# Example Material Passport with populated experiment_provenance[] +
# experiment_alignment_results[] + experiment_intake_declaration (#260).
#
# An ML paper that ran TWO experiments externally and brought the results back.
# Demonstrates the full intake + claim->experiment alignment layer:
#
#   - experiment_provenance[]: 2 entries, each with its own nested repro_lock,
#     planned_vs_executed[], negative_results[], known_limitations[].
#   - claim_intent_manifests[]: one manifest carrying THREE claims —
#       * C-001 pure-experiment claim (planned_experiment_ids only),
#       * C-002 mixed-evidence claim (BOTH planned_refs AND
#         planned_experiment_ids — allowed by EP-INV-3, audited by both paths),
#       * C-003 literature-only claim (planned_refs only, no experiment id).
#   - experiment_alignment_results[]: 2 rows, one per experiment-backed claim
#     (C-001 ALIGNED; C-003 is literature-only and gets none), including one
#     OVERSTATED row (C-002 against exp-quantization) so the worst-verdict-wins
#     gate decision is exercised even though the same claim's citation path is
#     clean.
#   - experiment_intake_declaration: status experiments_declared (symmetric to
#     the non-empty provenance array, EP-INV-4).
#
# Shape-validate the provenance entries with:
#   python scripts/check_experiment_provenance.py examples/passport_with_experiment_provenance.yaml
# Cross-array invariant check (convert to JSON first — that lint reads JSON):
#   the test suite (scripts/test_experiment_provenance.py) feeds this example
#   through check_claim_audit_consistency.validate_passport and asserts 0 findings.

origin_skill: academic-paper
origin_mode: full
origin_date: "2026-06-08T09:00:00Z"
verification_status: VERIFIED
version_label: "v1.0"
integrity_pass_date: "2026-06-08T09:30:00Z"
content_hash: "sha256:e1f2a3b4c5d6789012345678901234567890123456789012345678901234abcd"

experiment_intake_declaration:
  status: experiments_declared
  declared_at: "2026-06-08T09:05:00Z"
  declared_by: scholar

experiment_provenance:
  - experiment_id: exp-pruning
    title: "Ablation: remove attention-head pruning from the encoder"
    description: >-
      Measures the contribution of the head-pruning step to downstream macro-F1
      on the held-out test split, holding all other hyperparameters fixed.
    repro_lock:
      schema_version: "1.0"
      stochasticity_declaration: "LLM/training outputs are not byte-reproducible. This lockfile documents configuration, not a deterministic replay guarantee."
      ars_version: "3.11.1"
      model:
        family: custom
        id: encoder-v2-pruned
        weight_stable: true
      prompts:
        hash_timing: skill-load
        skill_md_hash: "sha256:a1a2a3a4a5a6789012345678901234567890123456789012345678901234abcd"
        agents_bundle_hash: "sha256:b1b2b3b4b5b6789012345678901234567890123456789012345678901234abcd"
      materials:
        list_hash: "sha256:c1c2c3c4c5c6789012345678901234567890123456789012345678901234abcd"
        count: 3
      external_protocols:
        s2_api_protocol_version: "3.11"
        s2_snapshot_available: false
      cross_model:
        enabled: false
        secondary_model_id: null
    planned_vs_executed:
      - planned: "macro-F1 on held-out test set with head-pruning removed"
        executed: true
        result_file: "results/exp-pruning/test_metrics.json"
        metric: "macro-F1"
        value: 0.812
      - planned: "per-class F1 breakdown across all 20 classes"
        executed: true
        result_file: "results/exp-pruning/per_class.json"
        metric: "macro-F1"
        value: 0.812
    negative_results:
      - description: "Pruning showed no measurable effect on inference latency (within noise)."
        result_file: "results/exp-pruning/latency.json"
    known_limitations:
      - description: "Single random seed; variance across seeds not characterized."

  - experiment_id: exp-quantization
    title: "8-bit post-training quantization of the deployed model"
    description: >-
      Measures accuracy retention and memory footprint after applying 8-bit
      post-training quantization to the released checkpoint.
    repro_lock:
      schema_version: "1.0"
      stochasticity_declaration: "LLM/training outputs are not byte-reproducible. This lockfile documents configuration, not a deterministic replay guarantee."
      ars_version: "3.11.1"
      model:
        family: custom
        id: encoder-v2-int8
        weight_stable: true
      prompts:
        hash_timing: skill-load
        skill_md_hash: "sha256:d1d2d3d4d5d6789012345678901234567890123456789012345678901234abcd"
        agents_bundle_hash: "sha256:e1e2e3e4e5e6789012345678901234567890123456789012345678901234abcd"
      materials:
        list_hash: "sha256:f1f2f3f4f5f6789012345678901234567890123456789012345678901234abcd"
        count: 2
      external_protocols:
        s2_api_protocol_version: "3.11"
        s2_snapshot_available: false
      cross_model:
        enabled: false
        secondary_model_id: null
    planned_vs_executed:
      - planned: "macro-F1 retention after 8-bit quantization vs fp32 baseline"
        executed: true
        result_file: "results/exp-quantization/accuracy.json"
        metric: "macro-F1"
        value: 0.806
      - planned: "INT4 quantization accuracy sweep"
        executed: false
        skip_reason: "INT4 kernel unavailable on the target hardware; deferred to future work."
    negative_results: []
    known_limitations:
      - description: "Quantization evaluated on one hardware target only (A100); CPU path untested."

claim_intent_manifests:
  - manifest_version: "1.0"
    manifest_id: "M-2026-06-08T09:10:00Z-1a2b"
    emitted_by: draft_writer_agent
    emitted_at: "2026-06-08T09:10:00Z"
    claims:
      # C-001 — pure-experiment claim: backed only by exp-pruning.
      - claim_id: C-001
        claim_text: "Removing attention-head pruning lowers macro-F1 by 4.2 points on the held-out test set."
        intended_evidence_kind: empirical
        planned_refs: []
        planned_experiment_ids:
          - exp-pruning
      # C-002 — mixed-evidence claim: BOTH a literature ref AND an experiment id.
      - claim_id: C-002
        claim_text: "8-bit quantization retains state-of-the-art accuracy while halving the memory footprint."
        intended_evidence_kind: empirical
        planned_refs:
          - dettmers2022int8
        planned_experiment_ids:
          - exp-quantization
      # C-003 — literature-only claim: no experiment id, classic citation path.
      - claim_id: C-003
        claim_text: "Post-training quantization is a widely adopted deployment-time compression technique."
        intended_evidence_kind: empirical
        planned_refs:
          - dettmers2022int8
    manifest_negative_constraints: []

experiment_alignment_results:
  # C-001 against exp-pruning — the claim matches the reported drop. ALIGNED.
  - finding_id: EA-001
    scoped_manifest_id: "M-2026-06-08T09:10:00Z-1a2b"
    claim_id: C-001
    claim_text: "Removing attention-head pruning lowers macro-F1 by 4.2 points on the held-out test set."
    experiment_id: exp-pruning
    result_pointer: "results/exp-pruning/test_metrics.json#macro-F1"
    manuscript_locator: "4. Results > 4.2 Ablations"
    alignment_verdict: ALIGNED
    rationale: "The reported held-out macro-F1 drop matches the 4.2-point claim verbatim."
    judge_model: gpt-5.5-xhigh
    judge_run_at: "2026-06-08T09:20:00Z"
    rule_version: EA-v1
  # C-002 against exp-quantization — OVERSTATED: provenance supports a weaker
  # version. "state-of-the-art" overstates what 0.806 macro-F1 shows. This is the
  # worst-verdict-wins case: C-002's citation path (dettmers2022int8) may be
  # SUPPORTED, but this OVERSTATED experiment row blocks the claim.
  - finding_id: EA-002
    scoped_manifest_id: "M-2026-06-08T09:10:00Z-1a2b"
    claim_id: C-002
    claim_text: "8-bit quantization retains state-of-the-art accuracy while halving the memory footprint."
    experiment_id: exp-quantization
    result_pointer: "results/exp-quantization/accuracy.json#macro-F1"
    manuscript_locator: "4. Results > 4.3 Quantization"
    alignment_verdict: OVERSTATED
    rationale: "Provenance reports 0.806 macro-F1 (a 0.6-point drop from baseline); 'state-of-the-art accuracy' overstates the retained-accuracy result."
    judge_model: gpt-5.5-xhigh
    judge_run_at: "2026-06-08T09:21:00Z"
    rule_version: EA-v1