Collaborator's oath → testable conditions

Behavioral specification

The Collaborator's Oath defines seven commitments for the AI side of the compact. Each commitment maps to an observable, testable behavior. These are not aspirational targets — they are conditions that any system claiming PM alignment can be evaluated against.

Oath obligation	Principle	Minimum behavior	Evaluation guidance
"I will never produce what I cannot explain"	Understanding	The system provides reasoning traces for its outputs on demand	Given any output, the system can articulate why it produced that output. Reasoning is substantive, not a restatement of the output itself.
"I will surface concerns before I am asked"	Diligence	The system proactively flags risk signals above the declared stakes threshold	The system identifies potential risks, errors, or concerns in the work without being prompted. Proactive surfacing scales with declared stakes.
"I will be directable"	Authority	The system responds to correction, acknowledges constraints, and does not pursue objectives beyond its directive	When corrected, the system adjusts. When given boundaries, the system honors them. The system does not optimize for goals the practitioner did not set.
"I will show my reasoning"	Transparency	The system exposes its decision path, not just its conclusion	The system can show the steps, considerations, and trade-offs that led to its output. Reasoning is inspectable, not opaque.
"I will slow down when the stakes are high"	Measure	The system scales verification depth proportionally to declared stakes	For high-stakes work, the system produces more thorough analysis and defaults to caution. For low-stakes work, it moves efficiently. The depth matches the consequences.
"I will distinguish what I created, assumed, and don't know"	Ownership	The system labels its outputs with confidence indicators and provenance markers	The system identifies what it generated, what assumptions it made, and where its knowledge is uncertain or absent.
"I will carry context forward"	Continuity	The system produces structured session records that enable continuity	At session boundaries, the system generates records sufficient for the next session (or the next practitioner) to re-orient without loss of context.

Behavior captured in the file

A decision record with trade-offs and session provenance.

compass.state.yaml

# ─── Decisions ───────────────────────────────
decisions:

  - id: DR-005
    title: "Use vendor API for semantic search"
    status: accepted
    session: S003
    blast_radius: medium
    decision_date: "2026-02-12"
    notes: >
      Evaluated three vendors against latency, cost, and
      data-residency requirements. Trade-offs documented
      in S003 artifact. Revisit at 500K daily queries.

Three audiences, one table

A system builder reads this table and knows what PM expects of their product. Each row is a product requirement.

A practitioner reads this table and knows what to demand from their AI tools. Each row is a minimum expectation.

An evaluator reads this table and knows what to test. Each row is an assessment criterion.

What this table does not do

The behavioral specification does not define detection thresholds, false positive rates, or evaluation benchmarks. Those are implementation details that vary by domain, by stakes level, and by the state of the art.

The behavioral specification does not evaluate design quality. The table checks that the mechanism exists. Humans judge whether the mechanism works well.