KUJO THE COURSESTAGE 6 / Agentic systems · LESSON 38
Mental model
A result says what happened; evidence lets another observer assess that claim. A useful receipt identifies the input, runtime, bounded actions, evaluation, and final disposition. It is compact enough to review and specific enough to reproduce.
Data contract
Use a versioned dictionary with stable fields. Separate factual observations from the verdict derived from them. Keep raw credentials, irrelevant private data, and hidden reasoning out of the record. A request hash or artifact digest can identify evidence without copying every input into every log.
The example records attempts, a verified count, and a PASS verdict. It tests that the verdict corresponds to evaluation, rather than trusting an arbitrary status string. The capstone persists a richer packet to a local work directory.
Ecosystem choices
RunLedger supplies run-record workflows; Casefile captures failure evidence; PatchBrief describes source changes; ChangeBucket measures change footprint. These solve different problems. Select a tool because the next reviewer needs its evidence, not to satisfy a repository checklist.
Dispatch and explicit handoffs should carry a bounded goal, context references, allowed effects, acceptance criteria, and a stop policy. Passing a task to another worker does not grant broader authority. The receiving host must resolve permissions independently.
Professional pattern
Include source/version provenance, deterministic inputs or references, executed checks, exit statuses, and a concise human summary. Preserve failed attempts when they explain the final disposition. Define what evidence to retain, for how long, and who may read it, separately from how telemetry is delivered.
Common mistakes
Do not store an entire transcript as a substitute for a useful receipt. Do not report an action as completed merely because it was dispatched. Do not accept an upstream handoff's claimed approval without the host's authorization state.
Working example
let passed := 2 == len(["a", "b"])
let receipt := {"schema": "course.run/v1", "runtime": "1.3.1", "attempts": 1, "evaluated": passed, "verdict": "PASS"}
assert_equal(receipt["evaluated"], true)
print(to_json(receipt))
Run it
From the course repository root, use the pinned Kujo 1.3.1 runtime.
kujo check examples/38.kujo
kujo run --untrusted examples/38.kujo
Captured output
{"attempts":1,"evaluated":true,"runtime":"1.3.1","schema":"course.run/v1","verdict":"PASS"}
Break it and diagnose it
The receipt claims PASS without passing evaluation. A status label cannot replace evidence.
let receipt := {"verdict": "PASS", "evaluated": false}
assert_equal(receipt["evaluated"], true)
kujo run --untrusted examples/38-break.kujo
Exit status: 4. Captured diagnostic:
[KUJOVM001] [vm] Runtime Error: Assertion failed: expected Bool(true), got Bool(false)
--> 0:0
Exercise
Design a receipt for your runner and a handoff for one bounded subtask. Include input identity, authority, limits, evaluation, and a stop result. Remove any field that stores secrets or irrelevant raw content.
Checkpoint
- I distinguish dispatched work from completed work.
- I tie verdicts to evidence.
- I carry explicit boundaries through handoffs.
Verified 2026-09-06 · Official Kujo 1.3.1 release · Source contract · Download example