We test the gate on our own work.
Instrumented file, shell, and Git operations from our founder's coding-agent workflow pass through Warm Winter in shadow mode. The July record below shows an early stage of that work. Shadow mode records; it does not block or grant execution authority.
Historical dogfood snapshot · July 19, 2026
This snapshot records the first shadow phase. Warm Winter has since added stronger CI and deployment outcome binding and is testing contextual competence in shadow. The historical record remains here to preserve the progression rather than rewriting it after the fact. These figures are not a current production reading.
Historical snapshot · recorded 2026-07-19 · Founder coding agents · shadow mode · weak labels (tool-exit). Counts include weak execution checks and cannot establish strong outcome coverage.
0 of 5 verified, and that's the point.
Warm Winter won't claim trust it hasn't earned. Provisional means there's signal but not enough verified evidence to support bounded autonomy. Ungrounded means it doesn't have the evidence yet, and says so instead of pretending. These historical labels do not describe the current production frontier.
The honest caveat: these are weak tool-exit labels. They show that the observation and reporting loop works; they do not prove that an edit was correct, a command was safe, or a push was valuable. Strong CI and deployment-health outcomes can supply better evidence when bound to the actual action. Verification also requires coverage, calibration, resolution, and freshness.
Read how evidence earns trustWhat shadow operation taught us.
Field note · September 2026
Average success is not case-specific judgment
- Observation
- Shadow operation exposed a limit in the first Gate design. Fixed-confidence action categories could learn their average success rate without showing that they could distinguish one specific action from another.
- What changed
- This finding is driving contextual prediction, stronger action/outcome identity, and competence conditioned on trusted runtime state and evidence. The Gate did not promote cells merely because sample count increased.
- Evidence boundary
- This is an internal engineering finding from founder dogfood, not customer validation or a causal value claim. Contextual Gate remains shadow-first.
- What remains unknown
- Prospective case-level calibration and resolution, reliable outcome attachment, and promotion readiness remain unproven.
See your own agent's frontier.
Guided setup, shadow mode, nothing to break. It installs, observes, and starts building your record from request #1.
Connect Codex, Claude Code, Hermes Agent, Gemini CLI, OpenCode, Goose, OpenClaw, Cursor Agent, Grok agents, or another local or remote agent through the same evidence contract. Observation is universal; governed execution is certified runtime by runtime.