Optional editor-access campaign

Labnote 003: Optional editor-access campaign🔗

Date: 2026-08-23

Hypothesis🔗

Giving a model optional access to constrained revision after a matched direct rewrite may improve judged output without forcing a mutation. Finalizing the initial candidate unchanged is a first-class treatment outcome.

Frozen protocol🔗

• Experiment: labnote_003
• Internal execution record retained; identifier withheld.
• Campaign digest: 6728008926b7d896ced277199b03748c3beda1400a0ea1a152a6dd8f879fadd3
• Repository revisions: 9a57f28, corrected by fd4c02f
• Model: qwen3:8b
• Arms: matched direct rewrite and optional editor access
• Planned trials: 48
• Blinded pairs: 24
• Accelerator ownership: one non-preemptive accelerator lease

The optional arm receives the same deterministic initial rewrite as its matched direct arm, followed by one editor decision. It may finalize unchanged or apply bounded revision operations. Invalid editor output is recorded as a protocol failure and transactionally falls back to the unchanged initial candidate, so failure cannot selectively remove a difficult case from human review.

Raw prompts, drafts, generated texts, editor operations, telemetry, review pairs, and the arm-reveal key remain in owner-only private state.

Runtime result🔗

The authoritative campaign completed all 48 trials successfully at the execution layer and released its accelerator lease normally.

All 24 matched comparisons remain in the blinded review set. The public review bundle digest is 534755a72c09fcc375f580bb0afbb36bd6801b8db35a575ea1371c999fdd6f8e.

The original pilot retained only valid optional-editor trials and would have created survivor bias. It is preserved as diagnostic evidence but is not an inferential campaign. The authoritative rerun uses the frozen transactional-fallback-v1 protocol revision.

Blinded human review result🔗

The owner completed all 24 locked judgments on 2026-08-23. The reveal gate opened only after the final judgment was durably committed against the original bundle digest.

Both non-ties preferred optional editor access, but an exact two-sided binomial test on only two non-ties gives p = 0.5. This is not evidence of an arm-level quality advantage. Both preferences occurred among the four voluntarily edited cases; the other two edited cases tied. All eight valid unchanged decisions and all twelve transactional fallbacks tied. No secondary-dimension scores were recorded.

The treatment voluntarily edited 4/12 protocol-valid cases (33.3%), or 4/24 of all treatment assignments (16.7%). Eight valid cases finalized unchanged and twelve invalid editor responses used the frozen unchanged fallback. The valid edits contained six replacements and two deletions in aggregate.

Optional access therefore used about 2.9 times the generated tokens, 3.5 times the prompt tokens, and 2.7 times the latency of direct rewrite in this campaign. Candidate text, operations, pair-level judgments, and treatment assignments remain in owner-only private state and are not reproduced here.

Disposition🔗

• Do not fine-tune or promote optional editor access from this result.
• Treat the overwhelming tie rate as the primary finding: matched deterministic

outputs were usually unchanged or judged equivalent.

• The two optional-editor preferences are localized to one case/style cell and

are too few for a general conclusion.

• Repair the 50% editor-protocol failure rate before another quality campaign.
• If optional editing is revisited, target tasks with a measurable residual

defect after direct rewrite and require the planned secondary scores.

• Preserve zero edits and transactional fallback as first-class outcomes; do

not filter either from future review.