108 — Declaring analysts correct their reports

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

Status: accepted Date: 2026-10-05 Amends: ADR 096 for the rule that analyst members are not rewritten, and ADR 098 for who corrects records and the memory return route. Lens-prefixed identities, declaration ownership, the separate reconciliation member and the three independent checks remain in force.

Context

Folding three analyst reports into one account produces three kinds of statement. A connection relates records in different reports, and no single report owns it. A correction changes a claim one report made, and that report also holds the fields, ledger rows and conclusions that depend on the claim. A standing disagreement is one the evidence cannot settle.

Writing corrections in the reconciliation member made it an overlay. The report kept the superseded text, and the verifier, the profile and synthesis jobs and readers of the published set each had to apply the correction and work out what else it affected. Two of the three runs that entered the record loop on the method of ADR 107 stopped at the third verification for this reason: each verification found report text that still asserted a value an amendment had replaced. Neither run used the memory return, the one path on which a declaring analyst could revise its own text.

The reconciler also judged single reports and corrected what it judged faulty, before an independent verifier judged the same records again. The verifier could not correct; its blockers could only start another reconciliation.

Decision

Each role has one function. The reconciler connects, the record verifier judges, and the declaring analyst corrects.

A record-verification blocker names the one report whose text must change: runtime, memory, epistemic or reconciliation. Code reads that word to schedule work and refuses a blocker without it. For each analyst report named, code runs that analyst again with its previous report and the verification. The analyst returns a complete report and one answer per blocker addressed to it: a correction, or a reason for declining. A runtime correction finishes before memory and epistemic corrections start, because both read the runtime report; they then read the corrected version. A new reconciliation and verification follow each correction step. The loop keeps three verifications, and blockers at the third stop the run.

Every version of a report stays in the run directory. Later jobs and publication receive one current version of each report; the published set carries no predecessor. A corrected report keeps every record ID its predecessor declared.

Reconciliation states identity between declared records, including supersessions in the existing Amendment: <ID> is superseded by <IDs> form, and describes disagreements between reports as unresolved conflicts for the verifier. It does not replace a record's value, judge one report's support or return findings to an analyst.

Operativity: the workflow consumes blocker addressees as scheduling authority. Analyst, reconciliation and verification jobs consume their revised instructions as authoring contracts. Acceptance checks consume the answer count, the retained record IDs and the supersession form with binding force. Profile, synthesis and publication consume the current reports as authoritative text.

Warrant: deterministic checks establish that each addressed blocker has an answer, that declared IDs survive, that an answer claiming a correction comes with changed text, and which versions were verified together. A code-computed text difference shows the verifier which lines changed. None of these establishes that a correction is right or that unchanged dependent text is still supported; that remains the verifier's model-dependent judgment.

Considered alternatives

Keep amendment overlays with stronger instructions. No report is rewritten and no job is added. Every consumer still applies corrections and finds their dependents by reading, which is where the stopped runs failed.

Let the reconciler write corrected reports. One job could correct several reports together without using a round. It gives a worker authority over findings it did not declare and does not re-derive them from the sources. It remains available if analyst jobs prove too costly for small corrections.

Let the reconciler also request corrections. Cross-report disagreements would be repaired a verification earlier. It adds a second request authority and a second correction cycle; the verifier reads the described conflict and addresses it.

Mark changes inside reports with tracked-change notation. It would add a notation with its own escaping and checking rules for information that record IDs and a computed difference already give.

Produce corrected reports only at publication. Public readers would get coherent text, but the record loop would still verify against an overlay.

Declare dependencies between findings so code routes dependents. This would remove the reading step that finds stale dependents. It was left out so that a trial can attribute its outcome to report correction alone.

Consequences

Consumers read reports as written, and the rule to judge records as amended is gone. A correction cycle costs one job per addressed analyst plus a reconciliation and a verification. A dependent passage that neither the verifier nor the analyst notices waits for the next verification, and the third verification still ends the run.

A defect the reconciler would have caught before verification now waits for the verifier. A record cannot be removed by a correction; a wrong record keeps its declaration and states the corrected finding.

Sets published before this decision can carry value amendments in their reconciliation member, and their readers still resolve those. The acceptance rules apply to new runs only.

The decision rests on two stopped runs and scripted workflow tests. It has not been exercised by a model run. Whether three verifications suffice, how often analysts decline, and whether reconcilers stay within the narrower role are open until a trial. The publication policy for unresolved issues is a separate decision and may change what happens to blockers at the limit.