AT Generation Procedure (RENAR §8A)

Agent-facing specification — consumed by AI assistants, not end users. No RU mirror is produced (see docs/README.md § Internal agent specs).

RENAR §8A (ADR-012) closes the one class of defect traceability (TC → SR → ADAPT → ТЗ) cannot catch: a wrong interpretation makes every TC pass because the system perfectly matches the wrong reading, and fails acceptance at the client. An AT (Acceptance Test) is the check that can still catch it — but only if it is produced the way the standard requires. tausik at createrecords an AT; it does not generate one. Generation is this procedure, followed by whichever agent is orchestrating the work.

The three properties that make an AT valid

1. Derived by an isolated agent from the final TZ alone. No ADAPT, BR, SR, SPEC, TC or code in its input. An agent that has seen the interpretation reproduces its mistake — isolation here is a generation mechanism, not a courtesy.
2. Regenerated before every trial from the current edition. A system tested against a year-old contract at the end of a long engagement is not tested against the contract. tausik at check-freshness names any AT whose source_as_of no longer matches the live final_tz_snapshot for its tz_ref.
3. tz_text is a verbatim quote, not a paraphrase — the field that ties an AT back to what the client actually signed, independent of whatever the scenario prose says about it.

Procedure

1. Pull the current reference, nothing else.

   tausik actz final-tz

   or tausik_actz_final_tz over MCP. For the tz_ref you are generating an AT for, take exactly two fields: governing_text (becomes tz_text, verbatim) and completed_at (becomes source_as_of).

2. Spawn a genuinely isolated agent. Use the Agent tool with a general-purpose subagent and a self-contained prompt containing only:

   • the tz_ref and the verbatim governing_text from step 1;
   • a generic instruction to write one or more acceptance scenarios that verify a system satisfies that clause, in plain language a client could read.

   Do not paste ADAPT interpretations, SPEC content, existing TC, or code into the prompt. Do not summarize "what we built" first — the isolated agent must not know. If the orchestrating session has already discussed the implementation of this clause in the current conversation, that context must not reach the subagent's prompt.

3. Transcribe, don't edit. The orchestrating agent records the isolated agent's output as-is:

   tausik at create &lt;slug&gt; &lt;tz_ref&gt; "<verbatim tz_text>" "&lt;scenario&gt;" \
     --as-of <completed_at from step 1> --by <this session's identity>

   Editing the scenario for style is fine; changing what it checks is not — that reintroduces exactly the contamination isolation exists to prevent. generated_by names the orchestrator (this session), not the isolated subagent — the isolated agent has no access to at create by design, and the field records who is vouching that the procedure was followed, not who wrote the words.

4. Regenerate before trial, not once. Run tausik at check-freshness before any acceptance run. A stale AT (its tz_ref's governing point changed since generation) must be regenerated via steps 1–3 before the trial counts as evidence. The at_freshness gate (warn) surfaces this at task-done automatically; it cannot regenerate anything itself.

What this procedure is not

It is not a way to generate AT content for this project's own SPECs as part of routine development — that is a separate, larger effort building on this mechanism, not a step of it. This document describes the mechanism only.