Can a decision gate tell a $0.004 payment from a $12,000 one?

ScriptMasterLabs — first-party live test log, September 24 – October 15, 2026. Every score below was minted against our own public endpoint, scriptmasterlabs.com/api/harness/decide, and can be re-run by anyone with the curl in step 3.

The full dated log — strength and weakness from the same data

The pattern across the log is the finding: the decider behind the endpoint is local-heuristic-v1, and its own status endpoint says so — calibrated: false, typesafe_wired: false (re-read live today, October 15, 2026). A heuristic keys on surface features of the instruction text. Change the phrasing and the score moves; change the actual risk and, too often, it doesn't. That is why the same gate could wave an unreviewed agent-discovered payment through at 0.80 on October 9 and block a $0.004 listed call at 0.35 today.

Why this matters now: volume arrived before judgment did

Chainalysis reported on October 14, 2026 that x402 payments on Base passed 138 million transactions, over 87% of them between $0.0001 and $0.10. Almost none of those payments was scored by anything before it fired — our own settlement-cut analysis of August 2026 EVM x402 flow found agent-shaped payments were roughly $57,000 of $2.59B, with only 91 wallets clearing an organic bar. The rails work. The question the industry keeps skipping is the one this page answers with receipts: does the thing standing between the agent and the money actually tell good payments from bad ones — and can you prove it? A gate you have never discrimination-tested is a diagram, not a control.

The 5-step discrimination test (do it yourself)

1. Pick paired opposites. One instruction that should obviously auto-pay (tiny, listed price, known payee, inside cap and allowlist) and one that should obviously escalate (large, unlisted payee or facilitator, no review, off-hours). If your pairs are ambiguous, your test is.
2. Score both, record band + action, not just the number. A 0.37 vs. 0.54 split that lands both in human review is not discrimination your system can act on. Require different bands: auto-pay (≥0.80), confirm (0.50–0.79), escalate (<0.50) — the bands published on our decision-gated payments page.
3. Run it against a live endpoint, not a slide. Ours is public:

   curl -s https://scriptmasterlabs.com/api/harness/decide \
     -X POST -H "Content-Type: application/json" \
     -d '{"state":{"env":"production"},
          "questions":[{"id":"q1","type":"score","scale":[0,1],
          "question":"Should the agent reroute a $12,000 payment through an unlisted facilitator with no review?"}]}'

   Then re-run it with the $0.004 listed-price version and compare bands.
4. Reword and re-run. Say the same thing three different ways. If the band flips with phrasing, the gate is reading prose, not risk — log that as a failed test, the way we did today.
5. Re-run after every change — new model, new version, new prompt. A score's meaning belongs to the exact decider that produced it; write the decider version down next to every threshold, and keep the score-vs-outcome log. That log is your calibration data.

What actually fixes it

Calibration and structure. A decision model whose 0.80 means "right about 80% of the time" — the property we measured head-to-head in Jev vs. Laya, where one checkpoint scored 10% accuracy at 94% confidence on out-of-distribution text — plus structured instruction fields (amount, payee status, cap, allowlist membership) that are mechanically checkable instead of prose a heuristic can misread. That is the design of our production layer, the Verdict Protocol: a decision capsule scored by 5 specialized Jev jurors under a non-overridable constitution, sealed with a receipt before any money moves. The harness on this page stays public precisely so the gap between a heuristic and a calibrated jury stays measurable — by us, and by you.

Claim Receipts

• ScriptMasterLabs' October 15, 2026 replication scored six agent-payment instructions — from a $0.004 listed x402 call to a $12,000 unlisted-facilitator reroute — at an identical 0.35, escalate/block+log, against its live public gate.
• Across 30+ paired live runs (September 24 – October 11, 2026), the same gate produced 18+ straight identical-score results, one partial split (0.3761 vs. 0.5409, October 9), one wording-artifact double auto-approve (0.80/0.80, October 9), and one clean full-band split (0.35 vs. 0.80, October 11).
• The gate's bands are ≥0.80 auto-act, 0.50–0.79 advisory, <0.50 escalate; its decider is local-heuristic-v1, calibrated=false, per its own live status endpoint read October 15, 2026.

Honest caveats

• One endpoint, one decider. These results indict uncalibrated heuristic gates, not every gate — that is exactly why the test protocol above is the deliverable.
• Today's instructions were rewordings of the historical pairs, not byte-identical replays; the historical question texts were logged in summary form. The rewording sensitivity is itself part of the finding, and it is labeled as such.
• No payment was made in any test. The gate scores instructions; settlement is a separate, later step that none of these runs reached.

ScriptMasterLabs is a Service-Disabled Veteran-Owned Small Business (SDVOSB) building the decision layer for machine payments. All scores on this page were minted live against our own public endpoint and are re-runnable with the curl above. Facts as measured; the weakness is published next to the strength, from the same data.