zorya-classifier 0.0.0-stage → 0.2.0

package/docs/evidence/jev-pricing-2026-10-18.json

@@ -0,0 +1,69 @@
+{
+  "checkedAt": "2026-10-17T21:31:17.256962+00:00",
+  "source": "https://openrouter.ai/api/v1/models/typesafe/jev-1.14/endpoints",
+  "response": {
+    "data": {
+      "id": "typesafe/jev-1.14",
+      "name": "TypeSafe: Jev 1.14",
+      "created": 1792021417,
+      "description": "Jev 1.14 is the latest version of Jev, TypeSafe's structured decision model and the first of its System One models. System One models make fast, structured decisions for software, returning...",
+      "architecture": {
+        "tokenizer": "Other",
+        "instruct_type": null,
+        "modality": "text+image->decisions",
+        "input_modalities": ["text", "image"],
+        "output_modalities": ["decisions"]
+      },
+      "endpoints": [
+        {
+          "name": "TypeSafe | typesafe/jev-1.14-20261014",
+          "model_id": "typesafe/jev-1.14",
+          "model_name": "TypeSafe: Jev 1.14",
+          "context_length": 262144,
+          "pricing": {
+            "prompt": "0.000000042",
+            "completion": "0",
+            "discount": 0
+          },
+          "provider_name": "TypeSafe",
+          "tag": "typesafe",
+          "service_tier": null,
+          "quantization": "unknown",
+          "max_completion_tokens": 235929,
+          "max_prompt_tokens": 262144,
+          "supported_parameters": [],
+          "supports_tool_choice": {
+            "none": true,
+            "auto": true,
+            "required": true,
+            "function": true
+          },
+          "status": 0,
+          "uptime_last_30m": 100,
+          "uptime_last_5m": 100,
+          "uptime_last_1d": 100,
+          "supports_implicit_caching": false,
+          "native_tools": {},
+          "supports_voice_cloning": false,
+          "supports_multiple_audio_references": false,
+          "supports_image_reference": false,
+          "supports_diarization": false,
+          "latency_last_30m": null,
+          "throughput_last_30m": null
+        }
+      ]
+    }
+  },
+  "calculation": {
+    "maxPromptTokens": 262144,
+    "usdPerInputToken": 4.2e-8,
+    "maximumInputUsd": 0.011010048,
+    "outputUsdPerToken": 0,
+    "reservedPerUnknownUsd": 0.02
+  },
+  "basis": [
+    "https://docs.typesafe.ai/models",
+    "https://openrouter.ai/docs/guides/community/multimodal-decisions"
+  ],
+  "interpretation": "Derived admission envelope: state and all questions share one per-request context; image requests use long context at normal input price. Not a provider-enforced account cap. Pricing checked before this run; future runs must re-check."
+}

package/docs/evidence/linux-check.json

@@ -0,0 +1,20 @@
+{
+  "status": "PASS",
+  "node": "v24.12.0",
+  "platform": "linux",
+  "command": "npm run check",
+  "checks": [
+    "secret worktree scan",
+    "typecheck",
+    "eslint --max-warnings 0",
+    "boundaries",
+    "dead-code:ci",
+    "160 tests, zero skipped",
+    "build"
+  ],
+  "sourceLogSha256": "652f6d6d3dd911fb39449b4102b83a0fa06c9dffcf0aa1f865508f6d73a421c8",
+  "runtimeDistSha256": "62d5a0075e5ccca6627ba81f8ca8be427f499ded7508570171c00520bcf369ca",
+  "runtimeDistFileCount": 34,
+  "scope": "Offline contract/lifecycle/geometry/provider-wire/secret guard tests, not model accuracy",
+  "note": "Native Windows runtime suite contains 142 tests; two later benchmark-only regressions are Linux-tested, and the 16 secret tests were run separately on Windows with one expected POSIX-only skip. Runtime dist remains identical."
+}

package/docs/evidence/linux-check.txt

@@ -0,0 +1,205 @@
+
+> zorya-classifier@0.1.0 check
+> npm run secrets:check && npm run typecheck && npm run lint && npm run boundaries && npm run dead-code:ci && npm test && npm run build
+
+
+> zorya-classifier@0.1.0 secrets:check
+> node scripts/secrets.mjs --worktree
+
+{"status":"PASS","mode":"--worktree","scanned":175,"skippedBinary":0,"totalBytes":1902577}
+
+> zorya-classifier@0.1.0 typecheck
+> tsc --noEmit
+
+
+> zorya-classifier@0.1.0 lint
+> eslint . --max-warnings 0
+
+
+> zorya-classifier@0.1.0 boundaries
+> node scripts/boundaries.mjs
+
+Boundaries and import cycles passed (17 modules)
+
+> zorya-classifier@0.1.0 dead-code:ci
+> knip --no-progress
+
+Configuration hints (2)
+Remove entry: eslint.config.mjs
+Remove project: eslint.config.mjs
+
+> zorya-classifier@0.1.0 test
+> node --import tsx --test --test-concurrency=2 test/*.test.ts
+
+✔ unknown and in-flight costs are summed, and duplicate settlement is rejected (0.66639ms)
+✔ reported cost overruns halt subsequent admissions and unknown errors retain reservations (0.216476ms)
+✔ zero-priced fixture requests still respect physical call admission cap (0.125403ms)
+✔ runner stops before native baseline and next cases after receipt-only evidence loss (72.938405ms)
+✔ controlled fixture neighbor cannot run under a live evidence label (0.318693ms)
+✔ classification keeps failures and operational unknown distinct from the semantic unknown class (1.416813ms)
+✔ all failed classifications produce zero accuracy and F1, not an empty successful denominator (0.260477ms)
+✔ macro F1 includes a declared class with no examples using zero-division zero (0.144784ms)
+✔ empty classification runs report null aggregate rates and preserve explicit zero confusion (0.131838ms)
+✔ class IDs cannot collide with evaluator buckets or object prototypes (0.172446ms)
+✔ classification rejects missing/duplicate taxonomy and unrecognized labels (0.383222ms)
+✔ latencies use nearest rank without mutating samples and include the slow failure tail (0.211551ms)
+✔ localization scores edge hits, any GT instance, distance, negatives and failed attempts separately (12.287563ms)
+✔ localization uses region center, not any region overlap or enclosing union of GT boxes (0.562178ms)
+✔ all image corners and right/bottom box boundaries count as closed edges (0.300795ms)
+✔ empty and all-failed localization runs do not fabricate zero distances or successes (0.172762ms)
+✔ localization rejects absent verification, missing GT, invalid geometry and invalid failure latency (0.281825ms)
+✔ native parity compares exact classes, raw probability differences and all-attempt denominator (0.344023ms)
+✔ unmatched probability keys are unavailable rather than implicitly filled with zeros (0.121188ms)
+✔ parity preserves empty/failure cases and rejects malformed probability measurements (0.137416ms)
+✔ prefix ablation resolves recorded labels to canonical source pixels, never generated coordinates (1.903325ms)
+✔ prefix presence is independent of spatial Choice, and absence at fine union is not global absence (0.503887ms)
+✔ baseline maxima use stable supplied order for equal probabilities and validate native winner (0.319196ms)
+✔ live CLI requires three explicit budget/admission flags; offline requires explicit artifact directory (1.438702ms)
+✔ recording retains failed attempt and unknown-cost reservation without retry or credential logs (12.984346ms)
+✔ budget rejection is recorded and never dispatches another provider attempt (2.77937ms)
+✔ late provider result is joined and recorded but cannot become a successful outcome (2.923929ms)
+✔ fixture response is content-blind and identical captured request baseline retains byte hash (4.717211ms)
+✔ receipt persistence failure is fatal and retains the primary provider failure and budget (3.218634ms)
+✔ late timeout plus receipt failure retains an independent terminal evidence marker (2.561243ms)
+✔ annotated rounded native distribution keeps raw values and stable ties; unannotated mass is rejected (0.180764ms)
+✔ prefix ablation accepts only explicitly annotated rounded scores without altering canonical geometry (0.211691ms)
+✔ architecture gate accepts ports and rejects reversed dependencies cross-platform (514.478991ms)
+✔ architecture gate rejects private, escaped and cyclic imports (743.341483ms)
+✔ classify validates before dispatch and preserves unknown cost (3.270214ms)
+✔ categorical maxima below 0.8 are accepted; exact ties use supplied order (0.52572ms)
+✔ contradictory native choice and missing probability are errors (0.436812ms)
+✔ one absolute deadline cancels provider and does not retry (15.069897ms)
+✔ cancelled queued operation never dispatches; physical work retains capacity (1.399111ms)
+✔ full queue fails with CAPACITY, then returns to idle (2.068873ms)
+✔ caller bytes are copied before queue awaits, including Buffer views (7.379434ms)
+✔ image limits and malformed media reject without provider calls (0.624833ms)
+✔ 1000 operation/dispose cycles leave no owned timeout handles (43.35027ms)
+✔ already aborted signal dispatches nothing (0.250312ms)
+✔ full queue rejects before reading or copying another image (2.817333ms)
+✔ rounded Jev mass needs explicit annotation and preserves raw categorical ties (0.408105ms)
+✔ leaked global/node intervals are rejected including aliases (18.494244ms)
+✔ stored interval cleanup and shadowed names are accepted (4.394401ms)
+✔ full locator has three calls with direct image parts and bounded candidate sets (259.903884ms)
+✔ global absence is one call, local absence does not claim global absence (256.98777ms)
+✔ none and presence contradiction remains unknown (236.465311ms)
+✔ 0.5 independent presence abstains, not a categorical threshold (35.953414ms)
+✔ ROI bounds checked before provider; tiny ROI reports layout limit (5.144767ms)
+✔ candidate refinement is opt-in, bounded and never substitutes area (448.485079ms)
+✔ source grids partition odd and narrow regions without edge gaps (0.554381ms)
+✔ beam keeps padded source geometry, canonical dedupe and at most196 cells (0.598029ms)
+✔ presentation preserves aspect ratio with padding and stays in Jev limits (5.677856ms)
+✔ adapter pins route, provider and no-fallback and preserves unknown metadata (20.146921ms)
+✔ accepts exact model identity typesafe/jev-1.14 and real zero cost (0.611645ms)
+✔ accepts exact model identity typesafe/jev-1.14-20261014 and real zero cost (1.03887ms)
+✔ uses header request ID only when response ID is unknown (0.568007ms)
+✔ rejects untrusted response metadata model {"model":"typesafe/jev-1.13-20260917"} (1.304323ms)
+✔ rejects untrusted response metadata model {"model":"typesafe/jev-1.14-untrusted"} (2.103559ms)
+✔ rejects untrusted response metadata model {"model":"typesafe/jev-1.14-20260230"} (0.431599ms)
+✔ rejects untrusted response metadata provider {"provider":"other-provider"} (0.378205ms)
+✔ rejects untrusted response metadata usage {"usage":{"cost":-1}} (0.633889ms)
+✔ rejects untrusted response metadata unexpected {"unexpected":"secret-token-do-not-return"} (0.386309ms)
+✔ sends direct top-level png image state unchanged (6.429089ms)
+✔ sends direct top-level jpeg image state unchanged (2.564998ms)
+✔ rejects unprepared image dimensions 279x280 (2.041021ms)
+✔ rejects unprepared image dimensions 280x279 (2.087383ms)
+✔ rejects unprepared image dimensions 1569x280 (4.474847ms)
+✔ rejects unprepared image dimensions 280x1569 (3.740591ms)
+✔ accepts maximum prepared image side (5.396898ms)
+✔ rejects unsupported image representation https://example.com/private.png (0.265924ms)
+✔ rejects unsupported image representation data:audio/wav;base64,AAAA (0.195845ms)
+✔ rejects unsupported image representation data:image/png;base64,AAAA (0.12849ms)
+✔ rejects unsupported image representation data:image/jpeg;base64,!!!! (0.154082ms)
+✔ rejects unsupported image representation data:image/jpeg;base64,/9j/= (0.170413ms)
+✔ rejects nested raw audio and extra request fields before transport (0.400849ms)
+✔ Choice enforces 2..255 options: 1 (0.271837ms)
+✔ Choice enforces 2..255 options: 2 (0.322199ms)
+✔ Choice enforces 2..255 options: 255 (2.553031ms)
+✔ Choice enforces 2..255 options: 256 (0.473632ms)
+✔ preserves categorical ties, raw scores and independent confidence (0.330622ms)
+✔ supports bounded Noul without conflating it with Choice (0.291917ms)
+✔ rejects mismatched question/answer distribution {} (0.254266ms)
+✔ rejects mismatched question/answer distribution {"other":{"type":"choice","choice":"a","probabilities":{"a":0.6,"b":0.4}}} (0.285431ms)
+✔ rejects mismatched question/answer distribution {"target":{"type":"choice","choice":"a","probabilities":{"a":0.6,"b":0.4}},"other":{"type":"choice","choice":"a","probabilities":{"a":0.6,"b":0.4}}} (0.303117ms)
+✔ rejects mismatched question/answer distribution {"target":{"type":"noul","noul":0.8}} (0.256523ms)
+✔ rejects mismatched question/answer distribution {"target":{"type":"choice","choice":"c","probabilities":{"a":0.6,"b":0.4}}} (0.219348ms)
+✔ rejects mismatched question/answer distribution {"target":{"type":"choice","choice":"b","probabilities":{"a":0.6,"b":0.4}}} (0.25795ms)
+✔ rejects mismatched question/answer distribution {"target":{"type":"choice","choice":"a","probabilities":{"a":1}}} (0.303071ms)
+✔ rejects mismatched question/answer distribution {"target":{"type":"choice","choice":"a","probabilities":{"a":0.6,"b":0.4,"c":0}}} (0.204174ms)
+✔ rejects mismatched question/answer distribution {"target":{"type":"choice","choice":"a","probabilities":{"a":1.1,"b":-0.1}}} (0.2671ms)
+✔ takes immutable request snapshot before any await (0.29756ms)
+✔ lower request cap rejects before fetch, including encoded JSON overhead (0.152823ms)
+✔ factory rejects unsafe options {"maxRequestBytes":33554433} (0.137247ms)
+✔ factory rejects unsafe options {"maxResponseBytes":1048577} (0.069426ms)
+✔ factory rejects unsafe options {"maxRequestBytes":0} (0.059932ms)
+✔ factory rejects unsafe options {"maxResponseBytes":-1} (0.061106ms)
+✔ factory rejects unsafe options {"apiKey":"bad\nkey"} (0.074425ms)
+✔ factory rejects unsafe options {"endpoint":"https://unapproved.example"} (0.060902ms)
+✔ HTTP 401 is sanitized, body cancelled, no retry (0.956054ms)
+✔ HTTP 403 is sanitized, body cancelled, no retry (0.301551ms)
+✔ HTTP 413 is sanitized, body cancelled, no retry (0.192749ms)
+✔ HTTP 415 is sanitized, body cancelled, no retry (0.278226ms)
+✔ HTTP 429 is sanitized, body cancelled, no retry (0.206133ms)
+✔ HTTP 500 is sanitized, body cancelled, no retry (0.17819ms)
+✔ HTTP 302 is sanitized, body cancelled, no retry (0.167717ms)
+✔ declared response cap cancels before reading provider body (0.305711ms)
+✔ actual streamed response cap works without or with false Content-Length (0.824344ms)
+✔ malformed successful response is rejected body (1.283253ms)
+✔ malformed successful response is rejected body (0.166169ms)
+✔ malformed successful response is rejected empty (0.142321ms)
+✔ malformed successful response is rejected body (0.127604ms)
+✔ fetch and stream exceptions cannot leak messages or details, including forged ClassifierError (0.468662ms)
+✔ pre-aborted signal prevents transport and sanitizes cancellation reason (0.147096ms)
+✔ retains TIMEOUT code while sanitizing caller reason (0.137947ms)
+✔ late fetch response after abort is cancelled/released and never returns success (0.320195ms)
+✔ abort during body read cancels once, joins physical cleanup, releases reader/listener (0.366724ms)
+✔ cancellation rejection is sanitized and visible without retaining reader (0.249836ms)
+✔ rejects prototype setter keys rather than silently dropping requested options (0.271893ms)
+✔ rejects prototype setter keys in provider answer rather than stripping them (0.218322ms)
+✔ abort racing with body EOF cannot publish successful answer (0.354991ms)
+✔ copies fragmented reused stream buffers without retaining per-chunk allocations (0.767713ms)
+✔ complete request byte cap includes criteria, instructions, keys and JSON escaping (0.55933ms)
+✔ accepts exact serialized request cap and rejects one byte below it (2.757233ms)
+✔ rejects oversized criteria before allocating full serialized body (120.128368ms)
+✔ rejects accessors, cycles and overlong collections before schema cloning (0.455123ms)
+✔ rejects unnormalized choice distribution without renormalizing it (0.34393ms)
+✔ replays actual Jev 1.14 wire: screen-locator-jev-latency-2026-10-16 (0.786599ms)
+✔ replays actual Jev 1.14 wire: screen-locator-jev-1-14-2026-10-16 (0.379678ms)
+✔ non-enumerable declared fields and proxies cannot bypass preflight byte accounting (0.243354ms)
+✔ owned snapshot never invokes caller array iterator: own-method (0.51811ms)
+✔ owned snapshot never invokes caller array iterator: own-getter (0.269403ms)
+✔ owned snapshot never invokes caller array iterator: inherited-getter (0.275344ms)
+✔ choice mass remains strict without explicit rounding provenance (0.119611ms)
+✔ explicit rounding validates clipped feasible mass, not a fixed sum tolerance (0.286918ms)
+✔ adapter annotates only rounded mass and preserves raw native ties (0.484918ms)
+✔ rounded transport still rejects malformed distribution {"type":"choice","choice":"a","probabilities":{"a":0.505,"b":0.485}} (0.263481ms)
+✔ rounded transport still rejects malformed distribution {"type":"choice","choice":"b","probabilities":{"a":0.5,"b":0.49}} (0.256775ms)
+✔ rounded transport still rejects malformed distribution {"type":"choice","choice":"a","probabilities":{"a":0.6,"b":0.4},"probabilityRoundingStep":0.01} (0.24355ms)
+✔ replays actual 153-option Jev 1.14 rounded response without normalizing scores (1.984932ms)
+✔ secret filename checks reject private inputs and keep exact example templates (401.591748ms)
+✔ known provider prefixes, private keys, database credentials and entropy are detected without excerpts (207.454083ms)
+✔ empty templates and obvious low-entropy samples do not trigger credential heuristics (47.424592ms)
+✔ staged secret remains rejected when the working file has already been cleaned (102.660539ms)
+✔ clean staged content is independent of a later working-tree secret (99.823983ms)
+✔ worktree scan includes untracked files and skips ignored env without opening it (88.844787ms)
+✔ forced-staged ignored env is rejected by filename, without content findings (55.885702ms)
+✔ example templates are not a content allowlist (67.146768ms)
+✔ binary content is explicitly counted, while private binary filenames are rejected (114.119753ms)
+✔ oversized text fails closed instead of silently omitting the file (50.151687ms)
+✔ staged symlink blobs are rejected without following a filesystem target (59.643092ms)
+✔ repository gitignore protects local credentials, caches and releases but not examples or source (91.208128ms)
+✔ git failures expose only a sanitized rule and no Git diagnostic text (29.01556ms)
+✔ unquoted environment references are code, while literal quoted credentials remain scanned (71.176543ms)
+✔ POSIX tracked FIFO replacement cannot block the scanner (56.663819ms)
+✔ leading BOM in an untracked Git filename cannot alias a clean tracked file (54.265448ms)
+ℹ tests 160
+ℹ suites 0
+ℹ pass 160
+ℹ fail 0
+ℹ cancelled 0
+ℹ skipped 0
+ℹ todo 0
+ℹ duration_ms 3914.369024
+
+> zorya-classifier@0.1.0 build
+> tsc -p tsconfig.build.json
+

package/docs/evidence/live-release-smoke-manifest.json

@@ -0,0 +1,82 @@
+{
+  "version": 1,
+  "startedAt": "2026-10-17T21:54:57.525Z",
+  "mode": "live",
+  "evidence": "real-provider; route and metadata recorded per call",
+  "options": {
+    "mode": "live",
+    "allowPaid": true,
+    "maxCalls": 14,
+    "maxCostUsd": 0.1,
+    "corpusDirectory": "benchmarks/corpus",
+    "outputDirectory": ".run/live-release-smoke",
+    "splits": ["dev"],
+    "limit": 2
+  },
+  "node": "v24.12.0",
+  "platform": "linux",
+  "model": "typesafe/jev-1.14",
+  "provider": "typesafe",
+  "fallbacks": false,
+  "operationTimeoutMs": 45000,
+  "retries": 0,
+  "maximumPlannedCalls": 14,
+  "planned": {
+    "photos": 2,
+    "texts": 2
+  },
+  "corpus": {
+    "corpus-manifest.json": "00f4459996db0f61a38235adaf512e1bfc52e02e1d599bb590b62a98217e4cdb",
+    "photo-manifest.json": "a99f47d4e1c064581c6cc9123b58a5c81b196d15227963bf015041544f713aa3",
+    "text/text-manifest.json": "5e74c8c13cd1f1f8a2e2ecc22e90e6c0c83ebf5f9407dfed5b87204d3b8831ac",
+    "text/inputs/policy.json": "9db3010c97fa056fbb4a43ea51a5c973bd75b9d21f8d07f2c15d6d176e712b8a",
+    "inputs/dev.json": "0c443231956e9578afa43d4bf884cc9741594ae69ac5f957657b9c5a24422bb6",
+    "ground-truth/dev.json": "c618e1a950ae89e0c9ced0c2b1ca249ea7adc891f61c0579468c479613ea4b26",
+    "text/inputs/dev.json": "e7d2bc07559cba70494d6e4672cc3e1113c070aba8b2694ee45e76ebcbc04fcc",
+    "text/ground-truth/dev.json": "4827650b560f20b32b0366dea14947681d29ef9fefb5a059301ba5abf5c031f0"
+  },
+  "sourceHashes": {
+    "src/app/create-classifier.ts": "55bfcf78b8c12c616d1144371541ec4eb1e56a6a4790700b366e41dbdf896491",
+    "src/cli.ts": "2a28730b48e07437549421513b9db3a23c1c1fa907831ae02c034db17aee8aec",
+    "src/entities/contracts.ts": "d48c488c986a136ddeb0e843c8ba85cf5dab4838f6f49b0e4010d217defe2631",
+    "src/entities/errors.ts": "1c92c94c9ed6917953ac5192bfd2d8896cdb6fe13b7d05593573b7f2103330b0",
+    "src/features/choice-beam-layout.ts": "91ec1be72e31e19c5e9f1b722b1ad3810e07b78bfe2e8126ae7b00869a433c35",
+    "src/features/choice.ts": "307858ff2ee0bc142eef953123ed35b80756a9d4118d306025db4365b4f9b31f",
+    "src/features/prompts.ts": "df8090d057e71641b628dcab8b9d2fd59aeb24c4f93af1cd042f8aab5caf57dc",
+    "src/features/uniform-grid.ts": "8cee38aee52ba2c4ac2eaed909d5f53cc0e5aa2c04fe87585c66b29573e21c4d",
+    "src/index.ts": "1cd73c74a874adf1a4a3b6edb48e230f54c671098b9a3d51bb3a01a0ad2866db",
+    "src/infrastructure/openrouter.ts": "b9d4e5686c2c0c3c3463154a4164b0c5e9034582d59abb71e97794ec6a532b86",
+    "src/infrastructure/raster.ts": "8217346f671b957540f3c01a7eb3d8ef7d0bfb6a5e18a47c96f0ccb08f4bcb08",
+    "src/openrouter.ts": "0fc784e949774d5517aee4c5b019afc0a9c6e488988be28ff1c1a8c272635232",
+    "src/processes/classify.ts": "fb2a13aec8f02853f5ba353babf9588cce7aaebecb9f4ae6339a2f40897d736f",
+    "src/processes/locate.ts": "bf54bef7a899f0aa422b63da9ff577102dd6fa105ad8aff718b4168fa9f8ea88",
+    "src/shared/choice-mass.ts": "7f645998d9cd2688ca425b4765ec0bf7d125e932f085f9e84599b4c3b97743a3",
+    "src/shared/gate.ts": "97e4e32749dc9d3a9def6faaea2566b5d0ed831ba4451999740c719fafa8e576",
+    "src/shared/scope.ts": "cc557e9fd83a838f5366c64d026859db119491d927dfb3a84ef4320b2f0b0beb",
+    "benchmarks/ablations.ts": "d4450a1c962bd96f89af3499fd0d5da52bb20eb2c6fdf8c79d2e6057d0d6325b",
+    "benchmarks/budget.ts": "538a7e89ca643d2c26ea7e13a4d213ee019a2df865fac6573c31e9fb1ea63166",
+    "benchmarks/corpus.ts": "0af52e2f896e74e8e9ab0937e34862cea5396ac0649f896299631160314848b4",
+    "benchmarks/metrics.ts": "8a836b35152afd24c24095669d3ca802fd57eafb452ebbd7e28dbb56853f9e07",
+    "benchmarks/overlay.ts": "9eced23935c5428d657df747d0f239c6fc8a61ef2f1582d88a86e08f71ddfbf5",
+    "benchmarks/recording.ts": "fd1913f50027c4f78ebb5daddfd9b563455b145e525ddd4dda18b453670358af",
+    "benchmarks/run.ts": "5ff28c02fd9bc6dc03aaf5934e7e28300f958696a2fb1fffca44edabf748ebfa",
+    "benchmarks/runner.ts": "fd95d5b720e11f4f22bd39e5044b456fcd3c42fffd6b4f308ee98bd99695558f",
+    "package.json": "0e7cb54459efb7f326d4d13a6416a08b29eeb626d41f962ce630c86f42c94acd",
+    "package-lock.json": "879d16df46b6fd0c95c837b9683422d7ff76d6a3f61790cf62288fcf010d4384",
+    "tsconfig.json": "3411e388c49ac70c5c50d0c2d85ee5bcc7270a16e8d3fb4a217c99212a783d0e"
+  },
+  "limitations": [
+    "Public photos may overlap training data; heldout is only held out from local development.",
+    "Text fixtures and labels are assistant-authored, not independent human annotations.",
+    "Locate returns one region; center-inside-any-box is not all-object recall or segmentation accuracy.",
+    "Prefix ablations reuse exact stage answers; they lack third-stage confirmation.",
+    "Request preparation and artifact writes are included in measured operation latency.",
+    "Known cost plus $0.02 per incomplete or unknown-cost provider invocation is an admission estimate, not a provider-enforced spending limit.",
+    "Public DecisionResult contains cost and route metadata; raw HTTP token usage is unavailable.",
+    "Photographs do not validate gaming UI, maps, OCR, or temporal tracking."
+  ],
+  "publicExport": {
+    "rawManifestSha256": "456dccfd3af1593d1a420453a5f5a5cbae47506542889156da81c83db3fd4487",
+    "removed": "Host-specific absolute output directory replaced by repository-relative path."
+  }
+}

package/docs/evidence/live-release-smoke-provenance.json

@@ -0,0 +1,9 @@
+{
+  "rawReportSha256": "53ded957128475306bbb8ab4c7a5a40d37382ee48cc51fe1df5301bd2c21d439",
+  "rawManifestSha256": "456dccfd3af1593d1a420453a5f5a5cbae47506542889156da81c83db3fd4487",
+  "status": "complete",
+  "physicalHttpCalls": 14,
+  "knownCostUsd": 0.004080510000000001,
+  "unknownCostCalls": 0,
+  "scope": "Final runtime smoke, 2 development photos and 2 development texts; not a quality re-evaluation"
+}