AIN-288 · one-command delta dry-run + CI

.github/workflows/delta-dry-run.yml

@@ -0,0 +1,18 @@
+name: delta-dry-run
+on:
+  push:
+    paths: ["benchmark/**", "eval/**", "datasets/**", ".github/workflows/delta-dry-run.yml"]
+  pull_request:
+    paths: ["benchmark/**", "eval/**", "datasets/**"]
+jobs:
+  delta:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-python@v5
+        with: { python-version: "3.12" }
+      - run: pip install -r benchmark/requirements.txt
+      - name: AIN-288 synthetic delta dry-run (must exit 0, deterministic seed=7)
+        run: ./eval/run_delta.sh 7
+      - name: assert delta table emitted
+        run: test -s preprint/results.md && grep -q 'ainfera_learned' preprint/results.md

eval/run_delta.sh

@@ -0,0 +1,15 @@
+#!/usr/bin/env bash
+# AIN-288 · one-command reproducible delta harness (synthetic dry-run).
+#
+#   ./eval/run_delta.sh [SEED]      # default seed 7 (CRN — deterministic)
+#
+# Generates the CC0 synthetic corpus then computes the routed-vs-baselines
+# delta table (routed/ainfera_learned vs single_best / cheapest / round_robin
+# (random) / agent_baseline (agent-default) / ainfera_static / oracle) and
+# writes preprint/results.md. Until a real judged corpus exists this is a
+# DRY-RUN (labeled synthetic, not a published claim).
+set -euo pipefail
+cd "$(dirname "$0")/.."
+SEED="${1:-7}"
+python datasets/generate.py --seed "$SEED"
+python eval/delta.py