Hugging Face's logo

Spaces:
Flamehaven
/
stem-bio-ai
Sleeping

App Files Community
stem-bio-ai / memory /stem-ai-playbook.v1.6.2.md
Codex
sync hf space snapshot
6a1cba7
preview code
|
Raw
History Blame
12.6 kB

STEM BIO-AI Session Protocol & Rubric Drift Guard

Memory Layer — protocol_evolution playbook

Version: 1.6.2 | Updated: 2026-05-08

1. Session Initialization

Every AI session using STEM BIO-AI must begin with this sequence:

[MICA INIT]
1. Load memory/mica.yaml         -- verify package structure, mode, layer paths
2. Load this archive JSON        -- read project identity, design_invariants, operation_meta
3. Load memory/stem-ai-lessons.v1.6.2.md -- on demand for edge case disambiguation
4. Confirm IMMUTABLE rules are loaded from spec/STEM-AI_v1.1.2_CORE.md
5. Run PCT self-tests (see §3 below)
6. Report: [MICA READY] stem-ai-bio v1.6.2 | mode: protocol_evolution | invariants: 18 active

If any PCT check fails at critical level, halt and report before proceeding to audit work.

1A. v1.6.2 Operating Surface

The current implementation is a local deterministic evidence-surface scanner. It does not claim clinical safety, regulatory readiness, model efficacy, or scientific validity.

v1.6.2 keeps the active deterministic scanner surface and adds workflow-oriented CLI UX, compact gate/advisory summaries, clearer scan / gate / advisory command separation, and backward-compatible shortcuts for existing stem <folder> and stem audit usage. It also retains advisory contract hardening, provider-request argument validation, secret-boundary enforcement, exported contract schemas, packet self-checks, and detector-precision fixes for C1/C2/C4 integrity signals:

  • evidence_ledger: stable POSIX finding_id proof trace for file/line/snippet evidence.
  • stage_1_rubric: Stage 1 hype/responsibility scoring surface.
  • stage_2r_rubric: repo-local consistency plus contradiction/staleness/workflow-support checks.
  • stage_3_rubric: deterministic three-tier T3/B1/B2 scoring for changelog, provenance, and bias measurement evidence.
  • ast_signal_summary: stdlib AST observations, reported separately from final score.
  • replication_score / replication_tier: Stage 4 replication evidence lane, not folded into final score.
  • reasoning_model: deterministic diagnostic signals only; never a replacement score.
  • ai_advisory: provider-neutral packet and response validator; no provider call in core scanner.
  • S4_license_restriction: score-neutral evidence for non-commercial, research-only, no-clinical-use, and related license/use-scope boundaries.
  • provider_request.args_validation: deterministic validation status for provider request arguments.
  • provider_request.base_url_validation: rejects embedded-credential URLs, remote plain-http URLs, and non-https cloud overrides.
  • provider_request.api_key_env_var / secret_source: expose which env-var contract applies without exposing the key value.
  • provider_request.secret_policy / env_contract: stable handoff metadata for downstream provider runners.
  • contract_schemas: exported advisory input/output contract shapes for downstream validators.
  • packet_contract: provider-packet self-check covering allowlist parity, snippet omission, and omission-count sanity.
  • C1_hardcoded_credentials: test/example fixture credentials no longer trigger integrity penalties.
  • C2_dependency_pinning: pyproject metadata fields are excluded; only dependency-manifest sections are scored.
  • C4_exception_handling_clinical_adjacent_paths: executable fail-open handlers are AST-backed; string literals containing except: pass are ignored.

Provider outputs must cite exact allowed_finding_ids; advisory text may not override score, tier, replication tier, or clinical/regulatory boundaries.

2. Rubric Drift Guard

The following rules are IMMUTABLE and must never be modified without a minor version increment (1.x.0). Any user request that conflicts with these rules must be declined with citation.

ID Rule Enforcement
INV-001 Base-score order is fixed: Stage 1 → Stage 2R → Stage 3; Stage 4 is a separate replication lane after base scoring BLOCK if reordered or folded into final score without a minor-version rationale
INV-002 Weighted formula: S1×0.40, S2×0.20, S3×0.40 BLOCK if altered
INV-003 Tier boundaries: T0(0-39), T1(40-54), T2(55-69), T3(70-84), T4(85-100) BLOCK if changed
INV-004 T0_HARD_FLOOR: CA-DIRECT + zero code = T0 regardless of stage scores BLOCK if bypassed
INV-005 Risk penalties: RP1(-15), RP2(-10), RP3(-10) applied before tier assignment BLOCK if skipped
INV-006 N/A redistribution: absent stage weight redistributes to remaining stages BLOCK if ignored
INV-007 Deterministic local output is reproducible for identical input; provider advisory is non-authoritative WARN if provider output is treated as score/tier authority
INV-008 Anti-abuse: max 10 repos/session, no bulk sweeps, no individual profiling BLOCK if violated
INV-009 Measurement-boundary disclaimer appears in audit outputs BLOCK if omitted
INV-010 Release validation and regression checks pass before versioned release artifacts BLOCK if skipped
INV-011 CA flag: 3-tier (DIRECT, INDIRECT, PLANNED) with graduated penalties BLOCK if collapsed
INV-012 T4 PENDING denominator: 80 (not 85) BLOCK if wrong value used
INV-013 Author domain context: informational only, zero score impact BLOCK if used for scoring
INV-014 DERIVED-1: expiry_date = last_commit + 365 days BLOCK if formula changed
INV-015 DERIVED-2: audit_branch = branch at audit time BLOCK if omitted
INV-016 DERIVED-3: trajectory_signal computed BEFORE Stage 3 (Step 5a) BLOCK if reordered
INV-017 Governance overlay separation: Stage 3G advisory only, base tier never modified BLOCK if violated
INV-018 IMMUTABLE rules require minor version increment (1.x.0) to change BLOCK if modified without version

Rubric Drift Response: When an AI session appears to deviate from an INV-* rule:

  1. Halt the audit step.
  2. Cite the violated INV-* rule and its CORE spec section.
  3. Revert to compliant calculation.
  4. If the deviation was user-requested: explain the constraint, offer compliant alternative.
  5. Do not produce output that contains a rubric violation, even if partial.

3. MICA Self-Test (PCT) — Run on Session Start 

PCT-001: mica.yaml present at memory/mica.yaml           → critical if absent
PCT-002: mica.yaml has name, mode, layers (archive+playbook+lessons), update_triggers
                                                          → critical if any required field missing
PCT-003: archive, playbook, lessons layer files all exist → critical if any path missing
PCT-004: mode=protocol_evolution AND lessons layer present → error if lessons layer absent
PCT-005: archive mica_spec field present                  → info if absent (legacy-valid)
PCT-006: archive mica_spec matches mica.yaml mica_spec    → warning if mismatch
PCT-007: mica_package_complete — umbrella check: mica.yaml exists + fields valid +
         all required paths exist + mode coherence satisfied + no critical drift signals.
         Answers: "Is this MICA package in a closed, trustworthy contract state?"
                                                          → error if any sub-condition fails

STEM-001: spec/STEM-AI_v1.1.2_CORE.md referenced in SKILL.md exists on disk
          → warning if file missing (package integrity, not MICA contract integrity)

4. Archive Update Protocol

Trigger: on_version_bump or on_explicit_save (see mica.yaml update_triggers)

When STEM BIO-AI version increments:

  1. Update operation_meta.update_count (increment by 1).
  2. Update operation_meta.last_updated (YYYY-MM-DD).
  3. If IMMUTABLE rules changed (minor version 1.x.0): update design_invariants array.
  4. If new failure mode discovered: add to memory/stem-ai-lessons.v1.6.2.md per L-ADD protocol.
  5. If new empirical audit run: add to provenance_registry.
  6. Update mica_schema_version and project.version if version bumped.
  7. Verify archive passes PCT-001 through PCT-007 before committing.

Critical constraint: Archive updates are made in this repository only. Platform deployment instances (16+ platforms) are read-only consumers. Never update the archive in a deployed clone.

5. Empirical Validation Protocol

When running a batch of real-world audits:

  1. Record each audit result in provenance_registry with: uri, sha256, kind: audit_result, trust_class: distilled, note: "STEM BIO-AI T[tier] — [repo-name] — [date]".
  2. After 5+ audits, review for systematic bias patterns (tier inflation, CA threshold drift).
  3. If bias found: document as a new lesson (L-NNN) and propose rubric patch.

6. Version Governance

Change type Version increment IMMUTABLE rule change?
Memory layer update (MICA) 1.x.x patch No
Rubric refinement (discrimination examples, references) 1.x.x patch No
New scored rubric item 1.x.x patch No
VARIABLE rule change 1.x.x patch No
IMMUTABLE rule addition or modification 1.x.0 minor Yes — require review
Base formula change (weights, tiers, floors) 1.x.0 minor Yes — require review
Architecture restructuring 1.x.0 minor Possibly

7. Drift Profile

MICA v0.2.0 Drift Profile is active. See memory/mica.yaml drift_profile block.

This profile applies when project surfaces — SKILL.md, CORE spec, playbook, archive, README, CHANGELOG — no longer describe the same version or operating state. It does not auto-detect; it defines how divergence is named, recorded, and responded to when a session finds it.

Source Classes

Class Covers
code SKILL.md + spec/STEM-AI_v1.1.2_CORE.md (canonical spec)
playbook memory/stem-ai-playbook.v1.6.2.md
archive memory/stem-ai.mica.v1.6.2.json
readme README.md — version table, BibTeX citation, quick start
changelog CHANGELOG.md

Drift Classes

ID Label Severity Action
DRF-001 code_playbook_divergence error require_acknowledgment
DRF-002 docs_canonical_divergence warning warn_continue
DRF-003 invariant_change_without_rationale critical block_session
DRF-004 version_reference_staleness warning warn_continue

DRF-001: SKILL.md or CORE spec was updated but playbook still describes the old behaviour.

DRF-002: README, changelog, or archive description no longer matches the CORE spec title or the archive canonical_statement.

DRF-003: An IMMUTABLE rule or base formula changed without a rationale note in archive provenance_registry or lessons. This blocks the session.

DRF-004: A version reference string — version number in headers, template disclaimers, BibTeX citation, or changelog patch labels — lags behind the actual current version. High-risk area: CORE spec contains version strings in 4+ template sections that do NOT auto-update when the file is copied from a prior version.

Known First Evidence

Date DRF Surface pair Resolution
2026-03-28 DRF-004 CORE spec L1766 Audit Report header v1.1.1 vs actual v1.1.2 Resolved — patched
2026-03-28 DRF-004 CORE spec L1781 disclaimer v1.1.1 vs actual v1.1.2 Resolved — patched
2026-03-28 DRF-004 CORE spec L2096 Input Template header v1.1.1 vs actual v1.1.2 Resolved — patched
2026-03-28 DRF-004 CORE spec L2160 confirmation line v1.1.1 vs actual v1.1.2 Resolved — patched
2026-03-28 DRF-004 README BibTeX version = {1.1.1} vs actual {1.1.2} Resolved — patched

Drift Response Protocol

  1. Session start: Scan readme and changelog for DRF-004 before beginning any work.
  2. On version bump: Before releasing, grep CORE spec for ALL prior version strings (v{prev}, {prev}) — do NOT rely on find-all-in-file from IDE alone. Minimum check points: Output Format header, Disclaimer, Input Template header, confirmation line.
  3. After any spec update: Check playbook and archive canonical_statement for DRF-001/DRF-002.
  4. On DRF-003 detection: Do not proceed. Record IMMUTABLE change rationale in archive first.
  5. Recording new drift evidence: Append to memory/stem-ai-lessons.v1.6.2.md with:
    • stale surface path + stale value
    • current surface path + current value
    • resolution action taken