Skip to content

Releases

SynapX-Harness 0.1.0 — Public Alpha

Pre-release

Choose a tag to compare

@moragu3001-ship-it moragu3001-ship-it released this 13 Sep 15:22

SynapX-Harness 0.1.0 — Public Alpha

Use your coding agent. SynapX checks whether the work is actually complete.

SynapX-Harness is an open-source assurance layer around coding-agent execution.

A core invariant is:

Agent DONE ≠ SynapX COMPLETED

Public Alpha scope

Public Alpha currently qualifies Codex Official Headless CLI through a governed execution path with:

  • explicit mutation scope
  • deterministic admission
  • controlled mutation
  • independent verification
  • evidence qualification
  • Harness-owned terminal completion
  • bounded execution timeout control
  • synapx doctor
  • minimal public Front Door

RQ8 clean-room qualification

The core governed execution path was qualified through RQ8 clean-room dogfooding using:

  • a sealed SynapX-Harness wheel
  • a separate Windows PC
  • a fresh external OSS checkout
  • a controlled regression
  • qualified RED evidence
  • scoped governed repair
  • deterministic verification
  • evidence qualification
  • Harness-owned terminal completion

The accepted RQ8 record reports:

  • pristine baseline: 297 passed
  • controlled RED: 24 failed, 273 passed
  • governed repair: VERIFIED, exit code 0
  • final independent GREEN: operator-observed 297 passed

The controlled regression was deliberately injected. SynapX-Harness does not claim to have discovered an upstream ItsDangerous defect.

Benchmark status

No benchmark superiority is claimed yet. Public Alpha begins the open measurement phase.

Future experiments will compare raw coding-agent execution with coding-agent execution governed by SynapX-Harness.

Known limitations

  • Codex is the only qualified coding-agent provider in Public Alpha.
  • Mutation tasks require an explicit allowed write set.
  • Paths requiring RED qualification fail closed when the required QualifiedRedExpectation is unavailable.
  • Qualification is currently Python/OSS focused.
  • Multi-agent orchestration, rich Execution Map, full Evidence Explorer, and additional agent providers are not Public Alpha capabilities.

Philosophy

Failure is permitted. Unqualified completion is not.

Reasoning proposes. Evidence qualifies. Authority decides.