# Research cycle optimization

Date: 2026-09-25. Reviewer: codex, self-review.
Question: can we avoid redundant inference while making the first application test
expose concrete research needs? No new paid live calls or source collection this cycle.

## Execution policy and verification

New conditional-v1 workspaces screen candidate ledgers before continuing. Zero/one
candidate produces an empty relation set deterministically; multiple candidates permit
one relation call which must pass endpoint/evidence-field checks. A deterministic
editorial checklist replaces automatic model self-critique. This prevents the repeated
claim-ID relation failure from propagating and avoids treating self-approval as validation.

Explicit reuse checks completed-run/cache integrity and fingerprints source, prompts,
workflow, adapter settings and code. Matching stages replay with origin provenance;
failed runs and changed inputs cannot silently qualify. Reuse is not freshness checking
of an external runtime profile. Current call/token accounting excludes cached responses.

Tests cover zero/one/two-candidate branches, valid/invalid edges, malformed candidates,
zero-call replay, changed source/prompt/fixture/model, cache corruption and failed-run
rejection. A local mock llm-connect HTTP service verifies execution/reuse accounting.
The checked-in synthetic smoke used two fixture calls; its replay used zero. Relative
to four historical calls, those branches remove two calls or all fresh calls respectively.
This is a call-count result, not a measured percentage reduction in live cost or quality.

Compact records and structured candidate/relation ledgers remain the output contract.
Source locators stay in the evidence ledger/prose. A structured report distinguishes
structural rejections from unassessed editorial corrections. The review-metrics template
requires individual findings before counts can mean anything.

## First application test

[VC-02 walkthrough](../walkthroughs/VC02-OfflineFieldNotes.md) connects Local Commit
Boundary, Pending Synchronization, Conflict Copy and Explicit Reconciliation. These
are four research hypotheses, not canonical patterns. It traces a restart, divergent
edits, preserved conflict, explicit resolution and lost acknowledgement. Removing
concurrent writers removes conflict-specific steps while retaining persistence/retry.

The baseline narrative can already describe the same architecture. The proposed benefit
of the vocabulary is exposing acknowledgement and recovery boundaries for review;
independent evidence of improved decision quality remains unknown. Explicit Reconciliation
may be merged into Conflict Copy if it adds no independent retrieval/application value.

## Editorial corrections made in this cycle

- Replaced one ambiguous saved/synced interpretation with three explicit boundaries:
  local commit, receiver acknowledgement and semantic resolution (VC-02 trace).
- Narrowed restart survival to a stated storage-transaction assumption; no power-loss
  guarantee is inferred from a local database being present (VC-02 assumptions/failures).
- Rejected the implication that a receiver acknowledgement proves all-device convergence
  or backup availability (VC-02 steps 9–10).

These are three recorded design corrections, not model-error-rate measurements. The
synthetic smoke has not received substantive research review and does not inherit them.

## Next questions, ranked by walkthrough impact

1. Select and verify a storage acknowledgement/atomicity contract under interruption.
2. Locate and test receiver deduplication plus revision ancestry under a lost response.
3. Define keyboard/screen-reader conflict discovery and resolution acceptance checks.
4. Contrast VC-01's global validation requirement before generalizing these relations.

T02 owns targeted evidence; T03 now owns the draft walkthrough and connected candidate
set. Runtime proof, canonical admission, VC-01 and the full M1 gate remain unfinished.
No further broad extraction is justified until it answers one of these questions.