Seeded mandatory-revision campaign

Labnote 002: Seeded mandatory-revision campaign🔗

Date: 2026-08-22

Hypothesis🔗

Schema-constrained revision can reliably transform a seeded draft through real virtual-buffer mutations and produce outputs suitable for blinded comparison with direct rewriting.

Frozen protocol🔗

• Experiment: labnote_002
• Internal execution record retained; identifier withheld.
• Campaign digest: e1dd20cac2d419017ebea302e47f4ea9ef06ac672c802f075bfa613183efd490
• Repository revision: 152f4a3
• Model: qwen3:8b
• Cases: six reviewed seeded drafts with known defects
• Arms: direct rewrite and schema-constrained mandatory revision
• Prompt styles: concise and preservation-first
• Repetitions: two
• Planned trials: 48
• Accelerator ownership: one non-preemptive accelerator lease

The committed manifest, protocol, and corpus are the authoritative inputs. Raw prompts, drafts, generated texts, checkpoints, review pairs, and the arm-reveal key remain in owner-only private state and are not reproduced here.

Runtime result🔗

The campaign completed all 48 trials successfully at the infrastructure level in approximately 43 seconds. It exited 0, produced an empty stderr log, released its accelerator lease normally, and left the queue idle and unblocked.

The 22 applicable schema-revision trials averaged 2.09 real revision operations. This closes Labnote 001's zero-edit limitation: the model exercised seeded-buffer mutations rather than supplying only a final replacement.

Both protocol failures occurred in the same matrix cell: bounded-safeguards with the preservation_first prompt style. Each produced a parseable operation document but omitted finalization. No retry-level campaign failure or infrastructure stop condition occurred.

Blinded review material🔗

The runner created 22 deterministic A/B pairs and a separate reveal key. The public experiment result contains only the pair count and bundle digest: 8abd30015163779d812487bfbe2d4201c1037e397acab2c3cc5414540a950b30.

Mechanical measurements do not establish which arm writes better. Draft similarity suggests that constrained revision preserved more source material, but whether that preservation was appropriate must be decided through blinded human scoring.

Blinded human review result🔗

The owner completed all 22 locked judgments on 2026-08-23. The reveal gate opened only after the final judgment was durably committed against the original bundle digest.

Among the 19 non-ties, direct rewrite was preferred in 68.4% of comparisons. An exact two-sided binomial test against equal preference gives p = 0.167, so this small review does not establish a statistically decisive arm advantage. No secondary-dimension scores were collected. Candidate text, pair-level judgments, and treatment assignments remain in owner-only private state and are not reproduced here.

Disposition🔗

• Do not fine-tune.
• Treat mandatory seeded revision as mechanically demonstrated, but below the

desired 95% protocol-success threshold.

• Investigate the narrow preservation-first/finalization interaction before

expanding the corpus.

• The sanitized blinded-review workflow and all 22 judgments are complete.
• Do not select schema revision as the default composition protocol from this

campaign. Direct rewrite led numerically, but the sample is not decisive.

• Before a broader confirmation campaign, repair the preservation-first

finalization failure and require the planned secondary-dimension scores so a future review can distinguish overall preference from adherence, correctness, concision, voice preservation, and unintended changes.