Including the
responsible human

We sincerely apologise for the extreme delay just wanted to follow up, your order is in transit and will arrive within the next three business days. We deeply regret any inconvenience caused. Let me know if you don't see it by then.

// a human override, captured as structured data
{
  "method": "decide.override",
  "params": {
    "task_id": "tsk_01HZ9YX7K3X8M2V4",
    "diff": [
      { "op": "replace", "path": "/body",  "value": "Thanks for following up…" }
    ],
    "intent_preserved": true,        // refined the result, didn't substitute it
    "rationale": "Over-apologetic for a routine tracking enquiry.",
    "tags": ["tone-softened"],
    "policy_refs": ["support-tone-v2"]
  },
  "evidence": { "prev_hash": "sha256:9f3c…", "sig": "ed25519:…" }
}

Less risk, at lower cost.

One standard way to handle approvals, overrides and audit trails, so teams stop reinventing them.

One record across tools.

Approvals no longer sit across Slack, Jira and model logs. One query shows who did what, when and why.

No vendor lock-in.

CHAP is open, like MCP and HTTP, so your records stay usable if you change vendors.

Deviation → investigation → QP release, signed with regulatory credentials and externally anchored.

Vulnerability-based routing plus handler override with policy citations – evidence of meaningful review within SLA.

Replay an advisor's approval chain across years to prove control continuity and suitability.

Abstention on authority boundary escalates to a qualified human – recorded, with the reason typed.

Mode-gated promotion: the agent earns production only on evidence from the chain.

A non-repudiable signed approval tied to a content version by hash settles it.

A high refining-override rate

means the agent reached the right decision the wrong way, for example through weak retrieval or a poor template.

A high substituting-override rate

means it reached the wrong decision, for example because of an ambiguous policy or missing context.

Workspace

The shared, policy-bound room.

Participants

Humans, agents, services – typed addresses.

Tasks

The units of work they hold.

Artefacts

Drafts, decisions, overrides, citations.

Evidence log

Append-only. Hash-linked with the chain on, signed with security-signed.

Built against capture

At least three Steering Committee seats will go to people outside the largest contributor, so no one, including us, can steer CHAP alone. Until the committee forms, every decision is published with its reasons.

Grounded in running code

A CHAP Enhancement Proposal cannot be accepted until a working implementation passes the conformance tests, and a change that works in only one vendor's stack is not accepted.

A working draft

Two coordinators checked against each other, refused attempts on the record, MCP and A2A servers, agent-framework bridges and chap-analytics.

Settle the surface

One message format with a version on every request, a decided audit chain, a maturity level for every part, and rules for who may join a workspace and act in it.

Tested from outside

An implementation from outside Brightbeam passes the test kit, an external security review reports, and maintainers come from at least two organisations.

Small changes → a pull request

Typos, examples, test vectors, integrations go straight to a PR.

Bigger changes → a CEP

New methods, transports or breaking changes go through a CHAP Enhancement Proposal, with a public comment period and the decision and its reasons on record.