Stochastic deduplicated optional editor campaign

Labnote 005: Stochastic deduplicated optional editor campaign🔗

Date: 2026-08-23

Hypothesis🔗

A modest nonzero sampling temperature can produce enough distinct optional editor outcomes to support a useful conditional blind review, provided exact duplicate answer pairs are removed before operator intake and remain visible as multiplicity telemetry rather than independent samples.

Frozen protocol🔗

• Experiment: labnote_005
• Internal execution record retained; identifier withheld.
• Campaign digest: 0927d8781e5e5bc262a72bf52e8d4840c5196f02389f50969608b9269825f002
• Repository revision: 73c9f2e
• Model: qwen3:8b
• Temperature: 0.35
• Base seed: 20260825, varied by repetition
• Matched assignments: 120
• Execution trials: 240
• Protocol revision: stochastic-deduplicated-review-v1
• Accelerator ownership: one non-preemptive accelerator lease

Each optional result is self-matched against its exact pre-editor candidate. Normalized-identical outputs become automatic ties. Normalized unordered changed pairs are content-hashed globally, and exactly one representative from each unique group enters blinded review. Duplicate multiplicity is retained as aggregate telemetry but cannot inflate the manual sample.

Raw prompts, drafts, generated text, edit operations, pair hashes, mappings, and review artifacts remain in owner-only private state.

Runtime and uniqueness result🔗

The campaign completed 240/240 execution trials with no campaign failures, empty stderr, and a normal accelerator-lease release.

The changed-pair multiplicities were sixteen singleton groups, three groups of two, and two groups of three. One voluntary edit normalized to no textual change, leaving 28 changed trials. The 21 unique pairs span seven task/style cells. The private operation aggregate contains one append, two deletes, and 61 replacements.

The public blinded bundle digest is 3628dae66aeb0ed93fe4e7f575f1f6f6b4f3e42a3178534c6d00c6fe75057929.

Blinded human review result🔗

The owner completed all 21 unique locked judgments on 2026-08-23. The reveal gate opened only after the final judgment was durably committed against the original bundle digest.

An exact two-sided binomial test gives p = 1.0. The deduplicated conditional review therefore provides no evidence of a preference advantage for either arm.

Restoring each unique changed pair's observed generation multiplicity gives a descriptive changed-trial frequency of 12 direct preferences and 16 optional preferences. Restoring the 92 automatic ties produces an all-assignment frequency of 12 direct, 16 optional, and 92 ties. These weighted counts describe the generated campaign; duplicate occurrences were not independently judged and must not be treated as independent preference evidence.

Cost🔗

Optional access used about 2.9 times the generated tokens, 3.5 times the prompt tokens, and 2.6 times the latency. It also had a 39.2% protocol-failure rate.

Disposition🔗

• Do not promote or fine-tune optional editor access from this campaign.
• Temperature 0.35 plus content-hash deduplication successfully improved review

diversity from 2 unique pairs in Labnote 004 to 21 in Labnote 005.

• Among genuinely changed unique outputs, human preference was effectively

even; the extra editor pass did not justify its compute cost.

• Preserve self-matching, automatic ties, pre-intake deduplication, and separate

unique versus multiplicity-weighted reporting as infrastructure.

• A future campaign should expand genuinely distinct tasks and target known

residual defects after direct rewrite rather than repeat this same corpus.

• Candidate text, pair-level judgments, hashes, and treatment mappings remain

in owner-only private state and are not reproduced here.