A real, confirmed opportunity we could not verify, and said so

This is not a savings-proof case study: there is no verified number here, and it should never be cited as one. What it is: a concrete example of the verification discipline itself, a real opportunity, diagnosed and re-confirmed with live data, that gets held back rather than shipped on the strength of an estimate alone, because the standard verification path structurally cannot reach it. That's a story about rigor, not about a delivered percentage.

The opportunity is real, not a dead end

The deep autoresearch loop found a genuine sharding opportunity in Angular's ci.yml: the adev job runs 161 independent test targets as one serial block, followed by adev-deploy's own production build. Composed estimate: 1386s → 666s (+51.9%) over two accepted changes. We shard-authored the first change for real: a genuine, working structural change, not a placeholder.

Then we checked whether the underlying hypothesis holds up against real data, not just the model. We pulled job-level timestamps from 5 independent real ci.yml runs.

Why it's unverified anyway

Live verification requires pushing a candidate branch to a fork and comparing real CI runs. Two structural facts block that here, neither fixable from our side: ci.yml only triggers on push to protected branches, never on a fork's pull request; and its jobs need credentials (including deploy secrets) that don't exist on a personal fork.

We piloted against the fork-safe PR-triggered workflow instead. That came back statistically null, but root-caused, not just measured: that workflow has no deploy job at all, and its own critical path is an unrelated 111-second job. The pilot tested the wrong workflow. It doesn't disprove the real hypothesis; it couldn't reach it in the first place.

The only remaining path

Verifying against the real workflow requires push access to Angular's protected branches and Angular's own deploy credentials, i.e., landing this as a genuine pull request against angular/angular itself, not an automated verification run. That's a different, public, reputational kind of action we don't take without explicit client sign-off. What's here documents a real opportunity the verification discipline correctly declined to claim without proof, not a dead end, and not a delivered result either.