Synopsis
$ zevron.exe --versionGive Zevron a plain-language description. It decomposes the task into units, and for each one a Specsmith writes the tests and a Codewright writes the code — blind, with different models — then runs and attacks them.
Example
zevron — build
zevron ❯ zevron build "a todo cli with persistence and search"
⟶ foreman decomposed into 6 units
⟶ specsmith authoring tests (visible + hidden) · 6 suites
⟶ codewright generating candidates · n=3 · visible tests only
cli.parser T0 ✓ T1 ✓ T2 ✓ T3 ✓ mutation 0.86 trusted $0.04
store.sqlite T0 ✓ T1 ✓ T2 ✓ T3 ✓ mutation 0.91 trusted $0.06
task.model T0 ✓ T1 ✓ T2 ✓ T3 ✓ mutation 0.84 trusted $0.03
export.markdown T0 ✓ T1 ✓ T2 ✓ T3 ✓ mutation 0.88 trusted $0.04
search.index T0 ✓ T1 ✓ T2 … verifying $0.05
import.csv T0 ✓ T1 ✓ T2 ✗ not trusted → forge $0.02
5 / 6 trusted · held-out pass-rate 77% · spend $0.24
run zevron trust <unit> for evidence.
Flags
| Flag | Default | Description |
|---|---|---|
| --mode | verified | Verification depth. hardened runs the full Crucible. |
| --n | 3 | Candidate programs generated per unit. |
| --budget | none | Stop once total spend reaches this cap. |
| --json | off | Emit machine-readable output. |
| --no-registry | off | Disable registry reuse for this build. |
Output
Each unit reports the tiers it cleared (T0–T3), its mutation score, a status, and a cost. The summary line reports how many units are trusted, the held-out pass-rate, and total spend. Units that fall short are marked not trusted and returned to the forge.
To see the evidence behind any verdict, run zevron trust.