Atomic-step grounding bounds unquoted sources per note

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

Archived (see archive README). Adopted by ADR 082: the bound, the validator rule, and the uncapped criteria are the live design. What remains here is the 2026-08-27 current-state anchor and the corpus measurement (344 notes, eight over five sources, median seven distinct artifacts) — design texture only.

Commonplace currently places its grounding reading limit on the review. The semantic/grounding-alignment criterion lets a pass inspect at most sixteen distinct linked artifacts, and sixteen was chosen because it is the corpus p90 offer (ADR 079). Two costs follow from that placement. The number is calibrated to a corpus, so it has to be recalibrated whenever the corpus's evidence needs move. And when it lags, the failure is silent: a reviewer that stops at its limit and returns a pass is indistinguishable, from the outside, from one that read everything and returned a pass. Disclosure conventions report where the budget stopped, not whether a finding was sitting behind it.

The option this proposal describes is the other placement of the same constraint. Bound the artifact instead: a note may cite at most N sources that a reviewer must open and search. A source quoted inline, where the quotation is resolved against the linked file by the validator (ADR 046), does not count against N, because the finding step has moved into code and what is left for the reviewer is a bounded judgment over a passage already on the page. Links to other notes do not count either, because a claim-titled note arrives with its own grounding already discharged.

Both exemptions are load-bearing and neither is asserted here: they are the transferable claims this proposal rests on, cited below.

Current state (as of 2026-08-27)

  • ADR 079 is in force. semantic/grounding-alignment (gate text) budgets sixteen distinct linked artifacts per pass, counts a repeated target once, and treats reaching the limit as disclosure rather than failure.
  • sentence/concept-attribution and sentence/misleading-link-text each read at most five distinct target notes per review. Those numbers are inherited constants, unchanged by ADR 079.
  • commonplace-validate resolves a quotation marked as such against the whole linked file by normalized substring containment (src/commonplace/lib/quote_verification.py, verify_content), pairing each quoted span with the nearest link in its paragraph.
  • Tracked ingests hold retained quotes in a ## Quotes section validated against the pinned snapshot; analysis elsewhere in an ingest is not source support (ADR 073).
  • Corpus measurement over kb/notes/: 344 notes, of which 256 link no ingest at all, 18 link more than three, and 8 link more than five (maximum 9). The eight are single-claim notes: theory-mediated-learning-may-improve-sample-efficiency-under-shifts (9), knowledge-storage-does-not-imply-contextual-activation (8), an-omitted-loop-function-and-a-frozen-one-need-different-repairs (7), evidence/real-self-improving-systems-occupy-combinations-no-rung-captures (7), a-proposal-selection-loop-requires-search-evaluation-and-retention (6), evidence/six-reported-self-improvement-paths-expose-bounded-redesign-surfaces (6), instantiation-alone-cannot-model-agent-learning-across-sessions (6), formal-systems-assess-explanatory-reach-through-causal-and-proof (6).
  • The comparable count over distinct linked artifacts of every kind has a median of seven across 337 measured targets, which is why the source-counting unit and the artifact-counting unit are not interchangeable at any N near five.

What the bound would count

The unit is the unquoted source: an artifact the reviewer must open and search to find the passage that carries the claim. Three consequences fix the count.

  • A source with a validated inline quotation is discharged. The reviewer still judges whether the passage supports the claim, but that judgment reads the note's own page.
  • A linked note is charged to the representation criteria (sentence/concept-attribution, sentence/misleading-link-text) rather than to grounding, so it does not consume N.
  • A source used through the (snapshot required) route has no bounded passage by construction, so nothing about it can be discharged by quoting.

Candidate mechanism

The binding consumer would be commonplace-validate: for each note, count the distinct kb/sources/*.ingest.md targets that carry no accompanying verified inline quotation, and signal when that count exceeds N.

One tightening is required rather than optional. verify_content currently matches a quoted span against the whole linked file, so a span aimed at an ingest can satisfy the check against the ingest's analytical prose. ADR 073 holds that an ingest's analysis is not source support, so for ingest targets the match must be confined to the ## Quotes section. Without that change the exemption can be earned by quoting the wrong half of the ingest.

Forces

For the artifact-side placement. A conforming note fits by construction, so the reviewer's limit never binds on it and the silent-lag failure mode does not arise. The number stops tracking a corpus statistic. And the verdict becomes auditable: a grounding result resting on three sources and their quoted passages can be re-checked from the same page by a second checker, including a person, whereas one resting on fourteen artifacts costs as much to audit as it cost to produce.

Against. The bound puts work on authors that the review currently absorbs, and it lands hardest on exactly the notes that are doing the most: the eight above are single-claim notes whose claim is established by reading six or more sources together. Whether they can be quoted or split without losing the claim is not known, and is the evidence this proposal needs.

The cost the review-side placement avoids. Some claims rest on the structure of a whole argument, on what a source does not say, or on material distributed across a document. Those cannot be carried by a bounded passage. Under an artifact-side bound they are pushed toward the snapshot route or toward fragmentation, and the bound offers them no relief.

Bytes stay unbounded either way. N counts artifacts, and a single (snapshot required) source can be very large. Nothing here prices bytes.

Free choices

  • N. Candidates: five, matching the original grounding cap; three, matching the corpus shoulder where 18 notes sit; or a number chosen as a stated convention. The theory this rests on fixes the kind of number rather than the value — N is a convention chosen so a second checker can redo the step, not a point at which reviewer performance degrades — so the choice is a commitment about auditability, not a measurement.
  • Fail or warn. Failing makes the bound a hard authoring contract; warning makes it a visible pressure that leaves the eight notes publishable while the corpus adapts.
  • (snapshot required) sources. They may count as one each against N, or be disallowed once N is reached, on the grounds that a note needing several whole sources at once is the case the bound exists to catch.
  • Articles. kb/articles/ cites many sources by nature and is not reviewed by the same gate. Articles may get their own N, or be exempt.
  • The representation criteria's own limits. The note-link exemption leaves the five-target limits in sentence/concept-attribution and sentence/misleading-link-text untouched. Whether those also want an artifact-side bound is open and not decided by adopting this one.

Interactions with existing decisions

ADR 079 stands. The review-side budget is not repealed by this option; it becomes a ceiling that never binds on a conforming note. A non-conforming note — one written before adoption, or one exempted — still gets a bounded review with disclosure. The two placements are compatible; what changes is which one normally decides.

The joint-support hazard is avoided rather than solved. Exceeding a review budget splits the task, since archived, identified that partitioning a review by link severs claims whose support spans several sources, so any partition would have to be by claim. Under an artifact-side bound the author, who knows which passages jointly carry the claim, quotes them onto one page and the check reads them together. The partition problem does not arise because nothing is partitioned; the hazard is sidestepped, not answered.

ADR 078 is not contradicted. ADR 078 rejected evidence living in the target artifact and kept the ingest ## Quotes pool as the evidence owner. An inline quotation is not a second evidence store: it is a checked copy of a quote the ingest already retains, admissible exactly because a validator re-derives it and fails on mismatch — the general rule that a derived copy of recomputable truth must be checked or absent. The ingest remains the owner, and the required ## Quotes tightening above is what keeps that true.

Operativity and warrant

The operativity path is the validator. commonplace-validate is the binding consumer, and its force is a build-time check on the note rather than a judgment instruction. cp-skill-write and cp-skill-write-multistage would need to know the rule at write time so a note is authored inside N rather than repaired after it; both already invoke cp-skill-ground at the point where quotes are retained (ADR 078), which is the same place. The grounding gate text itself is unchanged by this option.

What warrants the oracle: the substring test establishes that the words occur in the linked source, and nothing more. It does not establish that the passage means there what the note uses it for, or that it is representative of the source. That semantic half stays with the grounding reviewer, whose limit still applies. The bound therefore reduces what the reviewer must open; it does not reduce what the reviewer must judge.

Adoption criteria

This is the third mechanism candidate for the open review-attention-price question, alongside the reviewer spending a priced budget and code assembling the evidence pack to a pass size. Adopt when that question is decided in this option's favour.

The evidence to gather first is narrow: whether the eight notes above can be quoted or split without loss of the claim each makes. A bounded trial on those eight settles the force that most argues against this option, and it can be run without touching any production criterion.

Relevant Notes: