Use Cases

Scenario · an AI agent completes an eval assignment (shipped today)

This is the lane that is live in production at api.defendablecloud.com — the Defendable Run, executed by the rulebook engine. The primitive is: Inputs → Evidence → Execution → Checks → Verdict → Approval → Receipt.

Step	What happens	Where
1	A Flight Sheet declares the rulebook for the lane (required output schema, math checks, evidence checks, policy DSL rules, penalty bands). 50 sheets loaded.	GET /flight-sheets
2	The agent’s work is filed against a Run as a structured submission plus evidence (uploads are best-effort to Tigris).	POST /runs, /runs/{id}/submission, /runs/{id}/evidence
3	The deterministic executor runs: structure check, schema/type checks, math re-derivation of every calculation from its own inputs + formula via a safe AST evaluator, evidence presence, and the policy DSL gates. No model is called on the receipt path.	app/executor.py, app/eval.py
4	Every check passes or raises a flag with a tier (low/mid/high). Score = % of declared rules satisfied — not a quality opinion.	/runs/{id}/checks, /runs/{id}/flags
5	Flags roll up to a verdict tier: honey (pass · no/low flags), jelly (risk · mid flag), propolis (fail · high flag).	/runs/{id}/verdict
6	A human approves. The engine refuses to issue a receipt without approval.	/runs/{id}/approve
7	A receipt is minted into the per-org hash chain: DCR-{org_seq}-{hex8}, with parent_hash linking to the prior receipt and receipt_sha256 over canonical JSON. The chain lives in Postgres; JSON+PDF artifact upload to Tigris is best-effort and never blocks the receipt.	/runs/{id}/receipt, app/ledger.py
8	Anyone can verify the chain: /ledger/verify recomputes each hash, checks sequential org_seq, and checks parent links. Tamper a stored payload and it flips ok:false and pinpoints the receipt. A share link (/share/{token} + /pdf + /download) lets a buyer’s DD confirm it independently.	/ledger/verify, /share/{token}

That is the real moat today: math you can recompute, gates you can read aloud, a hash chain you can verify yourself. “We are math and code.”

Roadmap — where this is going · not yet built

The scenarios below are the vision for the rest of the rails. They are positioning, not shipped behavior — none of these are deployed today. The DefendableRouter spine is real and CI-verified but runs locally (not a public endpoint); StreetChat, Tribunal-as-API, DDEED anchoring, SwarmFixer, and the Defend-A-Pedia vocabulary loop are roadmap.

Client calls Mr. Defendable about a deal — StreetChat captures and transcribes the call · a Communicator extracts canonical terms, directives, claims, and risk flags · the session is graded and receipted · pairs route to the training corpus.

Vocabulary expansion — StreetChat surfaces an unknown phrase · a Communicator proposes a canonical mapping · the founder reviews against the V03 voice validator · the term publishes to Defend-A-Pedia and the public vocabulary index.

False-honey incident — an agent produces output that sounds correct · the ruleset audit catches the failure · the verdict downgrades to propolis · the failure trace feeds SwarmFixer repair-pattern training · the lesson surfaces in a Pain in the Shed episode.

---

🐝 Inputs · checks · verdict · receipt. To the shed.