@mlbottleneck/engine 0.4.0 → 0.5.0

package/README.md

@@ -78,6 +78,10 @@ const { ceiling } = engine.predict({ model: 'qwen3.6_35b_a3b', hardware: 'AMD St
              expectedTokensPerSecond,      // engine rate x peer correction (stock software)
              engineTokensPerSecond, correctionFactor, confidence, peers, verifiedPeers },
   memory:  { modelSizeGB, residentWeightsGB, kvCacheGB, availableGB },
+  measured: { nearest,                     // closest stock measurement on this model + hardware template (community gold run or lab stock row), or null
+              labTuned },                  // closest tuned neural.download lab run on the same machine (same template and device count), or null
+           // each: { tokensPerSecond, origin: 'community'|'lab', stack, model, hardware, deviceCount, runtime, quantization,
+           //         depthTokens, speculation, sameSetup, url, note }
   bottleneck: 'memory' | 'compute' | 'runtime' | 'coordination' | ...,   // devices[].coreBinding adds 'attention' for deep contexts
   power:   { watts, tdpWatts, costPerDay, costPer1KTokens },
   devices: [{ name, template, residentWeightGB, kvCacheGB, hasOverflow, overflowMode,
@@ -88,7 +92,9 @@ const { ceiling } = engine.predict({ model: 'qwen3.6_35b_a3b', hardware: 'AMD St
 }
 ```
 
-Numbers are *planning estimates*: the engine is calibrated so that the median community run lands on its prediction and ~85% land within 1.5×. Show users the ceiling ladder (`physical` → `optimized` → `expected`) rather than a single number when you can.
+Numbers are *planning estimates*: the engine is calibrated so that the median community run lands on its prediction and ~85% land within 1.5×. Show users the ceiling ladder (`physical` → `optimized` → `expected`) rather than a single number when you can, and `measured.nearest` / `measured.labTuned` beside it when they exist — they are real runs, with links.
+
+The evidence snapshot carries two collections: `goldCases` (community runs, the calibration corpus) and `labCases` (the author's [neural.download](https://neural.download/) Intel Arc Pro lab: `stock`, `lab-baseline`, and `tuned` rows). Lab rows never calibrate the engine; they only surface as measured references.
 
 ## Other methods
 
@@ -112,7 +118,7 @@ https://mlbottleneck.com/?model=qwen3.8_27b&hardware=Intel%20Arc%20Pro%20B70&cou
 
 - `package.json` `version` is the SDK version; `engine.version` reports it.
 - `npm test` rebuilds `dist/` from `engine.js` + `sdk/api.js` (and stamps the page's `engine.js?v=<hash>` cache key), so the committed bundle is always the one that passed the suite.
-- Bump the version when the engine, catalogs, or API change; `.github/workflows/release-sdk.yml` then publishes a `sdk-v<version>` GitHub release with the bundles, types, evidence snapshot, and checksums, and `.github/workflows/publish-npm.yml` publishes `dist/` to npm as `@mlbottleneck/engine` (name set by `sdkName` in `package.json`; needs the `NPM_TOKEN` repository secret and a LICENSE). The weekly evidence refresh updates `dist/localmaxxing-snapshot.json` in place without a release.
+- Bump the version when the engine, catalogs, or API change; `.github/workflows/release-sdk.yml` then publishes a `sdk-v<version>` GitHub release with the bundles, types, evidence snapshot, and checksums, and `.github/workflows/publish-npm.yml` publishes `dist/` to npm as `@mlbottleneck/engine` (name set by `sdkName` in `package.json`). It authenticates with npm trusted publishing (OIDC: the package's Trusted Publisher on npmjs.com points at `steveseguin/ml-bottleneck`, workflow `publish-npm.yml`), falling back to an `NPM_TOKEN` secret; a 2FA-gated token fails with `EOTP`, in which case publish by hand from `dist/` with `npm publish --access public --otp=<code>`. The weekly evidence refresh updates `dist/localmaxxing-snapshot.json` in place without a release.
 
 ## Limits