P8-A3.3 — experimental same-run typed evidence output implementation #153

Status

OPEN — EXPERIMENTAL_TYPED_EVIDENCE_OUTPUT_IMPLEMENTATION_GATE

Parent design: #151 / PR #152
Parent prototype: #149 / PR #150
External evidence authority: #114
Current main at gate open: 28f92d844dea34aca163b92b66b5d39ecde15db6

Authorized experimental surface

Implement only:

execsurface observe --evidence-output PATH -- COMMAND [ARGS...]

The flag is experimental/non-stable until a later promotion decision.

Invocation without --evidence-output must preserve current behavior.

Same-run binding decision

The evidence file will be a single envelope containing:

1. the original raw observation object from the same in-memory BackendObservation;
2. a typed projection derived from that exact object and internal collection-health envelope.

This avoids:

• a second workload execution;
• user-supplied binding identities;
• dependence on target stdout for structured evidence;
• a separate canonicalization/digest protocol merely to prove same-run association.

No new acquisition is authorized. The raw object is data ExecSurface already produced.

Candidate experimental envelope

Separate schema identity/version from raw observation and verdict schemas.

Required top-level classes:

• schema_version;
• report_kind;
• stability = experimental;
• backend;
• collection_health;
• unsupported_capabilities;
• bounded effects;
• limitations;
• does_not_assert;
• raw_observation.

No PASS/REVIEW/BLOCK field.

Gate chain

E0 — compatibility freeze

Before implementation, preserve current observe -- COMMAND output fixture and no-flag CLI behavior.

E1 — same-run envelope

One internal BackendObservation instance must supply both raw and typed sections.

E2 — typed semantics

Bounded fd-write effect may claim:

• runtime-fd path correlation;
• post-syscall successful positive-byte operation semantics;
• state-machine correlation where justified.

It may not claim exact byte count/content/final state/source-code provenance.

E3 — fail-closed attacks

Must test:

• incomplete/warning laundering;
• backend/profile inflation;
• missing capability;
• process binding mismatch;
• repeated effects;
• forbidden fields;
• output-path failure;
• unsupported option combinations / malformed CLI use.

E4 — privacy

Existing privacy sentinel must remain green. Evidence file must not add acquisition.

E5 — public-compatibility regression

Required:

• full CI;
• P9.3 compatibility;
• adversarial regression Ubuntu 22.04 + 24.04;
• legacy no-flag observe output unchanged for the frozen fixture.

E6 — Probity-style consumer replay

A test must demonstrate that one evidence file provides:

• original raw trace;
• typed collection health;
• successful-write semantics;
  while leaving state snapshots/authority/signing/publication policy outside ExecSurface.

E7 — decision

Choose exactly one:

• EXPERIMENTAL_TYPED_EVIDENCE_OUTPUT_IMPLEMENTATION_ACCEPTED
• IMPLEMENTATION_RESEARCH_ONLY_MORE_WORK_REQUIRED
• IMPLEMENTATION_KILLED_COMPATIBILITY
• IMPLEMENTATION_KILLED_AUTHORITY_RISK

Explicitly not authorized

• GitHub Action changes;
• changing check report schema;
• replacing observe stdout;
• stable-v1 promise;
• public release/tag;
• P8 A5 closeout;
• production-adoption claims.