@irisrun/evals 0.1.0 → 0.2.0

package/README.md

@@ -0,0 +1,27 @@
+# @irisrun/evals
+
+**Reproducible evals — because determinism makes scoring repeatable.** Run the
+same scenario and it replays **byte-identically** from its journal, so a score is
+a fact about the run, not a roll of the dice. The arbiter is reproducibility, not
+taste: the same case + scorer re-runs identically; a swapped tactic scores
+differently, but reproducibly.
+
+## What it is
+
+`runEval` runs a deterministic scenario (a fresh store + scripted performers per
+run) on the core `runTurn`, then scores the recorded session via
+`@irisrun/inspect`; `runSuite` runs a set; `reproduce` re-runs a case to confirm
+byte-identical replay. This is faithful record-replay of captured effects — not a
+claim that a live model is deterministic. Depends on `@irisrun/core` +
+`@irisrun/inspect`.
+
+## Use it
+
+```sh
+iris eval ./evals/suite.mjs    # reproducible scenario scoring
+```
+
+See **[docs/Audit & reproducible evals](../../docs/audit-and-evals.md)**.
+
+---
+Part of [Iris](../../README.md) — own, portable, verifiable state.

package/dist/evals.js

@@ -1,4 +1,4 @@
-// The reproducible-eval arbiter (spec 03 §7): "reproducible evals are the arbiter,
+// The reproducible-eval arbiter: "reproducible evals are the arbiter,
 // not editorial taste." An EvalCase is a DETERMINISTIC scenario; a Scorer reads the
 // recorded session (via @irisrun/inspect) and the last turn outcome. runEval calls
 // case.build() on EVERY invocation, so it gets a FRESH store AND fresh performers

package/dist/reproduce.js

@@ -1,4 +1,4 @@
-// reproduce() (roadmap P2-8): makes "reproducible evals" an EXPLICIT, provable
+// reproduce(): makes "reproducible evals" an EXPLICIT, provable
 // feature, not just an implicit property. It runs an EvalCase N independent times
 // (each `case.build()` is a fresh store + performers, index reset to 0 — the EvalCase
 // contract) and proves byte-identical {score, status, FULL-journal digest} across

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@irisrun/evals",
-  "version": "0.1.0",
+  "version": "0.2.0",
   "type": "module",
   "description": "Iris reproducible-eval arbiter (spec 03 §7) — run a deterministic scenario (fresh store + scripted performers per run), then score the recorded session via @irisrun/inspect. The arbiter is reproducibility, not taste: the same case+scorer re-runs byte-identically; a swapped tactic scores differently but reproducibly. Runner is core runTurn; deps @irisrun/core + @irisrun/inspect.",
   "exports": {
@@ -11,8 +11,8 @@
     }
   },
   "dependencies": {
-    "@irisrun/core": "^0.1.0",
-    "@irisrun/inspect": "^0.1.0"
+    "@irisrun/core": "^0.2.0",
+    "@irisrun/inspect": "^0.2.0"
   },
   "license": "MIT",
   "engines": {