| name | subjective-label |
| description | Human-grounded workflow for labeling a large text corpus on a vague subjective trait while jointly refining H/L/N gold labels, seven diagnostic regions, a human-and-machine-readable guideline, sealed-test model scorecards, and production provenance. Use for subjective annotation projects, calibration rounds, guideline optimization, final evaluation, corpus completion, or /subjective-label. |
Subjective Label router
Route the request to the smallest canonical command that can advance the project.
Canonical commands
| command | responsibility |
|---|
/sl-init | initialize corpus, seal final-test ids, embed the development pool, run random Round 1, and close the first checkpoint |
/sl-round | run or resume one later calibration round from candidate retrieval through checkpoint |
/sl-evaluate | freeze G*, create blind human gold on the sealed test, and score registered executors |
/sl-complete | run a validated production policy over the remaining corpus, reconcile outcomes, and perform final audit |
/sl-status | inspect current state, evidence, holds, and next valid action without writing |
Compatibility aliases:
/sl-iterate routes to /sl-round;/sl-validate routes to /sl-evaluate;/sl-scale routes to /sl-complete.
Explain the changed semantics when an alias is used. Do not reproduce the legacy
panel-majority, public-dataset convergence, or k-NN inheritance workflow.
Knowledge commands remain available when installed: lesson, feedback, and digest.
Core contract
Treat one identified human as the semantic authority. Use the strong calibration agent
to elicit, contrast, record, and generalize that person's judgment. Use weak language
models as independent executors whose sealed outputs diagnose guideline clarity; never
treat their consensus as gold.
Maintain these outputs:
- cumulative human-confirmed calibration gold
D_t, frozen as D_cal* at stopping;
- completed corpus
D* only after validated production, reconciliation, and audit;
- versioned structured guideline
G_t and eventually G*;
- seven diagnostic regions H, L, N, HL, LN, HN, and HLN, separate from uncertainty;
- complete Session, checkpoint, test, scorecard, production, and provenance records.
The canonical round is:
closed G_(t-1)
→ candidate pool C_t
→ sealed weak prelabels P_t
→ human batch B_t
→ blind Human-AI Session
→ human final Y*_t + guideline draft
→ checkpoint
→ closed D_t + G_t
Round 1 is random and has no model prelabels or inherited regions. Later rounds combine
targeted challenge items with a stratified random sample of consensus items to detect
shared model error.
Required references
Before routing a mutating command, read the command skill and the references it names.
The authority and lifecycle sources are:
../../ref/ref-contract.md../../ref/ref-stages.md../../ref/ref-assets.md../../ref/ref-architecture.md
Safety and implementation truth
Inspect the actual project state and available implementation before writing. Never
claim a phase ran because its conceptual contract exists. When a required engine,
custodian, schema writer, or verifier is absent, write no substitute evidence; return a
structured HOLD with the missing capability, preserved state, and next implementation
action.
