Nicholastempleman's picture
Agent: add CSOAI branded card header (csoai/brand-assets)
5d31e65 verified
|
Raw History Blame Contribute Delete
3.78 kB
metadata
license: cc-by-4.0
language:
  - en
task_categories:
  - text-classification
tags:
  - ai-oversight
  - human-in-the-loop
  - review-queue
  - gamification
  - measurement
pretty_name: HITL Academy Practice Review Queue (real PENDING artefacts)
size_categories:
  - n<1K

Council of AI (CSOAI)

HITL Academy — Practice Review Queue (the real one)

Reviewing these IS the real work. These 30 artefacts are actual entries from the Council of AI (CSOAI) Evidence-Admission-Trust (EAT) human review queue — 664 artefacts PENDING at publication time. Your verdicts feed that queue via the fail-closed review machinery (review.mjs + publish-manifest.mjs); this is not a simulation.

What a learner does

For each record: fetch the artefact named by artefact_key (its hash is the store's integrity binding), judge admissibility — does the measurement hold together: schema present, numbers consistent, separation justified, no shortcut (unsigned supersession, collapsed model, saturated bank) waved through — and give a reason. A correct, reasoned verdict, once accepted into the queue by a current human reviewer, earns xp_if_accepted = 50 XP, the same award the engine grants per signed oversight decision.

  • answer_key_status is PENDING — your decision joins the human review queue by design: there is no answer key here, because the gold standard is your judgement joining the real queue, producing an Ed25519-signed csoai.hitl-oversight-receipt/0.1 receipt (operator, axis, decision APPROVED/HOLD/REJECTED, note, competency flag).
  • Prerequisite: complete the frozen Micro-MOOC module for the artefact's axis first — csoai/hitl-academy-orb1 … orb5. The competency gate blocks unbriefed sign-off.

Composition (30 records, sampled deterministically from 664 PENDING)

  • Schemas: meok.bank-pilot/0.1, meok.council-interaction-pack/0.1, meok.measurement/0.1, meok.measurement/0.3, meok.offline-availability/0.1, meok.piece-consolidation/0.1, meok.read-plane/0.1, meok.s1-citation-verification/0.1, meok.takeover-audit/0.1
  • Axis hints:
  • care-refusal: 1
  • continuity: 1
  • estate-vs-board-consistency: 1
  • governance: 22
  • openness: 1
  • s1-citation-verification: 1
  • serving-fidelity: 3
  • Fields are public-safe only: filename key, store hash, schema name, axis hint, size, date.
  • No model outputs, no prompts, no rows files, no signatures beyond the store hash are included.

Sensitivity screening

Every candidate artefact was scanned for private keys, credential-like values (sk-, ghp_, AKIA…, bearer tokens), non-public email addresses and personal filesystem paths. 15 PENDING artefacts were excluded (mostly -hive-attack runs and ecosystem scans carrying home-directory paths or third-party addresses). Remaining 259 clean candidates; 30 sampled. Found something sensitive in an included record? Open an issue on this dataset and it will be pulled and re-screened.

The loop this dataset belongs to

Frozen datasets teach → this queue exercises → accepted verdicts sign receipts → XP and ranks accrue (ORB-I Observer 0 → ORB-V Arbiter 750) → competency-gated humans unblock gates (promotion gate, the sealed 120-item bank's G4 measure gate, working-group seats) → board measurement resumes → new artefacts regenerate the next practice set.

Licence & provenance

  • Licence: CC-BY-4.0 (the dataset records; artefact content remains governed by the CSOAI measurement estate).
  • Source: review/state.json (schema csoai.review-state/0.1), Council of AI, councilof.ai.
  • Measurement, not certification.