TR-2026-0028技术报告未经同行评审实验性已发布

Skill Factory: acquisition where the judge, not the teacher, is authority

Ran Tao (Octoryn Research)

本文未经同行评审。请将其视为工作文档,而非经验证的结果。

摘要

Skill acquisition gated by a model-free judge, not any teacher. On a small correctness probe, a model-free pattern-miner qualified all of its trajectories while a judged large teacher qualified none of its own: the teacher's structure was sound, but its output discipline failed the judge. Teachers only propose on the residual the miner cannot cover, and every proposal must clear a deterministic format-normalizer plus the same un-bypassable judge. Small-model distillation qualified none. Authority is the judge; this is a smoke test, and qualification gates output discipline, not skill quality.

本 Research Object 以其原始语言(英文)发表。

1. Contribution

Claimed. A sovereign skill-acquisition pipeline in which authority rests with a model-free judge, not with any teacher model. A model-free pattern-miner is the producer of record and, on a small probe, qualified all of its trajectories while a large teacher qualified none of its own. The teacher only PROPOSES on trajectories the miner misses; each proposal must pass a deterministic format-normalizer (JSON-repair plus schema-aware numeric-unquote) and then the SAME model-free judge before becoming a skill. The gate is shown to be un-bypassable on the probe: a malformed tool is rejected at the first gate and a leaky parameter at the second. Ordering observed on the probe: hybrid (miner plus judge-gated teacher) > miner-only > teacher-only; small-model distillation qualified none of its trajectories.

NOT claimed. This is a correctness/discipline-gating result on a tiny probe, not a benchmark-scale evaluation and not an algorithmic novelty claim. We do NOT claim the produced skills are useful or high-quality in deployment (the judge gates format/discipline, not task value); we do NOT claim the normalizer repairs wrong reasoning or structure (format only); we do NOT claim the large teacher is generally inferior (its structure was sound — only output discipline failed the strict judge); we do NOT claim acquisition works on a live fleet of real recurring tasks at scale; and we make no throughput claim beyond the qualitative "microsecond-scale" note for the miner.

2. Method

Field	Value
Pipeline	trajectory -> model-free pattern-miner (default producer) -> (miner-miss only) teacher PROPOSES -> format-normalizer -> same model-free judge -> accept iff qualifies
Producer of record	model-free pattern-miner (zero model)
Teacher role	proposes only, on the residual the miner cannot mine
Normalizer	deterministic JSON-repair plus schema-aware numeric-unquote (format only)
Judge	model-free, same judge applied to both miner and teacher outputs
Teacher model	a large capable model (identity/version not recorded)
Distillation model	a small model (identity/version not recorded)
Stimulus / inputs	trajectories (count/source not recorded beyond the probe denominators below)
Hardware	consumer-class workstation (not characterized in detail)
Stack / versions	not recorded
Precision	not recorded
Metric + gate	qualifies-as-skill (binary) under the model-free judge; un-bypassability checked via injected bad-tool / leaky-param cases

3. Results

Producer	Qualified	Note
Model-free pattern-miner	all trajectories	producer of record; "microsecond-scale, zero model"
Large teacher (raw, judged)	none	structure sound; output discipline (quoted numerics, malformed JSON) failed the strict judge
Small-model distillation	none	distillation not viable behind the judge

Gate behaviour on the probe: a bad tool is rejected at the first gate (normalizer) and a leaky parameter at the second (judge) — the judge is un-bypassable. Qualitative ordering: hybrid > miner-only > teacher-only.

4. Ablations / Negative controls

Control	Result	What it shows
Teacher-only (no miner), judged	none qualified	a capable teacher alone does not clear the discipline gate
Small-model distillation	none qualified	distillation does not substitute for the miner
Injected malformed tool	rejected at gate 1 (normalizer)	first gate catches structural format faults
Injected leaky parameter	rejected at gate 2 (judge)	second gate catches semantic leakage

5. Threats to validity / Limitations

Sample size is a smoke test, not a benchmark. Denominators are a handful of trajectories (miner / teacher) and a smaller set for distillation. The "all / none / none" results are directional, not statistically powered; a single different trajectory could move them.
Single-run, no variance. No repetition counts, seeds, or confidence intervals are recorded; the performance descriptor ("microsecond-scale") is qualitative with no measured number.
Parity/discipline != quality. The judge gates format and output discipline, not whether a produced skill is correct or useful in deployment. Qualifying every probe trajectory does not establish that those skills perform well.
Format-only normalizer. The normalizer repairs JSON and unquotes numerics; it cannot fix wrong reasoning or structure, which the judge then rejects. This bounds the teacher path to proposals that are already structurally sound.
Teacher result is discipline-specific, not capability-general. The large teacher's failure reflects output-discipline faults, not a structural deficiency; this should not be read as the teacher being a poor reasoner.
Unrecorded environment. Hardware, stack/versions, precision, teacher/distill model identities, and trajectory provenance are not recorded, limiting independent reproduction.
Scope bound. Acquisition from a live fleet of real recurring tasks at scale is untested; results are from a controlled probe.
Evidence level. Retained as "experimental"; given the probe sample size, treat the numbers as a correctness smoke test rather than experimental-at-scale.

6. Reproducibility / Provenance

Build: standard release build of the implementation; exact build identifiers, run commands, and version pins are not recorded.
Sovereignty: no third-party ML framework in the linear-algebra path (platform-native acceleration only).
Related context: a companion "Skill Runtime" effort covers activation — this note covers acquisition, not activation.
Validation caveat: the judge is the authority; any reproduction must apply the SAME model-free judge to BOTH miner and teacher outputs, and must re-confirm un-bypassability with injected bad-tool / leaky-param cases before trusting qualified counts.

声明边界

作者对范围的明确界定——本工作证明了什么、未证明什么——沿用自 Octoryn Research 的发表模型。

证明

On a small probe, acquisition is sovereign: a model-free pattern-miner is the producer of record, a judged large teacher qualified none on output discipline, and any teacher proposal must clear a format-normalizer plus the same model-free judge, un-bypassable on the probe.
On the probe, the hybrid configuration (miner plus judge-gated teacher) ordered above miner-only and above teacher-only.

未证明

That the normalizer fixes wrong reasoning or structure; it repairs format only.
That a small model can distill skills; small-model distillation qualified none on the probe.
Acquisition from a live fleet of real recurring tasks at scale.
That qualified skills are useful or high-quality in deployment; the judge gates discipline, not task value.
Any statistically powered claim: denominators are a handful of trajectories, single-run, with no variance reported.

适用于

Acquiring skills from trajectories where a model-free miner covers the bulk of cases and a capable teacher covers only the residual, behind the same model-free judge.

不适用于

Treating any teacher as authoritative, or relying on small-model distillation.
Extrapolating the probe counts to benchmark scale or live-fleet deployment, or reading qualification as proof of skill quality.

作者

Ran Tao — 调查研究, 写作

引用本文

引用格式

Tao, R., Octoryn Research. (2026). Skill Factory: acquisition where the judge, not the teacher, is authority (TR-2026-0028). Octopus Research Institute.

BibTeX

@techreport{oritr20260028,
  title       = {Skill Factory: acquisition where the judge, not the teacher, is authority},
  author      = {Tao, Ran and {Octoryn Research}},
  institution = {Octopus Research Institute},
  year        = {2026},
  note        = {Permanent ID TR-2026-0028. Not peer reviewed.}
}

披露

资助: 硬件与基础设施由 Octoryn / Octopus Core Pty Ltd 提供。
利益冲突: Octoryn 提供商业推理与治理工具;相关发现独立报告。