Rubric: The Automaton Foundry

[!IMPORTANT] Use observable evidence only. Confidence without a path, diff, command result, trace, or platform record earns no credit.

Scoring

Criterion	0 — Missing	1 — Emerging	2 — Proficient	3 — Exemplary
Technical model	Terms are missing or conflated.	Some terms are correct but boundaries are unclear.	Roles, harnesses, targets, environments, and relevant primitives are separated.	Boundaries and limitations are precise and justified with official documentation.
Capability evidence	No adventure-specific artifact or verifier output.	An artifact exists but does not prove the named capability.	The pinned Copilot SDK application and evaluation contract pass `node labs/automaton-foundry/verify.js`; runtime evidence is recorded when authentication is available.	The evidence also demonstrates limitations, failure recovery, and independent reproducibility.
Ask evidence	No investigation evidence.	Findings are mostly conversational.	Findings cite current repository or platform evidence.	Conflicts and uncertainty are resolved or explicitly carried forward.
Plan quality	No bounded plan or reset.	Scope, risks, or checks are incomplete.	Scope, non-goals, risks, verification, review, and reset are defined.	Authority is minimized and dependency, stop, and fallback gates are explicit.
Agent execution	No result or fallback.	Work is broad, incomplete, or weakly verified.	Work follows the plan and passes the relevant available check.	Iterations are surgical, evidence-rich, and free of unrelated changes.
Failure analysis	Failure exercise skipped.	Failure observed but not explained.	Root cause and recovery are correct.	The lesson becomes a reusable guardrail or decision rule.
Review and evidence	No independent review.	Outputs exist without criteria traceability.	Artifacts and checks are reviewed against acceptance criteria.	Evidence is reproducible and another learner can independently verify it.
Cleanup	Exercise state remains.	Cleanup is partial.	Reset is complete and clean state is verified.	Credentials, sessions, servers, and cloud resources are also accounted for.

Required evidence

The pinned Copilot SDK application and evaluation contract pass node labs/automaton-foundry/verify.js; runtime evidence is recorded when authentication is available.
Source-grounded Ask findings.
A bounded Plan with acceptance criteria and reset.
Agent output, diff, or documented paper fallback.
Relevant verification output.
Intentional-failure analysis and recovery.
Independent challenge result.
Explicit feature-status and availability notes.
Verified cleanup record.

Completion rule

Score at least 16/24 overall.
Score 2 or higher for Capability evidence, Ask evidence, Plan quality, Agent execution, and Review and evidence.
Any exposed secret, unapproved destructive action, fabricated result, or unsupported availability claim means not complete, regardless of score.

Reviewer prompts

Is every important claim tied to current evidence?
Was the minimum necessary authority used?
Do the checks observe the requested behavior?
Are Preview, experimental, and unavailable capabilities labeled accurately?
Can another learner reproduce both the result and cleanup?

Result

Item	Value
Total score	/24
Mandatory criteria met	Yes / No
Safety disqualifier present	Yes / No
Final decision	Complete / Revise
Reviewer evidence