Target duration: 4 minutes 50 seconds. All current metrics are visibly synthetic. This cut does not claim a real-world pilot, a consented participant quotation, or a continuously hosted backend.
This is one synthetic week in a small volunteer group: 214 messages, only three closed decisions, and a median decision time of 3.1 days. The problem is not a lack of communication. It is that commitments disappear inside the conversation. This is demonstration data, not a pilot result.
Most agents optimize for how often they can help. Quorum optimizes for how rarely it has to appear. It coordinates inside the group people already use, and gives every person a hard budget of two interruptions per rolling week. If the budget is exhausted, Quorum must reroute, batch, wait, or choose the published safe default.
Now I will replay that same synthetic week. Quorum first turns scattered promises into a Commitment Ledger. Every entry must retain its source-message reference and a verbatim evidence span. If the source cannot be proven, the ledger rejects the extraction instead of inventing certainty.
Three low-risk actions can complete without another group discussion: a tentative calendar event, a Gmail draft that is never sent, and a form request. Each action produces one short group receipt. Consequential changes remain reversible, and the undo path is signed, expires after 24 hours, and can be consumed only once.
When judgment is genuinely required, Quorum does not poll the room. Minimum-Quorum Routing asks only the smallest sufficient set of accountable people. Every question states its timeout and default. In this replay, only two people are contacted, and neither can exceed two decision requests in the rolling seven-day window. If no safe quorum remains, the action is deferred.
The main path is a five-node Strands Graph: Listener, Ledger Curator, Risk Appraiser, Quorum Router, and Executor. Risk and routing are deterministic, so model prose cannot change the score, quorum, timeout, autonomy, or interruption spend. Strands Swarm is used only for bounded semantic ambiguity. Before any tool call, a native hook interrupt re-reads the persisted policy and fails closed.
Two short-lived GitHub OIDC runs verified the AWS boundaries with both gates closed. AgentCore Runtime emitted one managed synthetic span, then returned HTTP 503 at the cost gate. AgentCore Memory reached ACTIVE, and AgentCore Gateway reached READY with exactly three IAM MCP tools. The probe reported zero model, Memory, Gateway, or external side-effect calls, and the workflows cleaned their temporary resources. PostgreSQL stays the authority for business facts; PII-safe OpenTelemetry records only allow-listed correlation fields.
Here is the same synthetic scenario side by side. Raw traffic starts at 214 messages and three closed decisions. Quorum appears six times in total, closes six decisions, and reduces the scenario's median decision time from 74.4 hours to seven hours. The scenario reports a 16.7 percent undo rate, which is why reversals stay visible instead of hidden. These numbers demonstrate the measurement contract and the product interaction. They are not measured community impact.
The repository also contains a 50-case synthetic gold set for commitment extraction. It reports precision, miss rate, and hallucination rate, and any prediction without valid source evidence is counted as hallucinated.
There is no participant quotation in this cut because no real organization has yet provided consented data or approved wording. Quorum makes that absence visible. A real quote will replace this segment only after the participant approves the exact sentence and its public use.
Quorum asks a different question of human-centered agents: not how much can they say, but how much coordination can they complete before they need to speak? The public synthetic replay and source are available now. The final submission screen will show the entrant's actual AWS Builder ID.
Use this only after written consent. Keep the current evidence-boundary segment otherwise.
On screen: organization type, study dates, sample size, and the exact approved quotation. Narration: “After the pilot, [ROLE — no unnecessary identity] told us, ‘[EXACT APPROVED QUOTE].’ This comparison covers [DATES] and [CONSENTED DATASET DESCRIPTION].”
Never describe synthetic replay metrics as pilot results, and never paraphrase a participant quote without approval of the final wording.