079-Grounding reviews budget sixteen distinct linked artifacts

Type: ../types/adr.md · Status: superseded

Status: superseded by ADR 082 Date: 2026-08-25

Context

The semantic/grounding-alignment criterion limited a review to five link occurrences. That rule could charge the same resolved target repeatedly, did not say whether tracked ingests shared the budget, and sat below the median seven distinct artifacts offered across 337 measured targets.

A paired twelve-note assay compared that criterion with an otherwise identical uncapped criterion under isolated gpt-5.4 workers. Five outcomes diverged. One warning recurred in a capped repeat that read all three available artifacts, so it was reviewer sensitivity rather than a cap effect. Fuller reading exposed four other findings after 6–16 artifacts, including two FAIL outcomes. The uncapped cases offering 21–23 artifacts all passed. The population's p90 offer was sixteen distinct artifacts.

Decision

semantic/grounding-alignment may inspect up to sixteen distinct linked artifacts in one pass. The budget is global across linked notes, ordinary linked sources, and tracked ingests. Repeated links to one resolved target consume one slot, and an ingest consumes one slot whether its Quotes route or its declared snapshot route is used. The target note and criterion are not linked material and do not consume the budget.

The limit constrains reading, not verdict semantics. Reaching it is not itself a WARN or FAIL. When a reviewer stops with a material support route unchecked, it names that route and scopes the verdict to what it inspected. The review protocol's consumption record remains provenance for the exact opened paths and whether budget or sufficiency stopped the pass.

Sixteen is an interim per-artifact ceiling justified by this assay and the observed p90 offer, not a derived attention price. The separate proposal to combine artifact count and bytes remains open until a measurement identifies their relative contribution. Split-pass review is not adopted. Reopen it if reviews at the new ceiling produce outcome divergence in the tail.

The operativity path is the hashed production gate criterion. Review prompts embed it as binding judgment instruction, and changing it makes existing pairs stale through ordinary criterion freshness.

Considered alternatives

Keep five. Rejected because four mechanism-aligned findings appeared only under fuller reading, and every one was reachable by sixteen artifacts. Two were material failures rather than differences in explanatory wording.

Remove the limit. Rejected because linked reading still consumes reviewer attention. Uncapped workers opened as many as twenty-two artifacts, while the three cases beyond sixteen produced no outcome divergence. The observed result supports moving the boundary, not deleting it.

Fail a pair when the budget is reached. Rejected because the budget prices a review pass, not note correctness. Automatic failure would penalize link density and would conflate unchecked coverage with demonstrated misgrounding.

Split an over-budget review into covering passes. Deferred. It requires partial-coverage, combination, and freshness semantics that the review model does not have. The assay found no divergent case above the selected ceiling. It also could not test split review's severed-support hazard because every capped baseline was PASS.

Select enforcement per run. Rejected for this decision because the setting would change verdict meaning and therefore review identity. The measured problem can be addressed in the criterion without adding a run dimension.

Adopt a count-and-byte price now. Deferred because capped or uncapped stop points do not identify the exchange rate between artifact-switching and bytes read. Choosing coefficients would restate the inherited-number problem in a more elaborate form.

Consequences

The budget now counts what the prompt and telemetry already expose: distinct resolved artifacts. Ingest applicability and deduplication are explicit, and reviewers can inspect every divergent case observed in the assay without special review-model machinery. Reviews may load more material than before, and existing grounding baselines become stale once because the judgment-bearing criterion changed.

The rule still ignores artifact-size heterogeneity. Reviews above sixteen may remain sampled, so disclosure and consumption telemetry stay necessary. The decision is bounded to the production grounding criterion and to the measured setting: twelve selected source-citing notes, the codex partition, and gpt-5.4. It does not establish sixteen as a capacity limit for other gates, models, or corpora.

Relevant Notes: