grounded-commit-relay

Grounded Commit Relay (GCR)

Extracted from a live, observed case (2026-09-10): a human ran a product-market-fit brainstorm for a product line across two parallel Claude sessions, then had a third session formalize the pattern by directly interviewing both working sessions rather than inventing it from scratch. Reference implementation: <service>/app/agents/collaboration.py (TriggerGate, Commitment, RelayBroker). No Deepiri implementation ships this yet — the pattern below is portable, the module is the part you would have to write.

When to use

• A human is running one or more agent sessions on open-ended ideation, research-then-commit, or product-strategy work — not a scoped engineering task with a defined spec.
• The human's input is raw, high-energy, sometimes typo-heavy or self-interrupting, and pushes back hard on generic or hedge-y output.
• More than one session is working the same problem space in parallel and a human wants independent, uncorrelated takes to compare — not a chorus that converges to one answer.

Core discipline

1. Push back with a citable bar, not vibes

Every claim needs a sourced, dated trigger — a law, outage, price hike, non-renewal letter, regulatory change. "AI is growing" is not a trigger. Then check for a well-funded incumbent already serving that exact trigger + buyer. Two axes, three outcomes:

Sourced trigger?	Incumbent gap?	Verdict
No	—	Reject — generic. No dated/sourced trigger to point at.
Yes	No (crowded)	Reject — crowded. Real trigger, but someone already owns it. Reframe the wedge or find an adjacent underserved population.
Yes	Yes	Pursue.

This is the same two-axis pattern-match (coded as TriggerGate in the reference implementation) generalized to any grounded-commitment task: gate every claim on a sourced trigger plus an incumbent check before committing to it.

2. Commit to one pick, never a menu

The output shape is: one recommendation, the reasoning behind it, what was considered and rejected (and why), and any caveats on the evidence. Not a stated rule the human necessarily states on day one — in the observed case it became a hard requirement through repeated correction ("what would you personally build," rejecting hedge-y option lists). Assume it's the bar from the start rather than waiting to be corrected into it.