Yao Reference Implementation Verification v0.1

Status: Draft implementation evidence

Steward: Newvar

Last updated: 2026-08-28

Chinese translation: zh-CN

1. Purpose

This document maps the Yao v0.1 Draft specifications to automated evidence in the Morphz reference implementation. It is an implementation verification record, not a conformance certificate and not a substitute for the normative specifications.

The test boundary deliberately includes more than successful examples. It exercises invalid and adversarial source, static authority rejection, deterministic identity, serialization and resume, partial parallel completion, failure aggregation, forged Host values, database migration, and legacy-source rejection.

2. Automated evidence

Requirement area	Principal automated evidence	Status
UTF-8 syntax, precise spans, escapes, comments, malformed input, source and nesting limits	yao::syntax::tests	Passing
Parser robustness	4,096 deterministic generated short inputs in arbitrary_short_inputs_never_panic	Passing
Canonical source and typed identity	yao::canonical::tests	Passing
Named records/unions, exhaustive match, collections, Option, Result, Ref, Program	yao::sema::tests, yao::eval::tests	Passing
Static effect inference, effect ceilings, Tool scope, Host effects, effect-normal form	yao::sema::tests, sexpr_eval::tests	Passing
Pure evaluation, checked decoding, overflow/failure paths, branch scope and deterministic par order	yao::eval::tests, sexpr_eval::tests	Passing
One unversioned typed source language and rejection of (version ...)	typed_source_is_the_only_public_source_language_and_rejects_version_forms and migration-only legacy fixtures	Passing
Exact type names and rejection of the historical lowercase text / json aliases	rejects_historical_lowercase_type_aliases, rejects_unknown_and_recursive_named_types	Passing
One shared model-visible Yao Language Card with a size budget	language_card_is_parseable_bounded_and_unversioned, Context protocol and eval Tool contract tests	Passing
Typed inference request/response, continuation serialization and exact resume	sexpr_eval::tests, plan_execution::tests	Passing
One complete typed BODY under either eval or infer, without task/evidence lowering	eval_and_infer_share_one_complete_typed_body, complete_yao_body_is_handed_to_the_model_without_task_request_lowering, eval_runs_a_submitted_program_and_hands_infer_back_to_the_model	Passing
Rejection of the removed fixed-field infer forms (task, evidence, tools, and model) at both evaluator roots	fixed_infer_request_syntax_is_rejected_at_every_root	Passing
Source-authorized lexical disclosure: only (captures ...) bindings cross the provider boundary; implicit Runtime and unlisted locals do not	nested_model_body_captures_only_explicit_parent_bindings, complete_yao_body_sends_only_source_authorized_lexical_captures, infer_discloses_only_source_authorized_parent_bindings_to_the_model	Passing
Durable infer Tool scope is derived from statically visible BODY calls and never exposes eval	plan_infer_tool_scope_never_inherits_parent_tools_implicitly, infer_may_gather_evidence_but_is_never_offered_eval, plan_infer_handoff	Passing
Durable par children, branch isolation, all-terminal barrier, ordered join, aggregate failure and restart	typed_par_* tests in sexpr_eval and plan_execution	Passing
Program Value raw-Yao transport, eval/infer root admission and owner dispatch, canonical hash, effective-effect/output bounds, caller-local isolation, durable child execution and shared depth budget	program_admission_requires_raw_yao_*, infer_root_program_value_*, program_*, generated_program_*, and nested_program_* tests	Passing
Typed Harness entry, quote-preserving canonicalization, explicit model Tool upper bound and actual Function Calling narrowing	typed_program_string_identity_survives_package_normalization, model_owned_entry_requires_an_explicit_tool_upper_bound, Harness and Orchestrator tests	Passing
Runtime environment injection and typed optional projections	runtime_context_is_injected_*, host_view_normalizes_*	Passing
Host receipt replay, candidate sealing, reference non-forgeability, same-Context Evidence and Objective authority	Host tests in plan_execution::tests and candidate tests in yao::eval::tests	Passing
Sealed typed ContextTransaction, Context Authority commit, deterministic identity, and crash-window replay	context_transaction_is_sealed_canonical_and_host_typed, evaluates_and_revalidates_sealed_context_transactions, typed_context_proposal_commits_once_and_recovers_the_commit_window	Passing
Applied Objective progress, wait, and completion operations through Objective Authority	objective_wait_proposal_uses_authority_and_replays_without_a_second_transition, objective_completion_proposal_consumes_committed_outcome_and_replays_intent	Passing
SQLite Plan wait migration, indexes, terminal fence trigger and restart lifecycle	SQLite migration and memory::sqlite::plan_execution::tests	Passing
Production Runtime, CLI, durable handoff, store conformance, evaluation package and vector extension regressions	the package and integration gates listed below	Passing

3. Reproducible gates

The 2026-08-21 verification run used:

cargo test -p yao-lang --offline
cargo test -p morphz --lib --offline -- --test-threads=1
cargo test -p morphz --bin morphz --offline -- --test-threads=1
cargo test -p morphz --test attempt_loop --offline -- --test-threads=1
cargo test -p morphz --test cli_contract --offline -- --test-threads=1
cargo test -p morphz --test objective_group_handoff --offline -- --test-threads=1
cargo test -p morphz --test plan_infer_handoff --offline -- --test-threads=1
cargo test -p morphz --test runtime_stability --offline -- --test-threads=1
cargo test -p morphz --test runtime_store_conformance --offline -- --test-threads=1
cargo test -p morphz --test terminal_handoff --offline -- --test-threads=1
cargo test -p morphz-evals --lib --offline -- --test-threads=1
cargo test -p morphz-memory-vector --offline -- --test-threads=1
cargo clippy --workspace --all-targets --offline -- -D warnings

The language crate completed 38 tests. The gates above completed 1,160 tests with no failures; six tests that explicitly require a live external login or manual terminal inspection remained ignored by their existing declarations. The shared Language Card also passed its hard 4,800-character artifact limit and 1,200-estimated-token Context Encoding limit.

3.1 Complete-BODY infer and capture-boundary delta (2026-08-28)

The implementation delta that made eval and infer share one complete typed BODY and added the source-level (captures ...) disclosure boundary was verified with:

cargo test -p yao-lang
cargo test -p morphz --lib -- --test-threads=1
cargo test -p morphz --test attempt_loop
cargo test -p morphz --test plan_infer_handoff
cargo clippy -p yao-lang --all-targets -- -D warnings
cargo clippy -p morphz --all-targets -- -D warnings
cargo fmt --all -- --check
git diff --check

The run completed 47 Yao tests, 1,028 Morphz library tests, 74 production attempt-loop tests, and four durable infer-handoff tests without failure. Six library tests retained their declared ignored status. The two Clippy gates, formatting check, and diff check passed. Two live GPT-5.6 Sol smokes additionally exercised a complete pure BODY and a complete BODY containing an ordinary Tool call; their non-secret evidence is recorded in yao_infer_complete_body_live_verification_2026_08_28.md.

Tests that bind local mock servers or exercise Morphz's own macOS sandbox must run outside a restrictive parent sandbox. Running them inside another sandbox fails at operating-system setup with Operation not permitted before the test subject executes.

4. Draft gaps

The following normative targets do not yet have complete implementation evidence and therefore remain Draft:

• parent cancellation propagation through every non-terminal par and Program child;
• injected and authorized ExecutionTargetView behavior;
• live PostgreSQL crash-window testing for each new Yao wait boundary;
• persistent fuzzing of semantic decoding, canonicalization, and Program admission beyond the deterministic parser robustness corpus;
• interoperability with an independent Yao implementation.

No release may claim complete Yao v0.1 conformance while these gaps remain or without a separately published conformance suite and versioned evidence bundle.