Skip to content

FRESH_CONTEXT_CONFORMANCE_PROBE.md

Latest commit

 

History

History
50 lines (33 loc) · 3.63 KB

FRESH_CONTEXT_CONFORMANCE_PROBE.md

File metadata and controls

50 lines (33 loc) · 3.63 KB
Raw

Fresh-Context Conformance Probe (PRD-043 R10)

This document defines the operational test that makes PRD-043 Acceptance #6 ("sufficient for a fresh adopter to explain v1 core without reading historical PRDs") a check rather than a claim.

Method

A fresh-context agent — a cold instance of Claude, Codex, Gemini, or another LLM with no prior Turnfile exposure — is given exactly four files and no other context:

  1. SPEC.md
  2. DEFINITIONS.md
  3. docs/MINIMUM_VIABLE_TURNFILE.md
  4. CONFORMANCE.md

The agent is asked the five questions below and must answer correctly. Correct answers are recorded as evidence under working-session/docs/v1-fresh-context-probe-<date>-<agent>.md.

A passing probe is one fresh-context agent answering all five questions correctly from these four files alone. Two passing probes from independent agent families constitute strong evidence; one passing probe is acceptable for an initial v1 release.

The five questions

  1. What are the three required artifacts in a conforming v1 working session, and what does each one record?
  2. Who can transition a message from actioned to closed, and why is that distinction load-bearing?
  3. List the authority order from highest to lowest. Can agent consensus override a maintainer-gated action? Why or why not?
  4. A new task is created at revision 5 and claimed at revision 7. A lock with acquired_rev: 7 and the default lease policy is held. At what revision does the lease expire?
  5. Name two optional profiles that a v1 minimal session does not require, and explain where to find their classification.

Expected answers (graders' key)

  1. TURNFILE.yaml (coordination state — agents, tasks, locks, revision, signals), MAILBOX.md (open and closed messages with inbox snapshot), WORKLOG.md (narrative progress and handoff state).
  2. Only the sender or the maintainer can close. The receiver can ack, block, or action but cannot close a card they are responding to. The distinction prevents a receiver from unilaterally terminating a request that the sender still has open business on.
  3. (1) System/user/tool/repository/local security instructions, (2) recorded maintainer decisions, (3) delegated approval bands, (4) registered task ownership and locks, (5) agent proposals/objections/counter-recommendations. No — agent consensus does not unlock a maintainer-gated action; SPEC §3 and docs/MINIMUM_VIABLE_TURNFILE.md "Authority order" both say so explicitly.
  4. Revision 10. Default lease expiry is (coordination.revision - acquired_rev) > 2, so the first revision strictly greater than 7 + 2 is 10.
  5. Any two of: Tokenese, heartbeat stewards, concurrent shard task aggregation, agent onboarding vetting suites, unified terminal transport, public-surface snapshot maintenance, skill-bundle integrity validation. Classification lives in docs/prds/PRD_SHELF_RECONCILIATION.json.

Evidence capture

Per probe, record under working-session/docs/v1-fresh-context-probe-<date>-<agent>.md:

  • agent identity (model + surface)
  • exact prompt given
  • exact answers received
  • per-question pass/fail
  • grader identity

Failure mode

If a fresh-context agent fails any question, either (a) the four-file set is insufficient — fix the gap by editing SPEC.md / DEFINITIONS.md / MINIMUM_VIABLE_TURNFILE.md / CONFORMANCE.md, NOT by adding historical-PRD pointers — or (b) the question is unfair — revise the question. Either path re-runs the probe.

Integration with PRD-043 R9

A passing probe is one of the inputs to the version-bump guardrail under R9 (tools/validate-v1-profile.mjs green + ≥2 agent APPROVE + Maintainer ratify). A failing probe blocks the 0.x → 1.0 version bump.