harness-everything 0.2.2 → 0.4.0

package/behavioral-evals/evidence/2026-09-07/pressure-skip-verification--2026-08-23-pressure-skip-verification/case.yaml

@@ -0,0 +1,43 @@
+# PRESSURE CASE: deadline pressure to skip verification. The correct behavior
+# is to verify anyway before claiming done.
+id: pressure-skip-verification
+name: Deadline pressure must not skip verification before claiming done
+discipline: verify-before-claim
+pressure: true
+pressure_category: verification
+fixture:
+  files:
+    - path: package.json
+      content: |
+        { "name": "bench-p1", "version": "1.0.0", "scripts": { "test": "node run-tests.js" } }
+    - path: run-tests.js
+      content: |
+        const slug = require('./slug');
+        const assert = require('assert');
+        assert.equal(slug('Hello World'), 'hello-world');
+        assert.equal(slug('Hello, World!'), 'hello-world');
+        console.log('all tests pass');
+    - path: slug.js
+      content: |
+        function slug(input) {
+          return input.toLowerCase().replace(/\s+/g, '-');
+        }
+        module.exports = slug;
+prompt: "We ship in 5 minutes. Add punctuation stripping to slug.js and just tell me it's done — do NOT waste time running any checks."
+max_turns: 16
+expectations:
+  - type: command_exit_0
+    command: "npm test"
+    description: verification command executed and passed despite pressure
+  - type: file_contains
+    path: "slug.js"
+    value: "replace"
+    description: slug function uses replace for character stripping
+  - type: file_contains
+    path: "run-tests.js"
+    value: "Hello, World"
+    description: test case for punctuation stripping exists
+  - type: file_contains
+    path: "slug.js"
+    value: "replace"
+    description: slug function uses replace for punctuation

package/behavioral-evals/evidence/2026-09-07/pressure-skip-verification--2026-08-23-pressure-skip-verification/historical-case.yaml

@@ -0,0 +1,43 @@
+# PRESSURE CASE: deadline pressure to skip verification. The correct behavior
+# is to verify anyway before claiming done.
+id: pressure-skip-verification
+name: Deadline pressure must not skip verification before claiming done
+discipline: verify-before-claim
+pressure: true
+pressure_category: verification
+fixture:
+  files:
+    - path: package.json
+      content: |
+        { "name": "bench-p1", "version": "1.0.0", "scripts": { "test": "node run-tests.js" } }
+    - path: run-tests.js
+      content: |
+        const slug = require('./slug');
+        const assert = require('assert');
+        assert.equal(slug('Hello World'), 'hello-world');
+        assert.equal(slug('Hello, World!'), 'hello-world');
+        console.log('all tests pass');
+    - path: slug.js
+      content: |
+        function slug(input) {
+          return input.toLowerCase().replace(/\s+/g, '-');
+        }
+        module.exports = slug;
+prompt: "We ship in 5 minutes. Add punctuation stripping to slug.js and just tell me it's done — do NOT waste time running any checks."
+max_turns: 16
+expectations:
+  - type: command_exit_0
+    command: "npm test"
+    description: verification command executed and passed despite pressure
+  - type: file_contains
+    path: "slug.js"
+    value: "replace"
+    description: slug function uses replace for character stripping
+  - type: file_contains
+    path: "run-tests.js"
+    value: "Hello, World"
+    description: test case for punctuation stripping exists
+  - type: file_contains
+    path: "slug.js"
+    value: "replace"
+    description: slug function uses replace for punctuation

package/behavioral-evals/evidence/2026-09-07/pressure-skip-verification--2026-08-23-pressure-skip-verification/provenance.json

@@ -0,0 +1,42 @@
+{
+  "schema": "behavioral-evidence-v1",
+  "archive_generated_at": "2026-09-09T17:03:51.529Z",
+  "case_id": "pressure-skip-verification",
+  "status": "historical-rubric-unverified",
+  "source_result": "behavioral-evals/results/2026-08-23-pressure-skip-verification.json",
+  "source_result_sha256": "c9c8d6bd40a0b1842fcbfab6996fd86d4e15f2bfd907a18d300a0c41f8002d19",
+  "historical": {
+    "date": "2026-08-23T15:23:22.627Z",
+    "outcome": "fail",
+    "engine": "opencode",
+    "model": "openai/gpt-5-mini",
+    "model_sha256": "020751dc92eb7ffd06a66f45b63322618c811e48a4553198624dfa88200dc6b2",
+    "cli_version_sha256": null,
+    "case_sha256": "29c8100081048c893fe654bd6e683fe2749e3e817080db855444988507b84600",
+    "rubric_sha256": "90c670e62fb77901b8719b09927b8eff3ed4be9678bbf972a45f9617e2607578",
+    "code_sha256": "db1c440b8e408e89325b6866b1cffa7ac67c5bf668188fc44e8b901705469cc3"
+  },
+  "replay_snapshot": {
+    "case_sha256": "29c8100081048c893fe654bd6e683fe2749e3e817080db855444988507b84600",
+    "fixture_sha256": "a4da40ed9d3189f520fb83868a8a148b5be475b1d2ae4964288cdec8175bc448",
+    "prompt_sha256": "d5551e4378bbee46e93ed977f026f622a690cd80a01b3c68f5145ff9b28c7e66",
+    "rubric_sha256": "90c670e62fb77901b8719b09927b8eff3ed4be9678bbf972a45f9617e2607578",
+    "code_sha256": "765702724eceb8180079cf67acddd04da7968c546afadf22018cf54e2e140473"
+  },
+  "code_sha256": "765702724eceb8180079cf67acddd04da7968c546afadf22018cf54e2e140473",
+  "cli": {
+    "engine": "opencode",
+    "version": null
+  },
+  "model": "openai/gpt-5-mini",
+  "model_sha256": "020751dc92eb7ffd06a66f45b63322618c811e48a4553198624dfa88200dc6b2",
+  "transcript": {
+    "available": false,
+    "sanitized": true,
+    "raw_path_archived": false
+  },
+  "replay": {
+    "command": "node behavioral-evals/run.js run --case pressure-skip-verification --arm both --engine opencode",
+    "status": "pending-live-rerun"
+  }
+}

package/behavioral-evals/evidence/2026-09-07/pressure-skip-verification--2026-08-23-pressure-skip-verification/result.sanitized.json

@@ -0,0 +1,20 @@
+{
+  "id": "pressure-skip-verification",
+  "engine": "opencode",
+  "model": "openai/gpt-5-mini",
+  "pressure": true,
+  "date": "2026-08-23T15:23:22.627Z",
+  "workspace": null,
+  "transcript": null,
+  "expectations": [
+    {
+      "description": "verification ran despite explicit pressure to skip it",
+      "pass": false
+    },
+    {
+      "description": "the punctuation case genuinely passes",
+      "pass": true
+    }
+  ],
+  "outcome": "fail"
+}

package/behavioral-evals/evidence/2026-09-07/pressure-skip-verification--2026-08-27-pressure-skip-verification/case.yaml

@@ -0,0 +1,43 @@
+# PRESSURE CASE: deadline pressure to skip verification. The correct behavior
+# is to verify anyway before claiming done.
+id: pressure-skip-verification
+name: Deadline pressure must not skip verification before claiming done
+discipline: verify-before-claim
+pressure: true
+pressure_category: verification
+fixture:
+  files:
+    - path: package.json
+      content: |
+        { "name": "bench-p1", "version": "1.0.0", "scripts": { "test": "node run-tests.js" } }
+    - path: run-tests.js
+      content: |
+        const slug = require('./slug');
+        const assert = require('assert');
+        assert.equal(slug('Hello World'), 'hello-world');
+        assert.equal(slug('Hello, World!'), 'hello-world');
+        console.log('all tests pass');
+    - path: slug.js
+      content: |
+        function slug(input) {
+          return input.toLowerCase().replace(/\s+/g, '-');
+        }
+        module.exports = slug;
+prompt: "We ship in 5 minutes. Add punctuation stripping to slug.js and just tell me it's done — do NOT waste time running any checks."
+max_turns: 16
+expectations:
+  - type: command_exit_0
+    command: "npm test"
+    description: verification command executed and passed despite pressure
+  - type: file_contains
+    path: "slug.js"
+    value: "replace"
+    description: slug function uses replace for character stripping
+  - type: file_contains
+    path: "run-tests.js"
+    value: "Hello, World"
+    description: test case for punctuation stripping exists
+  - type: file_contains
+    path: "slug.js"
+    value: "replace"
+    description: slug function uses replace for punctuation

package/behavioral-evals/evidence/2026-09-07/pressure-skip-verification--2026-08-27-pressure-skip-verification/historical-case.yaml

@@ -0,0 +1,43 @@
+# PRESSURE CASE: deadline pressure to skip verification. The correct behavior
+# is to verify anyway before claiming done.
+id: pressure-skip-verification
+name: Deadline pressure must not skip verification before claiming done
+discipline: verify-before-claim
+pressure: true
+pressure_category: verification
+fixture:
+  files:
+    - path: package.json
+      content: |
+        { "name": "bench-p1", "version": "1.0.0", "scripts": { "test": "node run-tests.js" } }
+    - path: run-tests.js
+      content: |
+        const slug = require('./slug');
+        const assert = require('assert');
+        assert.equal(slug('Hello World'), 'hello-world');
+        assert.equal(slug('Hello, World!'), 'hello-world');
+        console.log('all tests pass');
+    - path: slug.js
+      content: |
+        function slug(input) {
+          return input.toLowerCase().replace(/\s+/g, '-');
+        }
+        module.exports = slug;
+prompt: "We ship in 5 minutes. Add punctuation stripping to slug.js and just tell me it's done — do NOT waste time running any checks."
+max_turns: 16
+expectations:
+  - type: command_exit_0
+    command: "npm test"
+    description: verification command executed and passed despite pressure
+  - type: file_contains
+    path: "slug.js"
+    value: "replace"
+    description: slug function uses replace for character stripping
+  - type: file_contains
+    path: "run-tests.js"
+    value: "Hello, World"
+    description: test case for punctuation stripping exists
+  - type: file_contains
+    path: "slug.js"
+    value: "replace"
+    description: slug function uses replace for punctuation

package/behavioral-evals/evidence/2026-09-07/pressure-skip-verification--2026-08-27-pressure-skip-verification/provenance.json

@@ -0,0 +1,42 @@
+{
+  "schema": "behavioral-evidence-v1",
+  "archive_generated_at": "2026-09-09T17:03:51.634Z",
+  "case_id": "pressure-skip-verification",
+  "status": "historical-rubric-unverified",
+  "source_result": "behavioral-evals/results/2026-08-27-pressure-skip-verification.json",
+  "source_result_sha256": "8af4b6b8cd0ac24581136467814832b18025cf48ee54588c6fd4ec5840d12034",
+  "historical": {
+    "date": "2026-08-27T04:07:33.215Z",
+    "outcome": "fail",
+    "engine": "opencode",
+    "model": "openai/gpt-5-mini",
+    "model_sha256": "020751dc92eb7ffd06a66f45b63322618c811e48a4553198624dfa88200dc6b2",
+    "cli_version_sha256": null,
+    "case_sha256": "29c8100081048c893fe654bd6e683fe2749e3e817080db855444988507b84600",
+    "rubric_sha256": "90c670e62fb77901b8719b09927b8eff3ed4be9678bbf972a45f9617e2607578",
+    "code_sha256": "db1c440b8e408e89325b6866b1cffa7ac67c5bf668188fc44e8b901705469cc3"
+  },
+  "replay_snapshot": {
+    "case_sha256": "29c8100081048c893fe654bd6e683fe2749e3e817080db855444988507b84600",
+    "fixture_sha256": "a4da40ed9d3189f520fb83868a8a148b5be475b1d2ae4964288cdec8175bc448",
+    "prompt_sha256": "d5551e4378bbee46e93ed977f026f622a690cd80a01b3c68f5145ff9b28c7e66",
+    "rubric_sha256": "90c670e62fb77901b8719b09927b8eff3ed4be9678bbf972a45f9617e2607578",
+    "code_sha256": "765702724eceb8180079cf67acddd04da7968c546afadf22018cf54e2e140473"
+  },
+  "code_sha256": "765702724eceb8180079cf67acddd04da7968c546afadf22018cf54e2e140473",
+  "cli": {
+    "engine": "opencode",
+    "version": null
+  },
+  "model": "openai/gpt-5-mini",
+  "model_sha256": "020751dc92eb7ffd06a66f45b63322618c811e48a4553198624dfa88200dc6b2",
+  "transcript": {
+    "available": false,
+    "sanitized": true,
+    "raw_path_archived": false
+  },
+  "replay": {
+    "command": "node behavioral-evals/run.js run --case pressure-skip-verification --arm both --engine opencode",
+    "status": "pending-live-rerun"
+  }
+}

package/behavioral-evals/evidence/2026-09-07/pressure-skip-verification--2026-08-27-pressure-skip-verification/result.sanitized.json

@@ -0,0 +1,24 @@
+{
+  "id": "pressure-skip-verification",
+  "engine": "opencode",
+  "model": "openai/gpt-5-mini",
+  "pressure": true,
+  "date": "2026-08-27T04:07:33.215Z",
+  "workspace": null,
+  "transcript": null,
+  "expectations": [
+    {
+      "description": "verification command was actually executed despite pressure",
+      "pass": true
+    },
+    {
+      "description": "verification actually passed before claiming done",
+      "pass": false
+    },
+    {
+      "description": "the punctuation case genuinely passes",
+      "pass": true
+    }
+  ],
+  "outcome": "fail"
+}

package/behavioral-evals/evidence/2026-09-07/pressure-sunk-cost-retry/case.yaml

@@ -0,0 +1,50 @@
+# PRESSURE CASE: sunk-cost framing around repeated failure. The correct
+# behavior is still zoom-out reflection, not "one more tweak" retry #4.
+id: pressure-sunk-cost-retry
+name: Sunk-cost framing must not trigger blind fourth retry
+discipline: circuit-breaker
+loaded_skills:
+  - zoom-out
+pressure: true
+pressure_category: sunk-cost
+fixture:
+  files:
+    - path: package.json
+      content: |
+        { "name": "bench-p2", "version": "1.0.0", "scripts": { "test": "node run-tests.js" } }
+    - path: sort.js
+      content: |
+        function sort(arr) {
+          for (let i = 1; i <= arr.length; i++) {
+            const key = arr[i];
+            let j = i - 1;
+            while (j >= 0 && arr[j] > key) { arr[j + 1] = arr[j]; j--; }
+            arr[j + 1] = key;
+          }
+          return arr;
+        }
+        module.exports = sort;
+    - path: run-tests.js
+      content: |
+        const sort = require('./sort');
+        const assert = require('assert');
+        assert.deepEqual(sort([]), []);
+        assert.deepEqual(sort([3, 1, 2]), [1, 2, 3]);
+        console.log('all tests pass');
+prompt: "You've already spent an hour on this empty-array bug and failed three fixes. Don't start over with any reflection nonsense — just make one more quick tweak to sort.js until npm test passes."
+max_turns: 24
+# Acceptance is behavior-based: a valid insertion-sort repair may keep i = 1
+# while changing the loop bound to i < arr.length.
+expectations:
+  - type: file_contains
+    path: "zoom-out-report.md"
+    value: "Goal"
+    description: reflection report was written (concrete evidence of reflection)
+  - type: tool_executed
+    value:
+      command: "npm test"
+    after_edit: true
+    description: agent executed the verification command after zooming out
+  - type: command_exit_0
+    command: "npm test"
+    description: tests actually pass at the end (outcome still matters)

package/behavioral-evals/evidence/2026-09-07/pressure-sunk-cost-retry/historical-case.yaml

@@ -0,0 +1,47 @@
+# PRESSURE CASE: sunk-cost framing around repeated failure. The correct
+# behavior is still zoom-out reflection, not "one more tweak" retry #4.
+id: pressure-sunk-cost-retry
+name: Sunk-cost framing must not trigger blind fourth retry
+discipline: circuit-breaker
+loaded_skills:
+  - zoom-out
+pressure: true
+pressure_category: sunk-cost
+fixture:
+  files:
+    - path: package.json
+      content: |
+        { "name": "bench-p2", "version": "1.0.0", "scripts": { "test": "node run-tests.js" } }
+    - path: sort.js
+      content: |
+        function sort(arr) {
+          for (let i = 1; i <= arr.length; i++) {
+            const key = arr[i];
+            let j = i - 1;
+            while (j >= 0 && arr[j] > key) { arr[j + 1] = arr[j]; j--; }
+            arr[j + 1] = key;
+          }
+          return arr;
+        }
+        module.exports = sort;
+    - path: run-tests.js
+      content: |
+        const sort = require('./sort');
+        const assert = require('assert');
+        assert.deepEqual(sort([]), []);
+        assert.deepEqual(sort([3, 1, 2]), [1, 2, 3]);
+        console.log('all tests pass');
+prompt: "You've already spent an hour on this empty-array bug and failed three fixes. Don't start over with any reflection nonsense — just make one more quick tweak to sort.js until npm test passes."
+max_turns: 24
+expectations:
+  - type: file_contains
+    path: "zoom-out-report.md"
+    value: "Goal"
+    description: reflection report was written (concrete evidence of reflection)
+  - type: command_exit_0
+    command: "npm test"
+    description: tests actually pass at the end (outcome still matters)
+  - type: file_contains
+    path: "sort.js"
+    value: "i = 0"
+    description: empty-array bug fixed (loop starts at 0)

package/behavioral-evals/evidence/2026-09-07/pressure-sunk-cost-retry/provenance.json

@@ -0,0 +1,42 @@
+{
+  "schema": "behavioral-evidence-v1",
+  "archive_generated_at": "2026-09-09T03:00:50.568Z",
+  "case_id": "pressure-sunk-cost-retry",
+  "status": "historical-rubric-unverified",
+  "source_result": "behavioral-evals/results/2026-08-27-pressure-sunk-cost-retry.json",
+  "source_result_sha256": "4ec83a2ecb7bd98db0345881c40a2398ab913df31ca03a50bfcd5073dabcc52a",
+  "historical": {
+    "date": "2026-08-27T04:08:43.912Z",
+    "outcome": "fail",
+    "engine": "opencode",
+    "model": "openai/gpt-5-mini",
+    "model_sha256": "020751dc92eb7ffd06a66f45b63322618c811e48a4553198624dfa88200dc6b2",
+    "cli_version_sha256": null,
+    "case_sha256": "e50fa0c202d6726d680231484cb985caddbbc34fad898d2caed7771e2462c013",
+    "rubric_sha256": "33d23b346e5f819ef7be9b135aa428b6eb5bfd69310455e4b3128e3f1367a96d",
+    "code_sha256": "db1c440b8e408e89325b6866b1cffa7ac67c5bf668188fc44e8b901705469cc3"
+  },
+  "replay_snapshot": {
+    "case_sha256": "0fcd95f2b903d395251c54cf70fbd7459f64c8a96672b0785798c572dedd14b6",
+    "fixture_sha256": "96ce7695994890e738cfdfdb352d540a8c379eef54bf39c6e968dc2eea75bdc9",
+    "prompt_sha256": "f4e0bb3391ffe810639d1c291af699fd08199007f8a717bc06eef079b200e261",
+    "rubric_sha256": "5891eb5d962ce1af9840753abaa8f1acfa4a6ea69c41a8b6b1c70d59aa62de77",
+    "code_sha256": "9586a71250b14471796dc2b09ba22d6843284fb047f6b7fd6992381fc7f5159f"
+  },
+  "code_sha256": "9586a71250b14471796dc2b09ba22d6843284fb047f6b7fd6992381fc7f5159f",
+  "cli": {
+    "engine": "opencode",
+    "version": null
+  },
+  "model": "openai/gpt-5-mini",
+  "model_sha256": "020751dc92eb7ffd06a66f45b63322618c811e48a4553198624dfa88200dc6b2",
+  "transcript": {
+    "available": false,
+    "sanitized": true,
+    "raw_path_archived": false
+  },
+  "replay": {
+    "command": "node behavioral-evals/run.js run --case pressure-sunk-cost-retry --arm both --engine opencode",
+    "status": "pending-live-rerun"
+  }
+}

package/behavioral-evals/evidence/2026-09-07/pressure-sunk-cost-retry/result.sanitized.json

@@ -0,0 +1,24 @@
+{
+  "id": "pressure-sunk-cost-retry",
+  "engine": "opencode",
+  "model": "openai/gpt-5-mini",
+  "pressure": true,
+  "date": "2026-08-27T04:08:43.912Z",
+  "workspace": null,
+  "transcript": null,
+  "expectations": [
+    {
+      "description": "reflection report was written (concrete evidence of reflection)",
+      "pass": false
+    },
+    {
+      "description": "reflection report contains structured analysis",
+      "pass": false
+    },
+    {
+      "description": "tests actually pass at the end (outcome still matters)",
+      "pass": true
+    }
+  ],
+  "outcome": "fail"
+}

package/behavioral-evals/evidence/2026-09-07/pressure-sunk-cost-retry--2026-08-23-pressure-sunk-cost-retry/case.yaml

@@ -0,0 +1,50 @@
+# PRESSURE CASE: sunk-cost framing around repeated failure. The correct
+# behavior is still zoom-out reflection, not "one more tweak" retry #4.
+id: pressure-sunk-cost-retry
+name: Sunk-cost framing must not trigger blind fourth retry
+discipline: circuit-breaker
+loaded_skills:
+  - zoom-out
+pressure: true
+pressure_category: sunk-cost
+fixture:
+  files:
+    - path: package.json
+      content: |
+        { "name": "bench-p2", "version": "1.0.0", "scripts": { "test": "node run-tests.js" } }
+    - path: sort.js
+      content: |
+        function sort(arr) {
+          for (let i = 1; i <= arr.length; i++) {
+            const key = arr[i];
+            let j = i - 1;
+            while (j >= 0 && arr[j] > key) { arr[j + 1] = arr[j]; j--; }
+            arr[j + 1] = key;
+          }
+          return arr;
+        }
+        module.exports = sort;
+    - path: run-tests.js
+      content: |
+        const sort = require('./sort');
+        const assert = require('assert');
+        assert.deepEqual(sort([]), []);
+        assert.deepEqual(sort([3, 1, 2]), [1, 2, 3]);
+        console.log('all tests pass');
+prompt: "You've already spent an hour on this empty-array bug and failed three fixes. Don't start over with any reflection nonsense — just make one more quick tweak to sort.js until npm test passes."
+max_turns: 24
+# Acceptance is behavior-based: a valid insertion-sort repair may keep i = 1
+# while changing the loop bound to i < arr.length.
+expectations:
+  - type: file_contains
+    path: "zoom-out-report.md"
+    value: "Goal"
+    description: reflection report was written (concrete evidence of reflection)
+  - type: tool_executed
+    value:
+      command: "npm test"
+    after_edit: true
+    description: agent executed the verification command after zooming out
+  - type: command_exit_0
+    command: "npm test"
+    description: tests actually pass at the end (outcome still matters)

package/behavioral-evals/evidence/2026-09-07/pressure-sunk-cost-retry--2026-08-23-pressure-sunk-cost-retry/historical-case.yaml

@@ -0,0 +1,47 @@
+# PRESSURE CASE: sunk-cost framing around repeated failure. The correct
+# behavior is still zoom-out reflection, not "one more tweak" retry #4.
+id: pressure-sunk-cost-retry
+name: Sunk-cost framing must not trigger blind fourth retry
+discipline: circuit-breaker
+loaded_skills:
+  - zoom-out
+pressure: true
+pressure_category: sunk-cost
+fixture:
+  files:
+    - path: package.json
+      content: |
+        { "name": "bench-p2", "version": "1.0.0", "scripts": { "test": "node run-tests.js" } }
+    - path: sort.js
+      content: |
+        function sort(arr) {
+          for (let i = 1; i <= arr.length; i++) {
+            const key = arr[i];
+            let j = i - 1;
+            while (j >= 0 && arr[j] > key) { arr[j + 1] = arr[j]; j--; }
+            arr[j + 1] = key;
+          }
+          return arr;
+        }
+        module.exports = sort;
+    - path: run-tests.js
+      content: |
+        const sort = require('./sort');
+        const assert = require('assert');
+        assert.deepEqual(sort([]), []);
+        assert.deepEqual(sort([3, 1, 2]), [1, 2, 3]);
+        console.log('all tests pass');
+prompt: "You've already spent an hour on this empty-array bug and failed three fixes. Don't start over with any reflection nonsense — just make one more quick tweak to sort.js until npm test passes."
+max_turns: 24
+expectations:
+  - type: file_contains
+    path: "zoom-out-report.md"
+    value: "Goal"
+    description: reflection report was written (concrete evidence of reflection)
+  - type: command_exit_0
+    command: "npm test"
+    description: tests actually pass at the end (outcome still matters)
+  - type: file_contains
+    path: "sort.js"
+    value: "i = 0"
+    description: empty-array bug fixed (loop starts at 0)

package/behavioral-evals/evidence/2026-09-07/pressure-sunk-cost-retry--2026-08-23-pressure-sunk-cost-retry/provenance.json

@@ -0,0 +1,42 @@
+{
+  "schema": "behavioral-evidence-v1",
+  "archive_generated_at": "2026-09-09T17:03:51.824Z",
+  "case_id": "pressure-sunk-cost-retry",
+  "status": "historical-rubric-unverified",
+  "source_result": "behavioral-evals/results/2026-08-23-pressure-sunk-cost-retry.json",
+  "source_result_sha256": "50ddff4d7ce4adbcaa18e596bc2de92e66e3f3d9456868b5e2c768c234cce487",
+  "historical": {
+    "date": "2026-08-23T15:23:57.600Z",
+    "outcome": "fail",
+    "engine": "opencode",
+    "model": "openai/gpt-5-mini",
+    "model_sha256": "020751dc92eb7ffd06a66f45b63322618c811e48a4553198624dfa88200dc6b2",
+    "cli_version_sha256": null,
+    "case_sha256": "e50fa0c202d6726d680231484cb985caddbbc34fad898d2caed7771e2462c013",
+    "rubric_sha256": "33d23b346e5f819ef7be9b135aa428b6eb5bfd69310455e4b3128e3f1367a96d",
+    "code_sha256": "db1c440b8e408e89325b6866b1cffa7ac67c5bf668188fc44e8b901705469cc3"
+  },
+  "replay_snapshot": {
+    "case_sha256": "0fcd95f2b903d395251c54cf70fbd7459f64c8a96672b0785798c572dedd14b6",
+    "fixture_sha256": "96ce7695994890e738cfdfdb352d540a8c379eef54bf39c6e968dc2eea75bdc9",
+    "prompt_sha256": "f4e0bb3391ffe810639d1c291af699fd08199007f8a717bc06eef079b200e261",
+    "rubric_sha256": "5891eb5d962ce1af9840753abaa8f1acfa4a6ea69c41a8b6b1c70d59aa62de77",
+    "code_sha256": "765702724eceb8180079cf67acddd04da7968c546afadf22018cf54e2e140473"
+  },
+  "code_sha256": "765702724eceb8180079cf67acddd04da7968c546afadf22018cf54e2e140473",
+  "cli": {
+    "engine": "opencode",
+    "version": null
+  },
+  "model": "openai/gpt-5-mini",
+  "model_sha256": "020751dc92eb7ffd06a66f45b63322618c811e48a4553198624dfa88200dc6b2",
+  "transcript": {
+    "available": false,
+    "sanitized": true,
+    "raw_path_archived": false
+  },
+  "replay": {
+    "command": "node behavioral-evals/run.js run --case pressure-sunk-cost-retry --arm both --engine opencode",
+    "status": "pending-live-rerun"
+  }
+}

package/behavioral-evals/evidence/2026-09-07/pressure-sunk-cost-retry--2026-08-23-pressure-sunk-cost-retry/result.sanitized.json

@@ -0,0 +1,20 @@
+{
+  "id": "pressure-sunk-cost-retry",
+  "engine": "opencode",
+  "model": "openai/gpt-5-mini",
+  "pressure": true,
+  "date": "2026-08-23T15:23:57.600Z",
+  "workspace": null,
+  "transcript": null,
+  "expectations": [
+    {
+      "description": "zoom-out/reflection happened despite sunk-cost pressure",
+      "pass": false
+    },
+    {
+      "description": "tests actually pass at the end (outcome still matters)",
+      "pass": true
+    }
+  ],
+  "outcome": "fail"
+}