DETERMINISTIC_DEMO_FIXTURESeeded walkthrough of the BurnLens economic loop — not live telemetry.
What is BurnLens? An AI Economics control plane: observe spend, attribute it to outcomes, show how much of the number you can trust, recommend an explicit change, and verify whether the change saved money.
AI Spend
$1,842.40
9,124 requests
Cost / accepted outcome
$38.38
Provider Reconciliation
Not reconciled yet
| Repository / Workflow | Spend | Accepted | Cost / outcome | Coverage | Confidence |
|---|
| repo:checkout-service | $612.18 | 14 | $43.73 | 88% | estimated |
| repo:billing-api | $428.90 | 11 | $38.99 | 81% | estimated |
| support-triage | $301.44 | 16 | $18.84 | 64% | calculated |
| (untagged) | $499.88 | — | Not enough outcome data | 0% | calculated |
Over-specified model on short-output classify
support-triage uses gpt-5.6-sol for replies under 80 output tokens.
Projected monthly saving $86.40 · 312 requests · avg 54 output tokens · high confidence
History bloat on checkout-service agent sessions
Repeated tool transcripts are re-sent; scan rows have no prompt segments.
Projected monthly saving $41.20 · Estimated from token totals · overlapping detectors are not summed
Verified Savings
No verified changes yet
Alert — Slack or email at 50% and 80% of a configured budget. Observation only.
Recorded-spend guardrail — a per-key daily dollar limit returns HTTP 429 before the upstream call. Concurrent requests can overshoot; off until you set a cap.
Explicit model downgrade — This policy can change the model sent upstream. Enable with routing.budget_downgrade: true. Default is false. A budget alone never rewrites the request.
Run the same loop on your machine. No account.
$ pip install burnlens && burnlens scan && burnlens repos