and quietly so.

grounding-alignment judges whether cited material supports a claim. A claim whose support spans several sources is supported jointly. Split those sources across passes and neither pass sees the whole support, so each may fail a claim that is in fact adequately grounded — a false FAIL produced by the split itself, not by the artifact.

The corpus is full of such claims. Cohort work routinely found notes citing four or five ingests for one argument, and the rollout's own unit of work was deliberately the claim use, not the link, for exactly this reason.

So the partition must keep a claim's evidence together — it must be by claim, not by link. Which is the difficulty: grouping links by the claim they support is a semantic judgment, and the thing being budgeted is semantic judgment.

Options

A. Inventory then partition. A cheap first pass identifies which links support which claim; later passes each take whole claim-groups. This is the shape the grounding rollout already used successfully — inventory claims first, then split the work — and it keeps joint support intact. Costs one extra pass, and the inventory pass is itself bounded because it reads the note, not the sources.

B. Partition by link, accept severed support. Mechanical and cheap. Produces false FAILs on jointly supported claims, which is the failure mode least likely to be recognized as an artifact of the process.

C. One pass per claim. Maximum isolation, no severing possible. Many passes on a link-dense note, and repeated fixed cost since every pass re-reads the note.

D. Overlapping passes. Let claim-groups share links so no group is severed, at the cost of re-reading shared evidence. Simplest correct partition if the overlap is small; degenerates toward C when it is not.

What this changes elsewhere

It largely supersedes Review budget enforcement is a separable decision, now archived alongside this one. That proposal asked whether exceeding a budget should fail or merely require disclosure. Under splitting, exceeding does neither: it triggers decomposition. The disclosure requirement remains valuable — a pass should still say what it covered — but as provenance rather than as the thing that keeps a sampled verdict honest, because verdicts stop being sampled.

It does not supersede the measurement work. Review link budget prices reviewer attention, also archived, becomes more load-bearing, because a split needs a sizing basis to decide where to cut, and consumed cost is what says how large a pass can usefully be.

Adoption criteria

Reopen when reviews at the sixteen-artifact ceiling produce material outcome divergence in the tail, so a raised single pass no longer covers the observed failure range.

Do not adopt any option before deciding how outcomes combine and how a pair's freshness pins its passes. A split verdict whose provenance is unrecorded is weaker than the sampled verdict it replaces, because it looks complete.