The problem with an AI-generated answer is not only that it may be wrong. It is that, once the session closes, the path that produced it can disappear with it. Project Relay begins from a different premise: models can do useful work, but the evidence trail and the authority to accept a consequential result must remain visible to people.
The invisible paper trail: a crisis of causal history
In the current era of rapid AI deployment, organisations are pivoting toward disconnected AI sessions: a prompt is issued, an answer is produced, and the context of how the conclusion was reached evaporates when a window closes or a log rotates. This is more than a lack of transparency. It is an accountability crisis. When a model provides a black-box answer without a traceable path, people are left to manage an outcome they cannot properly defend.
Project Relay is not primarily another AI tool for execution. It is a governance and evidence layer that puts the bounded task and the human decision record back at the centre of the work. It moves beyond the how-to of automation to ask the harder question: what does an organisation need in order to trust the result?
Authority is non-transferable: models are execution surfaces
Relay draws a firm line between execution and authority. AI models, however capable, are execution surfaces. They can synthesise large datasets, compare proposals, and draft complex work, but they are not the signatory of the decision record.
Authority is not an emergent property of model consensus. A silent acceptance, where an output triggers a business process or code merge without a discrete human gate, is a governance failure. Because a decision record is not writable by remote model clients, Relay keeps the division explicit: the machine does the work, while a human remains responsible for accepting the result.
Disagreement is a feature: the advisory overseer
Conventional AI strategy often tries to eliminate friction, averaging away differences between models through prompt tuning or consensus voting. Relay takes the opposite approach: disagreement is one of the most valuable data points in the system.
Its Overseer component is advisory. It identifies contradictions and evidence gaps between outputs, then triggers human remediation when models diverge. It does not select a best answer or quietly turn disagreement into confidence. By preserving material conflict, the system makes independence auditable and prevents the false certainty that comes from forced machine agreement.
Why it matters: seeing exactly why models disagree reveals the fragility of a process. An inspectable history of conflict gives a human authority a way to locate insufficient evidence and move from blind trust to informed judgment.
Your tools are projections, not the truth
Coordination interfaces are often mistaken for the source of truth. Relay uses a rebuildable-projection philosophy instead. Notion, Slack, GitHub, and status dashboards can be useful control rooms, but they are views over the work, not the canonical record of authority.
The strategic distinction is simple: interfaces change, authority does not. If a Slack thread suggests that a task is complete but the validated evidence record is missing a required review, the evidence record wins. Dashboards can be rebuilt. Causal history cannot be improvised after the fact.
- Slack — conversational coordination over validated Git records.
- Notion — a knowledge projection over evidence hashes such as SHA-256.
- GitHub — reviewable public-safe status over causal decision history.
- Green dashboards — a projection over protocol-compliant evidence bundles.
Integrity is not the same as truth: the R/01 principle
A document can be perfectly structured and its digital signature can prove that it has not been tampered with, while the claim inside it remains false. Relay treats this distinction as foundational rather than technical fine print.
Two reviewers can hold a valid hash of the same artifact, settling the question of whether they have the same bytes, and still disagree about whether the method satisfies the task. Relay preserves that separation so structural checks do not create a veneer of factual correctness around the final human gate.
The free path is local-first: resilience over reliance
In a market dominated by cloud-heavy universal model gateways, Relay’s reference implementation is a local-first CLI and library. That is a deliberate choice for privacy and resilience. Local-first is the fallback when hosted tools are unavailable, compromised, or simply not appropriate for the work.
Hosted services and provider APIs become opt-in accelerators: useful for speed, but not the source of trust. By keeping transcripts and evidence local by default, teams can ensure that the cheapest path to evidence is also independent of any single vendor’s uptime. The record remains available to inspect even when the service that helped produce the work is not.
The future of inspectable work
As synthetic content becomes abundant, the specific answer an AI provides will matter less than the evidence supporting it. The most valuable asset a modern team possesses is no longer its library of prompts, but its catalogue of defensible decisions.
Project Relay shifts the organisational question from “Trust me” to “Show me the evidence.” The goal of AI integration is not to remove the human from the loop, but to keep the human capable of remaining accountable at automated scale.
If you had to defend an AI-generated decision to a stakeholder tomorrow, could you show them the evidence, or just the output?