BUNSEN · ISOLATED

Model routing receipts

A forced primary failure fell back to minimax-m3, and the usage receipt recorded the served model.

Hypothesis

A configured fallback can be functionally successful while observability still reports only the requested model. The isolated run must prove both behavior and receipt quality.

Method

Three fresh, identity-free Bunsen profiles exercised the same scoped Skynet key:

  1. primary minimax-m2.7-highspeed;
  2. direct minimax-m3;
  3. an intentionally invalid primary with minimax-m3 as fallback.

Result

  • Live roster: 9 models; both intended models present.
  • Primary: 15.16 s, expected response present, receipt names minimax-m2.7-highspeed.
  • Direct fallback: 14.32 s, receipt names minimax-m3.
  • Forced fallback: 25.95 s, expected response present, receipt names minimax-m3.
  • Canonical Bunsen config SHA-256: unchanged before/after.
  • Secret scan: PASS.

Context calibration

The Bunsen context window is 128,000 tokens with an 8,192-token completion reservation. The configured compression threshold is 0.70, but Hermes’s effective floor is 0.75. Compression begins at 89,856 input tokens and preserves a 10,782-token tail.

Falsifier

Fail this claim if the fallback response succeeds but the receipt names the invalid primary, the canonical config hash changes, or the model is absent from the authenticated live roster.