101 → 88
-12.9%Worse
Higher is better. The candidate materially reduces revenue per run.
ARRM Public Proof · Release Regression Case #1 · Reproducible repository fixture
A release can work — and still regress economically.
This is the canonical public ARRM verification artifact for one deterministic economic release-control case. It compares a baseline agent release with a candidate release across the same checkout scenario. Measurements are caller-provided; ARRM applies explicit rules to the evidence before production progression.
Canonical comparison
Same scenario. Worse economics.
Fixture scenario: standard_checkout. Baseline release: agent-checkout-v1.4.0. Candidate release: agent-checkout-v1.5.0.
0.425 → 0.365
-14.1%Worse
Higher is better. The candidate materially reduces conversion.
0.18 → 0.21
+16.7%Worse
Lower is better. The candidate materially increases cost per task.
Why ARRM blocks
The release decision is explicit, reproducible and machine-actionable.
ARRM evaluates configured metric direction and WARN/FAIL thresholds against baseline and candidate evidence. Here, adverse changes in revenue, conversion and cost for standard_checkout produce the deterministic aggregate result FAIL, mapped to BLOCK.
Control boundary: no LLM is used in the decision path. ARRM does not invent measurements, infer business intent or automatically collect production data. It gates caller-provided evidence using explicit deterministic rules. PASS maps to PROCEED, WARN to REVIEW and FAIL to BLOCK.
What this proves
Technical validity is not the same as economic release safety.
A candidate may remain technically functional while materially worsening revenue-sensitive outcomes. ARRM exists to convert comparable caller-provided baseline/candidate measurements into an explicit economic release-control decision before production exposure.
Claim boundary: this is a reproducible ZEODEC/ARRM repository fixture and public verification artifact. It is not a customer case study, realized revenue claim, savings claim, production guarantee or evidence of general availability. ARRM remains Early Access.
Try ARRM
Run the same gate against one real release.
ARRM Early Access includes the deterministic release-gate package, canonical GitHub Actions path, machine-readable report artifact and written setup support if needed.