docs: Phase F/A docs bundle -- negative result examples (F6), make docs-check target (A7)

Makefile

@@ -10,6 +10,7 @@ help:
 	@echo "Quick verification:"
 	@echo "  make pr-ready           Pre-PR check (agent-check + doctor)"
 	@echo "  make agent-check        Claim scan + doc links + benchmark deprecation"
+	@echo "  make docs-check         Doc links + doc index coverage (subset of agent-check)"
 	@echo "  make doctor             Environment diagnostic"
 	@echo ""
 	@echo "Pipeline:"
@@ -96,6 +97,10 @@ docs-index-check:
 	PYTHONPATH=src $(PYTHON) scripts/check_docs_index_coverage.py
 	@echo "OK (advisory — index coverage expected to improve over time)"
 
+.PHONY: docs-check
+docs-check: doc-links-check docs-index-check
+	@echo "--- All docs checks passed ---"
+
 docs-only-check:
 	@echo "--- Docs-only PR check ---"
 	PYTHONPATH=src $(PYTHON) scripts/check_docs_only_pr.py

docs/evidence/NEGATIVE_RESULT_EXAMPLES.md

@@ -0,0 +1,139 @@
+# Informative vs Non-Informative Negative Results (F6)
+
+This document provides worked examples contrasting informative and non-informative negative results, to help pipeline users interpret and record negative results correctly.
+
+## The distinction
+
+A negative result is **informative** when it:
+- Was produced by a well-calibrated simulation module
+- Used an explicit, controlled rejection criterion
+- Is traceable to a specific candidate ID and pipeline run
+- Has a rejection reason from the controlled vocabulary
+
+A negative result is **non-informative** when it:
+- Was produced by a module in a degraded or uncalibrated state
+- Used an implicit or ad-hoc criterion ("it looked bad")
+- Is missing a candidate ID or pipeline run reference
+- Cannot be linked to a specific rejection stage
+
+## Example 1: Informative — hemolysis screen rejection
+
+**NRR- record:**
+```json
+{
+  "nrr_id": "NRR-00147",
+  "candidate_id": "TOY-0049",
+  "pipeline_run_id": "RUN-2026-0003",
+  "simulation_module": "membrane_proxy",
+  "rejection_stage": "safety_screen",
+  "rejection_reason": "hemolysis_risk",
+  "confidence": 0.82,
+  "score": 0.87,
+  "threshold": 0.70,
+  "dry_lab_only": true
+}
+```
+
+**Why this is informative:**
+- `rejection_reason` is from `VALID_REJECTION_REASONS` (hemolysis_risk)
+- `confidence` is 0.82 — high enough that calibration can use this signal
+- `simulation_module` is named (membrane_proxy) — interpretable if module is well-calibrated
+- `score` and `threshold` are explicit — the criterion is reproducible
+
+**How to use it:** This NRR- can enter the FMS- failure mode similarity report to check whether future rejections cluster around this pattern.
+
+## Example 2: Non-informative — module in error state
+
+**Situation:** The membrane_proxy adapter returned an error code (408 timeout) during a run. A score of 0.0 was recorded and the candidate was rejected.
+
+**Why this is NOT informative:**
+- The "score" of 0.0 is not a biological prediction — it is a software artifact
+- The rejection was due to module unavailability, not candidate quality
+- If included in calibration, it would falsely inflate the rejection rate
+
+**Correct handling:**
+```json
+{
+  "nrr_id": "NRR-00148",
+  "candidate_id": "TOY-0050",
+  "pipeline_run_id": "RUN-2026-0003",
+  "simulation_module": "membrane_proxy",
+  "rejection_stage": "simulation_error",
+  "rejection_reason": "module_error",
+  "confidence": 0.0,
+  "score": null,
+  "threshold": null,
+  "is_software_error": true,
+  "dry_lab_only": true
+}
+```
+
+Flag with `is_software_error=True`. This marks the record as non-informative for calibration. The SUC- calibration report for membrane_proxy will record the error-state run.
+
+## Example 3: Informative — dual-use flag rejection
+
+**NRR- record:**
+```json
+{
+  "nrr_id": "NRR-00201",
+  "candidate_id": "TOY-0103",
+  "pipeline_run_id": "RUN-2026-0005",
+  "simulation_module": "safety_classifier",
+  "rejection_stage": "dual_use_screen",
+  "rejection_reason": "dual_use_concern",
+  "confidence": 0.91,
+  "score": 0.91,
+  "threshold": 0.50,
+  "dry_lab_only": true
+}
+```
+
+**Why this is informative:**
+- Dual-use concern is a controlled rejection reason from the safety vocabulary
+- Confidence is 0.91 — strong signal
+- The stage is `dual_use_screen` — directly interpretable and auditable
+
+**Important:** This record MUST NOT be omitted from negative result archives. Dual-use safety rejections are never discarded — they are especially important for audit trail integrity.
+
+## Example 4: Non-informative — undocumented ad-hoc rejection
+
+**Situation:** A researcher manually looked at the candidate list and removed 12 candidates that "seemed redundant." No rejection stage, reason, or criterion was recorded.
+
+**Why this is NOT informative:**
+- No `rejection_reason` — cannot enter controlled vocabulary
+- No `score` or `threshold` — criterion is not reproducible
+- No `simulation_module` — cannot be attributed to a calibrated assessor
+- "Seemed redundant" is an implicit judgment, not a machine-checkable criterion
+
+**Correct alternative:**
+Before removing candidates, record a rejection reason using the controlled vocabulary (e.g., `sequence_degeneracy` for redundancy). Run the similarity_neighbor_report to compute explicit novelty scores, then use those scores + a threshold to generate an NRR- for each removed candidate.
+
+## Example 5: Borderline — below-threshold but low-confidence module
+
+**Situation:** A candidate scored 0.72 on ensemble_checker, which has a pass threshold of 0.75. The SUC- report for ensemble_checker shows ECE=0.22 (overconfident, well above OVERCONFIDENCE_THRESHOLD=0.15).
+
+**Assessment:** The rejection reason is valid, but the calibration quality reduces its informational value.
+
+**Correct handling:**
+1. Record the NRR- with `rejection_reason=low_selectivity`, `score=0.72`, `threshold=0.75`
+2. Add a note: `calibration_note: "module SUC-0047 shows overconfidence (ECE=0.22); confidence estimate may be inflated"`
+3. The SUC- record for ensemble_checker should flag this as overconfident
+4. Do NOT omit this from the calibration loop — the rejection is still valid; the confidence just needs downweighting
+
+## Summary table
+
+| Example | Informative? | Key distinguishing factor |
+|---|---|---|
+| 1 — hemolysis rejection | Yes | Explicit criterion, calibrated module |
+| 2 — module timeout | No | Software error, not biological prediction |
+| 3 — dual-use flag | Yes | Safety screen, must not omit |
+| 4 — ad-hoc removal | No | No controlled vocabulary, no score |
+| 5 — low-confidence module | Partial | Valid criterion, reduced weight |
+
+## Agent behavior
+
+When processing simulation outputs, an agent MUST:
+- Record every rejection as an NRR- with a reason from `VALID_REJECTION_REASONS`
+- Set `is_software_error=True` for module error states
+- NOT drop borderline-confidence rejections — record them and add a `calibration_note`
+- NOT rerun a failing candidate without logging the initial failure