@axlsdk/eval 0.24.0 → 0.24.2

package/README.md

@@ -414,11 +414,13 @@ actually measured. The value is still validated — a non-negative finite number
 dropped with a `console.warn` — so `{ cost: 'free' }` can poison nothing.
 
 **Additive item metadata.** Returning `{ output, metadata: { category: 'billing' } }`
-preserves tracked models, tokens, agent calls, and workflow attribution, with or without
-trace capture. User keys override tracked defaults; nested objects are replaced wholesale.
-An explicit `models` or `workflows` list without corresponding call counts removes the
-inherited count map so run totals use the list fallback. Supply both list and counts for
-exact custom accounting. See the [metadata contract](../../docs/api-reference.md#evalitem).
+adds application keys beside measured models, tokens, agent calls, and workflow
+attribution. The reserved measurement keys (`models`, `modelCallCounts`,
+`tokens`, `agentCalls`, `workflows`, `workflowCallCounts`) are kept under
+`item.callerReport.metadata` if returned by the callback; they do not override
+the runtime's measurement. Model diagnostics and timing follow the accounting
+scope across runtimes and compatible ESM/CJS copies. Workflow names and captured
+traces still depend on the runtime hosting `trackOutcome`.
 
 ### Studio
 
@@ -717,6 +719,13 @@ could not price is reported as unknown (`unpriced`, with `accounting.reasons`) r
 than as zero — and because unknown spend cannot be enforced against, it does not consume
 a budget either.
 
+Compatible ESM/CJS or duplicate Axl loads in the same JavaScript realm share
+the accounting scope, budget controller, and request capture channel. A workflow
+using another compatible copy is measured and charged to this run. If a loaded
+copy cannot join the active accounting protocol, paid work is refused and the
+run is marked `incomplete` / `uninstrumented`; update older Axl copies to get
+this guarantee. See [cross-copy accounting](../../docs/observability.md#completeness-and-reasons).
+
 ### Captured requests (opt-in)
 
 When a score looks wrong, the next question is what the model was actually

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@axlsdk/eval",
-  "version": "0.24.0",
+  "version": "0.24.2",
   "description": "Evaluation framework for Axl agentic workflows",
   "type": "module",
   "main": "./dist/index.cjs",
@@ -25,7 +25,7 @@
     "dist"
   ],
   "dependencies": {
-    "@axlsdk/axl": "0.24.0"
+    "@axlsdk/axl": "0.24.2"
   },
   "peerDependencies": {
     "zod": "^4.0.0",