AI Agent Spend & Anomaly Guardrail Lab

Unconstrained Log Spend

$1,149.26

480 trace log events parsed

Spend Saved with Guardrails

$1,099.26

95.6% projected loss mitigated

Circuit Breaker Status

TRIGGERED Call #3

Terminated at threshold limit ($50.00 cap)

75-Scenario Policy Suite

75 Passed / 0 Failed

100% Policy Pass Rate VERIFIED

Execution Trace / Invoice Ingestion

Tesseract OCR Ready

Drop execution log file or billing screenshot/OCR receipt

Supports .txt, .log, JSON, or images (PNG/JPG)

Trace Stream Sample 480 Events Parsed

Guardrail Policy & Circuit Breaker Rule Engine

Active Enforcement

Trace Budget Cap ($) $50.00

Circuit breaker triggers immediately upon reaching cap.

Max Loop Depth 15 calls

Halts recursive tool searches exceeding depth threshold.

Token Cap per Request 25,000 tok

Prevents massive context inflation & prompt injection blowouts.

Breaker Action / Model Fallback

Action taken when anomaly or threshold breached.

Configured rules will automatically evaluate against 75 test scenarios.

Trace Timeline & Runaway Anomaly Visualizer

Cumulative spend projection vs. Circuit Breaker cut-off point across 480 trace calls

Unconstrained Trace

Guardrail Enforced

Failing Anomaly Call Index Call #3 (Loop Detect)

Anomaly Cause Recursive tool search loop step 12 (478 redundant queries follow)

Spend Intercept Point $50.00 max cap reached on Call #3

Automated Guardrail Policy Verification Suite (75 Scenarios)

Verifies breaker reaction timing, token ceiling safety, model fallback reliability & loop limits.

Canonical Verification Proof Surface

Total Spend Detected: $1,149.26

Guardrail Dollars Saved: $1,099.26

Test Suite Status: 75 Passed / 0 Failed

Breaker Triggered: true

A fixture guardrail intercepts a token surge before the spend cap

Your browser does not support video.

Read the explanation

The local fixture has four hundred eighty calls with assigned spend totaling eleven hundred forty nine point two six dollars. Three calls are explicitly described and four hundred seventy seven are generated to fill the remaining spend. At half a pixel per count explicit calls measure one point five, generated two hundred thirty eight point five and total two hundred forty. No uploaded invoice or real provider run is required for these figures; they are constructed browser data. Text upload merely saves raw text and image OCR needs a separate dependency, so parsed event labels do not prove the uploaded content changed this trace. The first three assigned costs are zero point zero three one five, zero point one one five eight and one point four two dollars. Their cumulative total is one point five six seven three. The default spend cap is fifty dollars, loop cap fifteen and input token cap twenty five thousand. Call three has twenty eight thousand four hundred input tokens, exceeding that token cap by thirty four hundred even though spend and loop depth remain below their caps. At one pixel per hundred tokens the request bar is two hundred eighty four, cap two hundred fifty and excess thirty four. The earliest matching condition triggers; the UI cap label alone does not identify which rule actually fired. Intercepting after the third assigned charge leaves approximately one point five seven spent and eleven hundred forty seven point six nine dollars labeled saved. At one fifth pixel per dollar spent measures about zero point three one four, savings two hundred twenty nine point five three eight and original total two hundred twenty nine point eight five two. These are retrospective fixture differences, not an enforcement bridge preventing real charges. Seventy five scenario cards are generated with passed true and render PASSED unconditionally, rather than executing seventy five tests. Native export can work before missing D3 fails chart rendering; preserve that distinction and never claim the cards prove a tested guardrail system.