fix target safety prescreen decision and scope review blockers

docs/handoff/2026-08-04-target-safety-therapeutic-window-prescreen.zh-CN.md

@@ -0,0 +1,69 @@
+# Target Safety and Therapeutic-Window Pre-screen Handoff
+
+## Status
+
+- Branch: `task_20260804_target-safety-prescreen-fix`
+- Base: latest `origin/main` at task start
+- Review: Round 1 returned `REQUEST_CHANGES`; remediation is in clean replacement PR #56
+- Data boundary: no source data, cache, result, model weight, or runtime output in the repository
+
+## Scope
+
+This GenModule is a target-level public-evidence pre-screen for ADC development.
+It asks whether public evidence contains a target-intrinsic hazard strong enough
+to kill, hold, or downgrade investment before antibody discovery and ADC assembly.
+It does not claim product-specific therapeutic-window prediction.
+
+## Implemented
+
+- Six evidence axes: normal tissue expression, surface accessibility, antigen density, soluble antigen/shedding/sink, existing modality toxicity, and tissue consequence/recoverability.
+- Evidence levels `A/B/C/D/U` and explicit risk directions.
+- Fatal-first rules for critical surface hazard, confirmed severe on-target toxicity, non-lower normal density, clinically demonstrated sink/exposure failure, and no exploitable differential.
+- Non-fatal material risk is explicitly retained and produces `HOLD`; `GO` requires all six axes to be resolved with no material risk or conflict.
+- Unknown direction auto-propagates to unresolved; surface, tissue criticality, and density evidence are aggregated across claims; differential status is structured rather than tag-driven.
+- Decision semantics: `KILL`, `HOLD`, `CONDITIONAL_GO`, `GO`.
+- Unknown, unresolved, and conflicting claims remain visible and produce next-experiment references.
+- All cross-boundary identities and evidence references require `external:` references.
+- `claim_ref` and `evidence_refs` are unique and must match exactly; mitigation
+  references must point to a `SUPPORTS_RISK` claim in the same request.
+- Runtime location is declared as `${BIOWORKSPACE_ROOT}/DATA/target_safety_therapeutic_window_prescreen/{raw,processed,result}`; no runtime writer is enabled in the repository.
+
+## Validation
+
+- Module tests pass.
+- Module tests: 21 pass; full suite: 228 tests pass.
+- `scripts/verify_repository_boundary.sh` passes.
+- `git diff --check` passes.
+- No `__pycache__` directory remains.
+
+## Known limitations
+
+- Evidence retrieval, source normalization, citation resolution, scoring calibration, and persistence remain external runtime responsibilities.
+- The first ruleset is deterministic and conservative; it is not a clinical safety model and must not be used as a product-level therapeutic-window claim.
+- The next implementation phase should add an external runtime adapter and benchmark fixtures under `DATA`, only after this contract PR is reviewed.
+
+## Round 1 remediation
+
+- Rebuilt the PR from `origin/main` so the CRC clinical-frame commit is not in scope.
+- Bumped the module and contract version to `0.3.0` for the structured differential, material-risk output, and hazard-context fields.
+- Replacement PR: `https://github.com/leezx/StelligenOS/pull/56`, head `d36e4a4`.
+
+## Round 2 remediation
+
+- Fatal aggregation is now context-aligned by shared `hazard_context_ref` or `(tissue, cell_type)`; unscoped claims cannot combine into a fatal decision.
+- `CONDITIONAL_GO` now requires every material-risk claim to be covered by a context-matched differential or explicit `mitigates_claim_refs`.
+- `NO_EXPLOITABLE_DIFFERENTIAL` requires an external comprehensive assessment reference; a single observation is downgraded to unresolved.
+- Added cross-context, unrelated-differential, partial-coverage, and unscoped-fatal regression tests.
+
+## Round 3 remediation
+
+- Read ChatGPT Round 3 `REQUEST_CHANGES`; the remaining blocker was that duplicate
+  claim/evidence references and invalid mitigation references could corrupt risk
+  coverage and permit a false `CONDITIONAL_GO`.
+- Tightened `AssessmentRequest` to require unique claim references, unique
+  evidence references, exact equality between the two reference sets, and
+  mitigation references that resolve to in-request `SUPPORTS_RISK` claims.
+- Added five contract-integrity regression tests covering duplicate claims,
+  duplicate evidence, missing mitigation targets, non-risk mitigation targets,
+  and the duplicate-reference conditional-go path.
+- Bumped module and contract versions from `0.3.0` to `0.4.0`.

genmodules/README.md

@@ -21,6 +21,10 @@ are not lifecycle stages or Gate implementations.
 - `gen_indication_endpoint_target@0.1.0`: defines data-free contracts for
   constrained ADC indication, endpoint, and target opportunity generation;
   generation, evaluation, ranking, and evidence remain external.
+- `target_safety_therapeutic_window_prescreen@0.4.0`: applies conservative,
+  fatal-first rules to externally supplied public-evidence claims for target-
+  intrinsic ADC safety pre-screening; it does not predict a product-specific
+  therapeutic window.
 
 ## Repository boundary
 

genmodules/target_safety_therapeutic_window_prescreen/README.md

@@ -0,0 +1,57 @@
+# Public-Evidence Target Safety and Therapeutic-Window Pre-screen Engine
+
+This GenModule performs a conservative, target-level ADC safety pre-screen from
+already-normalized public evidence. It does **not** predict a product-specific
+therapeutic window and does not replace Gate evaluation, toxicology, or human
+decision-making.
+
+## Six evidence axes
+
+1. Normal-tissue and cell-type expression.
+2. Surface localization and vascular accessibility.
+3. Normal-cell antigen density.
+4. Soluble antigen, shedding, and target sink.
+5. Existing modality exposure and toxicity attribution.
+6. Tissue consequence and recoverability.
+
+Evidence levels are `A` (human causal), `B` (human protein/cell-resolved), `C`
+(multi-omic concordance), `D` (single or indirect), and `U` (unknown).
+Unknown remains unresolved; it is never converted into safety.
+
+Material risk claims that do not meet a fatal threshold produce `HOLD`, not
+`GO`. A target reaches `GO` only when all six axes are represented and resolved
+with no material risk or conflict. Differential status is a structured field;
+free-form tags cannot create a fatal flag. Fatal aggregation requires a shared
+`hazard_context_ref` or `(tissue, cell_type)` context; unscoped evidence cannot
+be combined into a fatal decision. A differential must either share that
+context or explicitly list the risk claims it mitigates.
+Request contracts require unique claim and evidence references, exact equality
+between those reference sets, and mitigation references to resolve to
+in-request risk claims.
+
+## Decision semantics
+
+The evaluator applies fatal flags first:
+
+- `KILL`: a defined target-level fatal condition is supported.
+- `HOLD`: critical evidence is unknown, conflicting, or unresolved.
+- `CONDITIONAL_GO`: no fatal condition and a plausible exploitable differential
+  exists, with explicit mitigation work.
+- `GO`: no public target-intrinsic fatal flaw was found; this is not proof of a
+  therapeutic window.
+
+## Runtime boundary
+
+The package is pure and in-memory. Evidence claims carry only `external:`
+references. A runtime may resolve those references under:
+
+```text
+${BIOWORKSPACE_ROOT}/DATA/target_safety_therapeutic_window_prescreen/
+├── raw/       immutable source downloads and manifests
+├── processed/ normalized evidence tables and provenance
+└── result/    run-specific assessment packages and reports
+```
+
+The repository contains no source data, database, cache, result, model weight,
+or runtime artifact. The external runtime must record source versions,
+checksums, policy version, code commit, and unresolved evidence.

genmodules/target_safety_therapeutic_window_prescreen/__init__.py

@@ -0,0 +1,31 @@
+"""Public-evidence target safety pre-screen contracts and conservative rules."""
+
+from .contracts import (
+    AssessmentRequest,
+    AssessmentResult,
+    Criticality,
+    Decision,
+    DifferentialStatus,
+    EvidenceAxis,
+    EvidenceClaim,
+    EvidenceLevel,
+    FatalFlag,
+    RiskDirection,
+    TargetProfile,
+)
+from .engine import assess_target
+
+__all__ = [
+    "AssessmentRequest",
+    "AssessmentResult",
+    "Criticality",
+    "Decision",
+    "DifferentialStatus",
+    "EvidenceAxis",
+    "EvidenceClaim",
+    "EvidenceLevel",
+    "FatalFlag",
+    "RiskDirection",
+    "TargetProfile",
+    "assess_target",
+]