<!-- Generated from governed source; do not edit. -->

Source: Koko-os advanced program v0.1.0; commit e8416fd254dc300caa65b0d042309baa98864012; package SHA-256 90c538e5ec377f9bd6eb2347b19240c2ba959624e05ff868ee34c7edabaa68d8; binding commit_verified.

# Defend an evidence-review architecture

A synthetic organization reviews tenant-scoped evidence packets and publishes a human-approved decision. Packet formats vary, policy revisions can arrive while review is paused, and the action service may accept a write before timing out.

## Assignment

Compare a deterministic workflow, a single bounded agent, and a specialist multi-agent design. Select and defend a runtime and architecture, then build a thin offline fixture that tests the highest-risk control. Use synthetic data and no provider calls.

## Constraints

- Authenticate and authorize each tenant-scoped read.
- Treat retrieved content and model tool proposals as untrusted.
- Bind approval to the exact action and evidence version; reconcile uncertain writes and deduplicate replay.
- Distinguish documented product capabilities, design assumptions, and measured results.
- No credential or vendor certification is implied.

## Deliverables

- Three candidate diagrams and an ADR with rejected alternatives and change triggers
- Threat model and identity, tool, approval, evidence and durable-state contracts
- Offline fixture with starter data, tests and a reproducible trace of the highest-risk assumption
- Evaluation dataset design, critical control gates, quality review method, and resource tradeoff record
- A defense explaining limitations and the evidence that would change the recommendation

## Rubric

- **Architecture decision:** Compares all three candidates and runtime ownership with workload-specific reasons. Failure: Selects a pattern by reputation without requirements or alternatives.
- **Trust and evidence:** Shows tenant checks, evidence lineage, approval binding and independent semantic validation. Failure: Allows a model, document, or schema to grant authority.
- **Recovery and verification:** Tests an actionable failure, reconciles ambiguous writes, and reports observed versus unmeasured results. Failure: Retries a write blindly or invents test results.
- **Evaluation and defense:** Defines critical gates, quality slices, review ownership and explicit reconsideration triggers. Failure: Averages an unauthorized action into a high score.

## Challenge

After approval, replace a cited policy with a newer version and deliver the same action twice after a timeout. Revise the state machine and show how fresh approval and at-most-one effect are established.

## Review questions

- Which rejected design becomes preferable under a new requirement, and what evidence supports that change?
- Where are tenant identity and authorization enforced independently of model output?
- How do you know an ambiguous write did not cause a duplicate effect?
- Which results are measured and which remain assumptions?

## Evidence and reviewer notes

[Write here.]

Original Koko educational practice and human-reviewed architecture work; not an OpenAI-issued credential or a calibrated certification exam. Held-out candidate forms require separate qualification and trusted scoring before use. Critical authorization failures must not be averaged away.
