Relay turns disconnected AI sessions and human reviews into bounded tasks, reproducible evidence, explicit disagreements, and accountable decisions.
Why Relay exists
The workflow stays recognisable across domains: define a bounded question, compare attributed work against evidence, make gaps and disagreement visible, then leave the consequential decision with the person who owns it.
A researcher brings one bounded claim, its method, and the evidence needed to check it.
The claim, method, risk, and allowed evidence are recorded before review begins.
Independent workers inspect the method and evidence; missing steps and contradictions become review items.
The named researcher accepts, revises, or schedules another test with the reasoning preserved.
A team needs a traceable summary of sources, impacts, and unresolved questions.
The jurisdiction, scope, source date, fairness constraints, and accountable office are fixed.
Workers summarise the record independently; a reviewer flags unsupported, stale, or omitted claims.
The authorised official approves a bounded brief, requests remediation, or records why it is not ready.
A teacher wants consistent moderation without hiding uncertainty or appeals.
The task, rubric, moderation rules, and permitted student context are declared first.
Independent assessments attach reasons to the rubric; disagreement and missing evidence remain visible.
The teacher or moderator makes the grade decision and records any appeal or remediation path.
A panel wants comparable evidence without outsourcing the employment decision.
The role criteria, evidence policy, conflicts, and decision authority are locked before recommendations.
Workers map supplied evidence to each criterion; a reviewer checks omissions, consistency, and unsupported inferences.
The hiring panel decides whom to progress and records the rationale, including a request for human review where needed.
An incident lead needs an inspectable timeline rather than a confident story assembled from fragments.
Logs, symptoms, timestamps, hypotheses, and the safe scope of action are captured.
Workers build independent timelines; review tests chronology, causality, and gaps in the evidence.
The incident lead selects the next safe action, records the rationale, and keeps superseded hypotheses visible.
An owner needs to trace each conclusion back to the policy, artifact, period, and exception it relies on.
The control, scope, evidence window, exceptions, and accountable owner are declared.
Workers trace claims to artifacts; an independent reviewer checks scope, gaps, and conflicting interpretations.
The accountable owner accepts, remediates, or escalates with a durable record of what remains open.
An MCP/plugin or AI-client developer brings selected schemas, manifests, logs, process receipts, test fixtures, and bounded traces—not an entire private transcript archive.
Relay records the runtime, dependency versions, connector boundary, expected behaviour, permitted files, and safe reproduction scope.
Action chronology, reproduction matrices, failure clusters, schema/config diffs, tool maps, and security-boundary checks keep observations linked to evidence.
The developer or maintainer chooses the next code change, test, rollback, or connector action; Relay does not execute debugging commands or grant authority automatically.
Every Relay workflow starts by deciding what can be shared, with whom, and for what purpose.
Classify the record, name the authority, and choose the local, Notion, Slack, GitHub, or hosted projection.
Workers see only the approved, redacted context; receipts show what was attempted and what failed.
A named human approves publication, execution, retention, or deletion. Model agreement never substitutes for that decision.
01
One accountable record
Tasks, events, evidence, reviews, and decisions have stable identifiers, deterministic hashes, and causal history.
02
Evidence without execution
Methods, commands, environments, artifacts, failures, and hashes travel with a claim without Relay executing them.
03
Disagreement stays visible
Independent review and reproduction are explicit. Contradictions trigger remediation; they are not silently averaged away.
04
Humans keep authority
Models can contribute and compare. A named person still accepts, remediates, rejects, or defers consequential work.
05
Static live console
A credential-free projection of project and task status for GitHub Pages and local review.
06
Provider-neutral tools
Local read-only tools, the Overseer round-table chat, and guarded adapters for validate, verify, and prepare—without publishing or holding secrets.
01
Roadmap counts, pilot readiness, and current task projections from the canonical state snapshot.
Open status →
02
Default evidence gate, interactive claim trace, trust boundary, and deliberate non-goals.
Open protocol →
03
Local-first, Notion, Slack, and GitHub as rebuildable projections—never alternate authority.
Open control rooms →
04
Release route, build map, and what is deliberately deferred until the next gate.
Open roadmap →
05
Clone, install, check, and run the local multipage console—public-safe only.
Open start →
06
Understand the problem, the people it helps, and the boundaries that keep it accountable.
Open overview →
07
Plain answers about privacy, models, decisions, and what is available today.
Read answers →