SuperAPI AI Directory
Partner with us中文AI newsFavorites
Skillslaohan-jiaozhun

laohan-jiaozhun

Writing & research

View repository

Updated 10-15 22:12

About this skill

Content Calibration Adapter

Contract Status (2026-09-20 governance ruling)

This skill is acalibration adapter auxiliary tool, and is not registered in the executor-registry—the ④ node contract entry point is upstreamcheat-on-content(registry node 4), and the ⑤ deep-scan contract entry point isdbs-content-review+laohan-shencha-content-claims(node 5). The host executes ④⑤ according to registry/whois, and does not route to it merely because this skill directory is visible; this skill's lane routing, anti-contamination, and registration process serve as an operations manual reference.

This is not a second Cheat.cheat-on-content is the only calibration engine; this skill only connects the episode and its calibration lane, avoiding drift in locally forked formulas, prediction rules, or migration rules.

Boundaries that must not be crossed

  • It is forbidden to score here by fixed tutorial-style weights, forbidden to estimate and then treat the score as a prediction, and forbidden to separately write PREDICT / RETRO / BUMP.
  • It is forbidden to take tutorial-video 's historical samples to evaluate opinion-video; the two kinds of content must be independent lanes.
  • Blind prediction must be written by the upstream cheat-on-content into its predictions/before shooting, publishing, or viewing this episode's actual data; the prediction section must not be rewritten.
  • If the state schema does not match after an upstream version update, only cheat-migrate dry-run may be executed first. Historical lanes must not be migrated before backup and user confirmation.

Routing

The verify-blind-rubric.mjs before scoring must not merely verify the header:blind-rubric.md must verbatim preserve the source rubric's current complete dimension definitions and placeholder rules. If the dimension semantics are abbreviated, omitted, or rewritten, keep PENDING, and after repair, re-evaluate in a brand-new isolated context.

  1. First run laohan-bianpai vendors. When Cheat or dbskill upstream has updates, only record them as to-dos; run vendors --synconly before a new episode or during an independent maintenance window. When it reports a schema migration, stop; do not bypass it, and do not switch rules midway through this episode.
  2. Read episodes/<slug>/01-口播稿.md and 00-选题.md, and determine the content form.
    • Tutorials, explicit operation paths, resource delivery → tutorial-video lane。
    • Opinions, commentary, issues, personal judgment → opinion-video lane。
    • When it cannot be determined, stop; it must not default to the tutorial type.
  3. Read the target lane's .cheat-state.json, and confirm that content_form is consistent. Opinion videos always use the workflow project's calibration/opinion-video/; when there is no lane, first create one in that directory using the upstream cheat-init , and do not reuse a mismatched old directory.
  4. First run node scripts/verify-blind-rubric.mjs calibration/<content_form>. For the current 01-口播稿.md , run the upstream cheat-score: the main scorer may use the complete rubric_notes.md to parse the formula, but the isolated sub-scorer may only read the verified blind-rubric.md and the current episode draft. Any contamination warning must be discarded and a brand-new isolated scoring must be restarted. The complete output must retain the original isolated JSON (complete script hash, rubric version, scores/reasons for each dimension, input_status, self_check, refusal); the registration script will re-check it, and merely registering the Markdown's SHA is not allowed. Save the warning-free output into the lane, then run node scripts/register-cheat-score-evidence.mjs ...。
    • When the lane does not yet have real published samples (calibration_samples=0), the status must be labeled as the enum value COLD_START_DIAGNOSTIC, not the plain-language "cold-start diagnosis"; the register/prepare scripts and the bianpai gate mechanically judge by that enum value, must not claim quality or traffic has been verified, and must not overwrite the original template, catchphrases, scenes, rhythm, or punchlines in order to improve composite. A scoring suggestion enters the revision candidate list only when it points out a specific comprehension/delivery problem.
  5. Enter ⑤: fixed calls to dbs-script-flow、dbs-resonate、dbs-ai-check and laohan-shencha CONTENT_CLAIMS; call only when the first 5 seconds are judged weak: dbs-hook. dbs-ai-check only diagnoses the specific hit sentences and does not automatically rewrite; once dbs-hook triggers, fully use the three upstream methods, 3–5 entries each, for a total of 10–15 differentiated candidates. Fabricating author experience, number of interviews, data, results, or cases is strictly prohibited; when evidence is lacking, you must write [需真实证据]. After the report is complete, run node scripts/stamp-episode-script-hash.mjs episodes/<slug> 04-深扫报告.md 04-事实核验.md, and only then will the two ⑤ reports be written to the same script_hash. Hash matching alone is insufficient: ③ must have no explicit unresolved high risk; ⑤ the deep scan must be review_status: CLEAR and unresolved_issue_count: 0, and the fact verification must be fact_check_status: CLEAR、contradicted_count: 0、unverifiable_count: 0, otherwise prepare or registration of the final blind prediction is not allowed. ⑤ must also output 04-事实主张.json: any PROOF beat must cite the SUPPORTED claim_id bound to the current draft/fact report and same-source evidence;INFERRED(even with inference_note) must not enter a PROOF beat; a source link by itself does not constitute factual proof. Note: the current stamp script only binds the hashes of the two ⑤ reports; the PROOF beat boundary of04-事实主张.json is a descriptive-layer constraint, and this time the script is not changed.
  6. After the ⑤ reports match the current draft hash, first run node scripts/prepare-cheat-prediction-input.mjs episodes/<slug> calibration/<content_form>. Use the output lane snapshot as the sole input for the upstream cheat-predict , and write **Script**: <snapshot path> and **Script Hash**: <sha256>in the prediction metadata. Then run node scripts/register-cheat-prediction.mjs episodes/<slug> calibration/<content_form> predictions/<file>.md v1. It verifies that the prediction is later than the snapshot and that the metadata matches the snapshot, and only then writes 03-预测证据.json and marks RECORDED. Any revision invalidates the hash; you must rerun ⑤, prepare a new snapshot, and register a new prediction revision; you cannot rewrite an old prediction.
  7. 03-校准报告.md must begin with:
---
calibration_engine: cheat-on-content
content_form: opinion-video # 或 tutorial-video
calibration_root: <相对于本报告的 lane 路径>
script_hash: <01-口播稿.md 的 sha256>
score_status: COMPLETE
prediction_status: PENDING # ⑤完成后才改为 RECORDED
prediction_file: null # RECORDED 时必须是 predictions/...md
---

Then only summarize this issue's score, key revision points, and the upstream prediction file path; do not copy a full rubric. Run laohan-bianpai status; missing upstream state, hash, or final prediction means BLOCKED.

Historical compatibility

~/.openclaw/workspace-shared/content-ops is the old tutorial-video calibration data, not the default source for all episodes. It remains read-only until it is upgraded to the upstream schema; opinion videos such as FDE must not use its old tutorial-style composite as a basis for release.

When to return here

  • After publication:laohan-yunying only archives the real publication/data/comments, then hands them to the upstream cheat-publish and cheat-retro。
  • When samples are sufficient: full re-scoring and independent review are performed only by the upstream cheat-bump .
  • ⑭ retro-handoff must have feedback_hypotheses: each entry bound to a unique hypothesis_id, intervention, expected metric, T+N window, status, and feedback target (① or ④); if fields are missing, bianpai must reject the full workflow. This field is produced by the upstream cheat-retro ; this skill serves as the routing entry from ④ to ⑭, and after registering the blind prediction it must prompt that ⑭ needs this field completed (a single episode may only OBSERVATION/HYPOTHESIS, and only with at least two episodes of the same intervention, a fixed window, and an explicit control may it be marked UPGRADE_EVIDENCE）。