082-Grounding is bounded on the artifact by unquoted sources per note

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

Status: accepted Date: 2026-08-27

Context

Commonplace placed its grounding reading limit on the review. semantic/grounding-alignment budgeted sixteen distinct linked artifacts per pass, and sentence/misleading-link-text and sentence/concept-attribution each read at most five distinct targets. Every one of those numbers tracks a corpus statistic, so it has to be recalibrated whenever the corpus's evidence needs move. When it lags, the failure is silent: a reviewer that stops at its limit and passes is indistinguishable from one that read everything and passed.

The two representation caps had a sharper problem. Under the note-link exemption those criteria are what makes the exemption sound, and 55% of kb/notes/ links more than five distinct notes. The cap bound on the majority of the corpus, so the exemption's precondition — that the representation check actually runs — was false for most notes.

A third force is auditability. A verdict resting on fourteen opened artifacts costs as much to audit as it cost to produce, so it is trusted rather than re-checked, and Commonplace's own measurement of verdict instability on unchanged input is why re-checking is worth paying for.

Quote verification carried a related gap. A span marked verbatim was resolved against the whole linked file, so a quotation aimed at a tracked ingest could be satisfied by that ingest's analytical prose, which ADR 073 holds is not source support.

Decision

This adopts the proposal Atomic-step grounding bounds unquoted sources per note.

  1. Grounding is bounded on the artifact, not the review. A note may cite at most five distinct tracked sources (kb/sources/*.ingest.md) without a verified verbatim quotation paired to each. A source the note quotes verbatim is discharged and does not count. A source linked with (snapshot required) anywhere in the note always counts: the marker attaches per source rather than per link, because one such use already forces the reviewer to open the snapshot. Links to other notes and to other library artifacts never count. A linked note has passed its own grounding review, and the citing note owes it representation, not re-grounding.
  2. Enforcement is a commonplace-validate type rule on notes, severity FAIL. It shipped as WARN while the eight notes then over the bound were conformed and the two authoring conventions named below were written into the notes collection contract; both were done the same day, and the rule fails from then on.
  3. Quote verification is confined for ingest targets. A verbatim quotation aimed at an *.ingest.md must occur in that ingest's ## Quotes section, not anywhere in the file. Otherwise the exemption could be earned by quoting the half of the ingest that is not source support. A quotation that wraps across source lines inside one paragraph is now recognised; before this decision such spans were never verified at all, and an unverified span does not discharge a source.
  4. The three link-following criteria lose their reader-side caps. semantic/grounding-alignment drops the sixteen-artifact budget and its disclosure clause. It reads a linked library artifact head-first for representation — title and opening paragraph, and the target's own treatment of a concept only when the invoked claim is an interior concept the note does not quote. It judges quoted sources on the note's page and keeps the two existing routes for unquoted sources. sentence/misleading-link-text checks every link and sentence/concept-attribution every identity claim, under that same head-first rule. A head check needs no N, because auditability — the reason N exists — is trivially satisfied when a second checker can redo the step from the same heads. The paraphrased-interior case in concept attribution remains a finding-step outside the artifact-side bound. It is tolerated because the sentence shape is rare, no note carrying more than four such claims, with quote-or-title as the author-side remedy if it stops being rare.
  5. Five is a convention chosen for auditability, not a measured degradation point. What a small step buys is that a second checker, including a person, can redo the grounding step from the note's own page; the uncapped arm of the paired experiment read twenty-two artifacts with no observed loss, so the number is not a reviewer capacity. Five was selected because the eight-note trial reached it from already-retained quotes in every case, at 2–3% word growth, with no split, no snapshot escape, and no new grounding run. Three would put eighteen notes over.

The operativity path has two channels. commonplace-validate consumes the count deterministically on every validation run, with warn force that does not fail the process. The three hashed criteria consume the reading rules as binding judgment instruction inside review prompts; changing them stales their populations once through ordinary criterion freshness.

Considered alternatives

Keep the sixteen-artifact budget. It is calibrated to a corpus statistic, carries the silent-lag failure mode, and leaves a passing verdict unauditable because redoing it costs what producing it cost.

Price reviewer attention by count and bytes, computed by a sizing command. The proposal Review link budget prices reviewer attention restates the inherited-number problem in a more elaborate form, since nothing identifies the exchange rate between artifact-switching and bytes read. The assay that would calibrate it becomes unnecessary once the artifact fits by construction.

Assemble the evidence pack in code and split the review. The proposal Exceeding a review budget splits the task and the review-attention-price workshop's Mechanism B both partition an over-budget review. Partitioning severs joint support, so a claim carried by several sources together fails in fragments a whole pass would pass, and the routes need partial-coverage and combination semantics the review model does not have.

N = 3. Rejected: it would put eighteen notes over the bound, against eight at five, without a corresponding gain in what a second checker can redo.

WARN indefinitely. Rejected: a bound that only warns is a visible pressure, not a contract, and the trial showed conformance is cheap. WARN was kept only until the eight over-bound notes were conformed, because two of them cited sources only from a trailer or a table cell, where there was no hosting sentence for a quotation.

Per-link (snapshot required) semantics. Rejected because a source could then be both discharged by a quotation at one link and snapshot-required at another, which contradicts the cost the bound counts — artifacts the reviewer must open.

A synthesis-trait exemption for casebook notes. Deferred rather than rejected. The two joint casebooks are the strongest candidates, but both conform numerically, so nothing yet forces the exemption.

Keep the two representation caps. Rejected because they bind on 55% of notes, leaving the majority of the corpus only partly checked and making the note-link exemption's precondition false.

Free choices resolved: the two authoring conventions — that a source cited as evidence is cited in the body rather than only in the trailer, and that a casebook table hosts its retained evidence in an accompanying prose block — live in the notes collection contract. kb/articles/ shares the same N (resolved 2026-08-29): both live articles sat at two and three unquoted sources, so the rule bound on nothing, and the articles contract admits the verbatim marker inside a citation parenthesis so an article has the same discharge route as a note. Nothing is left open.

Consequences

Grounding verdicts become auditable from the note's page: the passages that carry a claim are on it, and a second checker can redo the support judgment without opening the sources. The reviewer caps no longer need recalibration as the corpus grows, and the note-link exemption makes note-dense notes cheap to review rather than expensive.

Authors pay what the review used to absorb. A note over the bound must quote or split, and every quoted passage lengthens the note, so the remedy spends against the co-loading bound. The eight notes over the bound at adoption were conformed from already-retained quotes at 2–10% word growth.

Two risks stay. Occurrence is not representativeness: a validated quotation establishes that the words appear in the source, not that they mean there what the note uses them for, so a passage that reads as lifted from a qualifying context still costs a trip to the source. And an ingest whose retained quotes are paraphrase would discharge the bound falsely; one such ingest has been found and needs a fresh grounding run.

A review-side budget is no longer available as the mechanism that protects reviewer attention. ADR 079 is superseded. The review-attention-price workshop closes; link and usage telemetry remains a measurement of review cost, not a cap.

Where the decision stops applying: it covers kb/notes/ under Commonplace's grounding gate, in a system where verbatim quotations are mechanically validated against their cited source. Another check would have to re-derive its own unit of cost before any count transferred to it. Five is not a capacity claim about other corpora, gates, or models, and the conformance trial that selected it measured eight notes on one date.

Relevant Notes: