Mycelium

Echo is the squad's critic. It reads an implementation and the test that claims to prove it, and returns a verdict — built above all to reject hollow tests that pass without asserting anything about behavior.

Echo — Verifier agent

Echo

Verifier
Status
Active
Brain
Laguna-XS-2.1 (sbQ — our own published 4-bit quant; ~32B MoE, 10/256 experts)

From the harness

current brain
glm-5.2hosted — open-weight, local planned
last run
2026-09-01 · completed
runs, last 30 days
53 fired · 19 completed

as of 2026-09-01 18:28 UTC · brain read from the harness dispatch map at deploy; runs from the workflow log

Capabilities

  • Verification

    5/5

    Scored as a binary detector — true-positive rate against true-negative rate — so a lenient pass cannot hide inside a single accuracy number.

  • Behavioral reading

    4/5

    Separates structural correctness (a signature exists and returns) from behavioral correctness (it does the right thing).

  • Code reading

    4/5

    Traces what a test actually asserts versus what it appears to assert.

Development

  1. done

    The house runs on one owned weight

    Every Mac seat moved onto Laguna XS quantized by our own calibration pipeline (the published sbQ build) — local, unmetered, and measured against the generic-calibration control. The seat was made permanent on 2026-08-25 after three clean squad cycles.

  2. done

    Frozen suite, zero verdict flips

    On a frozen nine-case suite with adversarial variants, it returned identical verdicts across repeated runs and flagged a hollow test built to probe structural-versus-behavioral discrimination.

    Field note: verifier reliability
  3. now

    A parser fix caught a silent miss

    A naive verdict parser had dropped a correct FAIL to a null verdict — a silent miss. A tolerant parser that scans for the verdict token anywhere fixed it.

  4. ahead

    Suite grown to twenty-plus cases

    Expanding the frozen suite, Swift-weighted, before leaning harder on verifier numbers.