[jithinAB]

Law 8 says anything a run creates passes the same checker before it can be asked, and names three such things: a run-code snippet, a narrowed-new helper and a skill a run wrote (97 §4, law 8). The reference checked a learning proposal's instructions: (check_instructions) and none of the three. This adds the documents they are checked as, so a runtime that instantiates them checks them the way this one checks a proposal, and two runtimes check the same document.

What changes

• checking.with_helper(document, starter, instructions, description=, uses=): the starter's own entry under <starter>-helper (the next free -2, -3… when taken), without its name:, told instructions, on the starter's model. uses may narrow it; a name the starter lacks raises NotNarrower with delegation.beyond_its_starter's sentence, so a model reads the same refusal from either.
• checking.with_skill(document, agent, name, description, content, use_when=): a skills: entry, named in the agent's uses: when the agent writes one.
• checking.with_program(document, language, code, description=, runs_for_at_most_ms=): a nondeterministic program under run-code-<first 12 hex of the code's sha256>, its fuel the run's own ceiling and stop-and-say-so. PACT records a program's body and never opens it, so the code is not part of what is checked. A language with no engine raises Uncheckable.
• checking.literal(text, document): words a model wrote are checked literally. Under the default template language {{ is written \{{; under template-language: jinja the words go in raw blocks, an {% endraw %} inside them written between two blocks as a string.
• check_document on the result is the check, as for a proposal.

Evidence

• tests/test_what_a_run_instantiates_passes_the_same_checker.py (6 tests, real checker): a helper of examples/kernel/phases passes and is its starter told something else; a narrower helper passes and a wider one is refused in beyond_its_starter's words; the next free key; a written skill passes and one whose description says nothing is refused; a snippet is a program with the run's ceiling, and a language with no engine is uncheckable; both template languages read model words literally.
• The same rules, in TypeScript, check the helper, skill and program of every runnable agent of this repository's examples and test trees against the pinned checker in Bud Meta Harness (packages/pact/pact-runtime/src/instantiate.ts, packages/pact/pact-loader/tests/instantiate.pact.ts).
• The full adapter suite on this branch: 3127 passed, 4 failed, of which one fails on master too (test_the_second_port_reports_a_spend_cap_it_cannot_price) and three were this change's own bookkeeping, fixed in the same commit: the headline test count (now 4923: 1642 Rust + 3281 adapter) and the host-API rows (with_helper, with_skill and with_program added; delegation.beyond_its_starter is reached now, so its row goes).

Nothing changes in the specification or the crates.

🤖 Generated with Claude Code

[jithinAB]

Marked ready and merged on the owner's instruction (BMH lane pact-kernel). BMH's PK-13b holds what a run creates (helpers, skills, programs) to this check; merging it lets the next re-pin retire BMH's own copy of the instantiation rule.

• Head ddea93f merges master f5d4d05 (The TypeScript port fences words nobody reviewed as the reference does (law 7) #19). gate/everything passes on this head (run 43561579253).
• gate/newest-pydantic-ai fails with the same 5 tests/test_parity.py cases as master (the S4 registry rows for capabilities released after the pinned pydantic-ai, and O7-1); this branch does not touch that path.