YUREI RESEARCHPrompted and constrained editing baseline
An existing instruct model can produce deterministically applicable virtual-buffer operations reliably enough that fine-tuning is not yet justified.
Labnote 001: Prompted and constrained editing baseline
Date: 2026-08-22
Hypothesis
An existing instruct model can produce deterministically applicable virtual-buffer operations reliably enough that fine-tuning is not yet justified.
Frozen protocol
- Experiment:
labnote_001 - Internal execution record retained; identifier withheld.
- Model:
qwen3:8b - Seed:
20260822 - Temperature:
0 - Corpus: six frozen, non-sensitive writing cases
- Arms: append-only, unconstrained JSON editing, and schema-constrained JSON
- Accelerator ownership: one non-preemptive accelerator lease
editing
The committed protocol and corpus are the authoritative inputs. Generated text remains in a private artifact store and is not reproduced here.
Runtime result
The run completed successfully with exit code 0 in approximately 24 seconds. It released its accelerator lease normally, left the queue empty, produced an empty stderr log, and retained its stdout artifact with owner-only permissions.
| Editing arm | Parse success | Semantic application success | Finalize success |
|---|---|---|---|
| Unconstrained JSON | 6/6 (100%) | 5/6 (83.3%) | 5/6 (83.3%) |
| Schema-constrained JSON | 6/6 (100%) | 6/6 (100%) | 6/6 (100%) |
The unconstrained arm failed the friendly-technical case because it did not finalize the document. The schema-constrained arm completed all six cases.
Interpretation
The schema-constrained arm cleared the frozen protocol's 95% mechanical reliability gate. On this evidence, fine-tuning is not justified. Constrained decoding is sufficient to continue testing the interface with the existing model.
This is not evidence that iterative editing improves writing. Both editing arms reported zero edit operations: successful cases primarily supplied a final buffer rather than exercising append, replace, or delete transitions. The experiment therefore established protocol reliability, not composition quality or the value of revision.
The six-case corpus is intentionally small. A single deterministic run cannot establish general reliability, compare semantic quality, or support a model training decision.
Disposition
- Do not fine-tune.
- Preserve this run as the mechanical baseline.
- Keep generated texts private until they receive a blinded quality review.
- In the next experiment, begin from a seeded draft and require at least one
- Compare constrained revision against direct generation using blinded human
- Revisit training only after repeated constrained-protocol failures that
valid revision operation before finalization.
judgments for instruction adherence, correctness, concision, and voice.
cannot be resolved through the interface, schema, or prompt design.