test(evals): add technical-design bad pointwise fixtures

evals/cases/case-aerial-delivery-shipping-v1/candidate-bad-pointwise.md

@@ -0,0 +1,13 @@
+# Expected-Bad Pointwise Candidate
+
+This candidate is intentionally defective for manual pointwise calibration. It is not a reference design and should remain adverse.
+
+## Candidate Design
+
+The design uses Shipping for shipping-facing workflow coordination, Delivery for active pickup and delivery execution, Package for package tagging, Scheduler for assignment policy, Accounts for business registration records, and Drone Management for drone availability.
+
+## Intentional Calibration Defect
+
+Delivery owns long-term delivery history.
+
+This contradicts the source-visible retention boundary: active delivery execution may be separate from long-term delivery history ownership.

evals/cases/case-cloudevents-core-contract-v1/candidate-bad-pointwise.md

@@ -0,0 +1,13 @@
+# Expected-Bad Pointwise Candidate
+
+This candidate is intentionally defective for manual pointwise calibration. It is not a reference design and should remain adverse.
+
+## Candidate Design
+
+The design keeps a CloudEvents Event Contract as the interoperability seam, supports structured and binary JSON encodings, and allows domain-specific payloads to travel between producers and consumers.
+
+## Intentional Calibration Defect
+
+CloudEvents defines event-sourcing lifecycle.
+
+This invents event-store and event-sourcing ownership that the contract fixture explicitly keeps out of scope.

evals/cases/case-customer-credit-order-saga-v1/candidate-bad-pointwise.md

@@ -0,0 +1,13 @@
+# Expected-Bad Pointwise Candidate
+
+This candidate is intentionally defective for manual pointwise calibration. It is not a reference design and should remain adverse.
+
+## Candidate Design
+
+The design has an Ordering boundary that creates orders in `PENDING`, sends `reserveCredit` to Customer Credit, and records `APPROVED`, `REJECTED`, `UNKNOWN_CUSTOMER`, and `INSUFFICIENT_CREDIT` outcomes. Messaging and Integration Infrastructure carries commands and asynchronous replies.
+
+## Intentional Calibration Defect
+
+The design uses 2PC for consistency between Order Service and Customer Service.
+
+This contradicts the local-transaction saga model and the separate service persistence boundary.

evals/cases/case-fineract-loan-lifecycle-v1/candidate-bad-pointwise.md

@@ -0,0 +1,13 @@
+# Expected-Bad Pointwise Candidate
+
+This candidate is intentionally defective for manual pointwise calibration. It is not a reference design and should remain adverse.
+
+## Candidate Design
+
+The design separates Loan Account Lifecycle, Loan Product Configuration, Loan Charges, and Accounting Integration. Loan Account Lifecycle tracks submitted, approved, active, closed, and overpaid states while reading loan product terms.
+
+## Intentional Calibration Defect
+
+Loan Account Lifecycle owns product configuration.
+
+This collapses loan product template authority into the account lifecycle boundary.

evals/cases/case-kubernetes-sidecar-containers-v1/candidate-bad-pointwise.md

@@ -0,0 +1,13 @@
+# Expected-Bad Pointwise Candidate
+
+This candidate is intentionally defective for manual pointwise calibration. It is not a reference design and should remain adverse.
+
+## Candidate Design
+
+The design treats sidecar behavior as part of Pod lifecycle semantics, keeps init-container ordering visible, and discusses Job completion, readiness, resource accounting, and termination ordering.
+
+## Intentional Calibration Defect
+
+Sidecars are ordinary app containers.
+
+This contradicts the source-visible rule that sidecars are restartable init containers under `initContainers` with `restartPolicy: Always`.

evals/cases/case-openfeature-evaluation-api-v1/candidate-bad-pointwise.md

@@ -0,0 +1,13 @@
+# Expected-Bad Pointwise Candidate
+
+This candidate is intentionally defective for manual pointwise calibration. It is not a reference design and should remain adverse.
+
+## Candidate Design
+
+The design keeps OpenFeature as a vendor-agnostic evaluation API with providers as backend adapters, no-op default behavior, hook lifecycle stages, and provider readiness/error/configuration/stale events.
+
+## Intentional Calibration Defect
+
+OpenFeature owns flag management UI.
+
+This invents product-management scope that belongs outside the evaluation API seam.

evals/cases/case-tiny-laundry-pickup-v1/candidate-bad-pointwise.md

@@ -0,0 +1,13 @@
+# Expected-Bad Pointwise Candidate
+
+This candidate is intentionally defective for manual pointwise calibration. It is not a reference design and should remain adverse.
+
+## Candidate Design
+
+The design has Scheduling own booking lifecycle, overlap detection, cancellation rules, appliance availability reads, maintenance hold reads, and the public booking APIs.
+
+## Intentional Calibration Defect
+
+Identity owns scheduling conflict logic.
+
+This contradicts the source-visible boundary: identity may prove resident verification, but scheduling conflict decisions stay with the booking lifecycle owner.

evals/model-judge-calibration-notes.md

@@ -8,15 +8,15 @@ Deterministic eval verdicts remain authoritative. Pointwise model-judge evidence
 
 These notes summarize the existing local provider-backed runs from July 4, 2026. No fresh provider calls were made for this curated summary.
 
-| Field             | Value                                        |
-| ----------------- | -------------------------------------------- |
-| Config            | `evals/eval-kit.model-judge.config.json`     |
-| Model             | `gpt-5.4`                                    |
-| Provider          | `openai:codex-app-server`                    |
-| Reasoning effort  | `medium`                                     |
-| Prompt version    | `pointwise-prompt-v1`                        |
-| Rubric version    | `pointwise-coverage-rubric-v1`               |
-| Candidate profile | expected-good `reference-design.md` fixtures |
+| Field             | Value                                                                                                                 |
+| ----------------- | --------------------------------------------------------------------------------------------------------------------- |
+| Config            | `evals/eval-kit.model-judge.config.json`                                                                              |
+| Model             | `gpt-5.4`                                                                                                             |
+| Provider          | `openai:codex-app-server`                                                                                             |
+| Reasoning effort  | `medium`                                                                                                              |
+| Prompt version    | `pointwise-prompt-v1`                                                                                                 |
+| Rubric version    | `pointwise-coverage-rubric-v1`                                                                                        |
+| Candidate profile | expected-good `reference-design.md` fixtures; expected-bad candidates now committed for future false-pass calibration |
 
 ## Calibration Summary
 
@@ -32,7 +32,7 @@ These notes summarize the existing local provider-backed runs from July 4, 2026.
 
 ## Human Calibration Notes
 
-- False pass risk: not yet calibrated for technical-design because this run set contains expected-good pointwise candidates only. Phase C4 is the planned follow-up for expected-bad pointwise calibration fixtures.
+- False pass risk: not yet provider-calibrated for technical-design because the summarized provider run set contains expected-good pointwise candidates only. Phase C4 added committed expected-bad pointwise candidates for future false-pass calibration.
 - False fail risk: no false-fail evidence observed in the expected-good run set.
 - Ambiguity: no `partial` or `unknown` verdicts appeared in these runs.
 - Verbosity bias: no evidence that longer wording was required for coverage, but expected-bad fixtures are needed before drawing stronger conclusions.
@@ -41,4 +41,18 @@ These notes summarize the existing local provider-backed runs from July 4, 2026.
 
 ## Decision
 
-Keep the technical-design pointwise lane manual and advisory. The current expected-good evidence is usable for smoke calibration, but false-pass detection remains untested until Phase C4 adds expected-bad pointwise calibration fixtures.
+Keep the technical-design pointwise lane manual and advisory. The current expected-good evidence is usable for smoke calibration. The committed expected-bad candidates below are ready for a future provider-backed false-pass calibration pass, but they do not add fresh model evidence by themselves.
+
+## Expected-Bad Candidate Set
+
+These candidates are committed for manual pointwise false-pass calibration. They intentionally contain technical-design-specific semantic defects only and are expected to remain adverse.
+
+| Case                                    | Candidate                                                                      | Targeted defect                                    |
+| --------------------------------------- | ------------------------------------------------------------------------------ | -------------------------------------------------- |
+| `case-aerial-delivery-shipping-v1`      | `evals/cases/case-aerial-delivery-shipping-v1/candidate-bad-pointwise.md`      | Delivery owns long-term delivery history.          |
+| `case-cloudevents-core-contract-v1`     | `evals/cases/case-cloudevents-core-contract-v1/candidate-bad-pointwise.md`     | CloudEvents owns event-sourcing lifecycle.         |
+| `case-customer-credit-order-saga-v1`    | `evals/cases/case-customer-credit-order-saga-v1/candidate-bad-pointwise.md`    | Order and Customer services use 2PC.               |
+| `case-fineract-loan-lifecycle-v1`       | `evals/cases/case-fineract-loan-lifecycle-v1/candidate-bad-pointwise.md`       | Loan Account Lifecycle owns product configuration. |
+| `case-kubernetes-sidecar-containers-v1` | `evals/cases/case-kubernetes-sidecar-containers-v1/candidate-bad-pointwise.md` | Sidecars are ordinary app containers.              |
+| `case-openfeature-evaluation-api-v1`    | `evals/cases/case-openfeature-evaluation-api-v1/candidate-bad-pointwise.md`    | OpenFeature owns flag management UI.               |
+| `case-tiny-laundry-pickup-v1`           | `evals/cases/case-tiny-laundry-pickup-v1/candidate-bad-pointwise.md`           | Identity owns scheduling conflict logic.           |

evals/tests/public-reference-cases.test.mjs

@@ -44,6 +44,37 @@ const newPublicCaseIds = [
   "case-fineract-loan-lifecycle-v1",
 ];
 
+const expectedBadPointwiseCandidates = [
+  {
+    caseId: "case-aerial-delivery-shipping-v1",
+    findingId: "CTX-005",
+  },
+  {
+    caseId: "case-cloudevents-core-contract-v1",
+    findingId: "FACT-005",
+  },
+  {
+    caseId: "case-customer-credit-order-saga-v1",
+    findingId: "FACT-008",
+  },
+  {
+    caseId: "case-fineract-loan-lifecycle-v1",
+    findingId: "CTX-001",
+  },
+  {
+    caseId: "case-kubernetes-sidecar-containers-v1",
+    findingId: "FACT-001",
+  },
+  {
+    caseId: "case-openfeature-evaluation-api-v1",
+    findingId: "FACT-006",
+  },
+  {
+    caseId: "case-tiny-laundry-pickup-v1",
+    findingId: "FACT-011",
+  },
+];
+
 describe("public reference cases", () => {
   it("accepts the new public reference designs", () => {
     for (const caseId of newPublicCaseIds) {
@@ -56,6 +87,17 @@ describe("public reference cases", () => {
     }
   });
 
+  it("keeps committed expected-bad pointwise candidates adverse", () => {
+    for (const { caseId, findingId } of expectedBadPointwiseCandidates) {
+      const candidate = readText(caseId, "candidate-bad-pointwise.md");
+      const findings = findingsForCaseCandidate(caseId, candidate);
+
+      expect(candidate, caseId).toContain("Expected-Bad Pointwise Candidate");
+      expect(verdictFor(findings, findingId), caseId).toBe("contradicted");
+      expect(verdictForFindings(findings), caseId).toBe("red");
+    }
+  });
+
   it("passes the case rubric into pointwise judge variables", async () => {
     const config = loadConfig(path.join(packageRoot, "eval-kit.config.json"));
     const caseId = "case-tiny-laundry-pickup-v1";