feat: Phase B B3 baseline_caveat field -- forces cheap-explanation visibility in every certificate

docs/evidence/METRICS_CURRENT.md

@@ -9,6 +9,14 @@ Machine-readable snapshot: `outputs/metrics_snapshot.json` regenerated with `mak
 
 ## Changelog
 
+### v0.10.28 — Phase B B3: baseline_caveat field in certificates
+- Added `baseline_caveat` field to `build_certificate()` in `src/openamp_foundry/evidence/certificate.py`
+- Auto-computes three cheap-baseline flags: charge≥4, length 10-40aa, hydrophobic_fraction≥0.30
+- Warns when all three pass: "simple conjunction rule would select this candidate without ML scoring"
+- 63 tests in `tests/test_certificate_baseline_caveat.py` covering presence, charge/length/hydro flags, all-YES/NO paths, claim discipline, integration
+- BASELINE 6331→6394
+- Closes Phase B B3 — cheap-explanation visibility is now mandatory in every certificate
+
 ### v0.10.27 — Phase B B2: CertificateClaimBoundary schema
 - Added `CertificateClaimBoundary` (CCB-) schema
 - 10 fields: ccb_id, pipeline_version, certificate_id, candidate_id, boundary_date, unsupported_claim_classes (8-value vocab), boundary_statement, dry_lab_only, all_listed_classes_unsupported, notes

docs/research/NEXT_100_PR_MAP.md

@@ -43,7 +43,7 @@ Make candidate evidence packages impossible to confuse with biological proof.
 |---:|---|---|---|
 | B1 | Add certificate field for `proof_ladder_level` (complete). | PLC- schema: 14 fields, 14 validation rules, dry_lab_only caps level at multi_signal_candidate_evidence, human_review_required enforced for levels ≥expert_reviewed_assay_proposal, unsupported_claims required as anti-overclaim guard; makes level claim machine-verifiable. | B |
 | B2 | Add `unsupported_claims` field to certificates (complete). | CCB- schema: 10 fields, 13 validation rules, 8 claim class vocabulary, ≥3 classes required, dry_lab_only+all_listed_classes_unsupported enforced, no-duplicate check; negative complement of PLC- closes score-to-proof drift gap. | B |
-| B3 | Add `baseline_caveat` field. | Forces cheap-explanation visibility. | B |
+| B3 | Add `baseline_caveat` field (complete). | Added to build_certificate(): auto-computes charge/length/hydrophobicity flags and warns when all three cheap baselines pass; forces cheap-explanation visibility at certificate issue time. 63 tests in tests/test_certificate_baseline_caveat.py. | B |
 | B4 | Add `release_status` field. | Supports staged release. | C/D |
 | B5 | Add certificate quality-tier validator. | External-review readiness becomes measurable. | B |
 | B6 | Add human-readable certificate report. | Domain experts can inspect faster. | B |

src/openamp_foundry/evidence/certificate.py

@@ -7,14 +7,16 @@
 from openamp_foundry.types import ScoredCandidate
 from openamp_foundry.utils.hashing import stable_json_hash
 
+# Cheap-baseline thresholds for AMP selection heuristics.
+# These define what a simple rule-based filter would pass —
+# a candidate passing all three is explainable without ML.
+_BASELINE_MIN_CHARGE = 4
+_BASELINE_LENGTH_MIN = 10
+_BASELINE_LENGTH_MAX = 40
+_BASELINE_MIN_HYDROPHOBIC_FRACTION = 0.30
 
-def _infer_proof_ladder_level(scored: ScoredCandidate) -> str:
-    """Infer highest proof-ladder level supported by this candidate's evidence.
 
-    Pipeline outputs are multi_signal_candidate_evidence (Level 4) by default:
-    they have feature extraction, multiple independent scorers, novelty checking,
-    and diversity selection. Lower levels apply only if key signals are missing.
-    """
+def _infer_proof_ladder_level(scored: ScoredCandidate) -> str:
     scores = scored.scores
     has_multi_signal = (
         scores.get("activity", 0) > 0
@@ -26,6 +28,54 @@ def _infer_proof_ladder_level(scored: ScoredCandidate) -> str:
     return "multi_signal_candidate_evidence"
 
 
+def _build_baseline_caveat(scored: ScoredCandidate) -> str:
+    features = scored.features
+    charge = float(features.get("net_charge_proxy", 0))
+    length = int(features.get("length", 0))
+    hydro = float(features.get("hydrophobic_fraction", 0.0))
+
+    charge_passes = charge >= _BASELINE_MIN_CHARGE
+    length_passes = _BASELINE_LENGTH_MIN <= length <= _BASELINE_LENGTH_MAX
+    hydro_passes = hydro >= _BASELINE_MIN_HYDROPHOBIC_FRACTION
+
+    flags_yes = sum([charge_passes, length_passes, hydro_passes])
+
+    charge_flag = "YES" if charge_passes else "NO"
+    length_flag = "YES" if length_passes else "NO"
+    hydro_flag = "YES" if hydro_passes else "NO"
+
+    if flags_yes == 3:
+        residual_note = (
+            "All three baseline flags pass — a simple conjunction rule "
+            "(charge\u22654 AND length 10-40 AND hydrophobicity\u22650.30) would select "
+            "this candidate without ML scoring. Residual ML signal must exceed "
+            "this baseline to justify pipeline use."
+        )
+    elif flags_yes == 2:
+        residual_note = (
+            "Two of three baseline flags pass — candidate partially explainable "
+            "by heuristic rules. Verify ML residual is non-trivial."
+        )
+    elif flags_yes == 1:
+        residual_note = (
+            "One of three baseline flags passes — candidate unlikely to be "
+            "selected by simple heuristics alone."
+        )
+    else:
+        residual_note = (
+            "No baseline flags pass — candidate cannot be explained by the "
+            "standard cheap heuristics; ML signal is dominant."
+        )
+
+    return (
+        f"Cheapest-explanation check: "
+        f"charge={charge:.1f} (\u2265{_BASELINE_MIN_CHARGE} passes: {charge_flag}), "
+        f"length={length} ({_BASELINE_LENGTH_MIN}-{_BASELINE_LENGTH_MAX} aa range: {length_flag}), "
+        f"hydrophobic_fraction={hydro:.2f} (\u2265{_BASELINE_MIN_HYDROPHOBIC_FRACTION:.2f} passes: {hydro_flag}). "
+        f"{residual_note}"
+    )
+
+
 def build_certificate(
     scored: ScoredCandidate,
     config: dict[str, Any],
@@ -50,4 +100,5 @@ def build_certificate(
         "pipeline_version": __version__,
         "config_hash": stable_json_hash(config),
         "proof_ladder_level": _infer_proof_ladder_level(scored),
+        "baseline_caveat": _build_baseline_caveat(scored),
     }