Verification standard
“Verified” is a receipt, not a slogan.
A claim is verified only when its checks pass, its fixture is public, and its receipt matches the commit served by the site.
The four required links
- Commit identityThe receipt records the exact source commit and site commit. They must match.
- Named fixtureThe scenario, harness, pack version, inputs, and expected outputs are public.
- Passing checksRepository, installer, reference, and outcome checks all report pass without a hidden baseline.
- Limits and rollbackThe receipt states what was not tested and how to remove the installed files safely.
Current status
The public site remains pending until the full repository gate and an outcome fixture pass on the same main commit. Local improvements alone do not turn this label green.
Runtime evidence: clean Claude Code client
Per ADR-0003, structural and runtime evidence are separate scorecards. The results below are real and reproducible, but they are not the public outcome fixture the site requires, so the label above stays pending. They make no claim about any harness other than Claude Code.
| Field | Value |
|---|---|
| commit | ad6582f8621bf3c21a116828626869b6c481a09a |
| package | agi-super-team@1.6.0 |
| harness | Claude Code 2.1.207 |
| fixture | none — ad-hoc reproduction, not a public fixture |
| isolation | CLAUDE_CONFIG_DIR pointed at <home>/.claude; empty project directory |
The five cleanClientReceipt.requiredChecks declared in config/harness-adapters/claude-code.json, measured on a fresh install:
| Check | Result | Observed |
|---|---|---|
fresh-config-directory | passed | Install into an empty home succeeded; 106 agent definitions and 170 Skill directories written. |
agents-discovered | passed | Session initialization listed all 106 ast-* agents and 169 skills, including agi-super-team-orchestrator. |
orchestrator-semantic-trigger | passed | An unprompted cross-functional request caused the model to invoke the agi-super-team-orchestrator Skill on its own. |
ceo-manager-leaf-dispatch-observed | passed | Nested dispatch observed to depth 2: the CEO selected five C-suite managers, and managers dispatched their own leaves, in parallel branches. |
governor-independent-review-observed | passed | ast-governor gathered its own evidence, returned a Block verdict, and escalated two unexamined veto points. |
Contract and artifact agree
The delegation block declares 11 managers plus the coordinator. In the installed artifacts exactly 12 definitions omit disallowedTools: Agent (ast-ceo plus those 11), and the remaining 94 — every leaf and ast-governor — carry it. Nested dispatch reached depth 2, matching requiredMaxDepth.
Other primary harnesses
Codex, OpenClaw, Hermes, and DSH each install and generate their connection receipt, but none has been run against a clean client. The table records what was verified in this pass and what was not, so an unmeasured surface is not read as a measured one.
| Harness | Installs and lands where the README says | Receipt declared | Clean-client run |
|---|---|---|---|
| Codex | verified | 5 checks | not run |
| OpenClaw | verified | 7 checks | not run |
| Hermes | verified | 4-step flow | not run |
| DSH | verified | 4-step flow | not run |
"Installs and lands where the README says" means a real run into an isolated root produced the documented paths, not that a client loaded them. The receipt column names the check list each adapter emits; none has been executed.
Limitations
- One local run by the maintainer, not a public fixture. A different model, prompt, or harness version may behave differently.
- Only Claude Code was exercised. Codex, OpenClaw, Hermes, and DSH remain
runtimeEvidence: pending. - Prompts were written by the maintainer to elicit dispatch deliberately; this is not an unbiased quality measurement.
- Reproduce by installing into a clean
CLAUDE_CONFIG_DIR, then rerunning the commands below with--output-format stream-json --verboseand inspectingparent_tool_use_idon eachtool_use. - Rollback: remove the installed
.claude/agents,.claude/skills, and.claude/agi-super-teampaths, or delete the throwaway config directory entirely.
Receipt contract
{
"commit": "full Git commit SHA",
"siteCommit": "commit used for this Pages artifact",
"pack": "starter kit and version",
"harness": "tested runtime and version",
"fixture": "public fixture identifier",
"checks": [{ "name": "repository", "result": "pass" }],
"result": "pass",
"limitations": ["explicit untested conditions"]
}
Run the repository evidence yourself
npm test
npm run validate
npm run validate -- --warnings-as-errors
bash -n install.sh
git diff --check
Run these commands from a clean checkout. A public receipt can still become stale when main changes, so the site checks receipt identity before showing a verified state.
Rollback
The installer never overwrites existing personas, user files, or skills. Review the preview output, keep the listed destination paths, and remove only files introduced by the selected kit. If a step fails before publish, the staged directory is discarded and the destination stays unchanged.