098 — Separate analysis reconciliation from synthesis

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

Status: accepted Date: 2026-10-01

Amended 2026-10-05: ADR 108 moves record corrections to the declaring analysts and removes the memory return route. The separate reconciliation member, the record loop's limit and the three independent checks remain in force; the correction descriptions below record the prior decision.

Amended 2026-10-06: The synthesis and all three independent verifications are declared retained members, pinned and published as authored. The overview is a code-written entry page with identity and disposition, the synthesis's description, links to every member, the amendment index and deterministic validation account. It copies no member sections. Earlier versions, correction answers and packets, change diffs, round checks and run state remain outside the set and are not published. The separate judgments and bounded loops remain in force.

Context

Record repair and public writing have different dependencies. A public author needs settled records; a reconciliation author must still resolve disagreement and may return findings to the memory analyst. Combining these tasks writes public text before the records that support it have passed independent review. In one local Instinctual Memory run, reconciliation grew from 3,314 to 6,406 bytes across three rounds while synthesis stayed between 4,695 and 4,775 bytes. That observation motivates the split; it does not establish its frequency or net cost.

Decision

Reconciliation is a separate retained member. It authors amendments, supersessions and explicitly marked unresolved conflicts. An independent record judge checks the three analyst members and reconciliation before a separate author writes the public synthesis. A second independent judge checks that synthesis against its supporting records. The overview remains the entry point, with a code-written index of amended records and links to the synthesis and all three judgment members.

Keep the record loop's two correction rounds and the memory return route. Allow one synthesis correction. A record fault found during synthesis checking is a limitation; if it cannot be stated faithfully without making the synthesis misleading, stop rather than reopen reconciliation.

The synthesis author and judge load the synthesis type plus the shared source and record contracts; the judge also loads its verification type. Reconciliation and record verification load the analyst and reconciliation types. Synthesis reads the reports as evidence without loading their authoring contracts.

Considered alternatives

Keep reconciliation in the overview. Delaying synthesis could preserve four members, but would keep a type shared by two authors and require a provisional overview or an extraction boundary. A separate member gives reconciliation one output contract and permits direct record checks.

Reopen records from synthesis checking. This permits deeper repair, but makes the correction budget and the meaning of settled records harder to follow. Bounded limitations or stopping preserve the phase boundary.

Append verification to reconciliation. Readers would find amendments and their judgment together. It gives that member two author roles and makes the overview's completion account less direct. Keep each judgment as a separate member linked from the entry point.

The unresolved-conflict marker and separate record check are execution choices. Cited-record extracts and removing overlap between reconciliation and its judge are deferred until trial evidence supports another change.

Consequences

Record corrections no longer rewrite public text. A complete set retains all declared members, including the synthesis and three verifications, and every synthesis is checked independently. Extra synthesis and verification jobs cost calls even when records need no correction. Separating contracts reduces what public-writing workers must load, while retaining the whole reports as evidence.

The workflow consumes judge blockers as scheduling authority. Validators and publication consume the declared layout and amendment index with binding force. Analysis workers consume their role-specific contracts as authoring instructions. Review rendering, handoff and comparison loaders consume the accepted members and manifest; transfer and landscape workers resolve amended records through reconciliation. Published reviews remain projections of the overview, with the public synthesis and judgments reached through its links.

This decision applies to new frozen-source analysis runs. Structural checks establish shape, identity and reference resolution; both semantic judgments remain model assessments, not an independent experimental oracle. Tests cover both bounded loops and failure before publication. Both fresh Luna trials stopped at final record verification, before synthesis, on source-traceability judgments, an altered quotation and lost reconciliation markers. They exercise the bounded stop path; they do not test the synthesis contract selection or establish reduced total work. Subsequent Sol medium runs reached verified synthesis, and one published on 2026-10-01. Separate Luna medium and Sol medium runs both published on 2026-10-02. These establish the bounded publication path, without a matched estimate of reduced work or improved analytical quality.

This ADR adopts the design proposal Separate analysis reconciliation from synthesis and amends ADR 096's reconciliation location and public-writing role.