Skip to content

P3_PRODUCT_DECISION.md

Latest commit

 

History

History
146 lines (95 loc) · 6.13 KB

P3_PRODUCT_DECISION.md

File metadata and controls

146 lines (95 loc) · 6.13 KB
Raw

ExecSurface — P3 Product Decision

Date: 2026-09-29 Tracking: #103 / #105 Parent program: #100 Branch: development/post-alpha4-behavioral-integrity Status: CLOSED — P3_EPHEMERAL_ONLY_VALUE_ESTABLISHED

Decision

P3_EPHEMERAL_ONLY_VALUE_ESTABLISHED

P3 has established bounded value for an evidence-backed, producer/role-constrained ephemeral identity projection. P3 has not established product value for automatic multi-run variance authorization.

No public integration, release, stable-tag movement or public-v2 reinterpretation is authorized by this decision.

Team and review boundary

Fixed roles throughout the closing gates:

  • Innovation Scientist / Systems Architect
  • Anti-Drift / Scientific Integrity Reviewer
  • Independent Falsifier / Red Team
  • Independent Milestone Reviewer

Closing specialist set:

  • Go/GCC nondeterminism
  • Linux ptrace clone/fd lifecycle
  • canonicalization / baseline / diff semantics
  • Rust research-tool correctness
  • provenance / artifact integrity
  • supply-chain false-PASS analysis
  • CI reproducibility

Evidence chain

V2 bounded grammar

The accepted GCC candidate is not path-prefix suppression. Eligibility requires the frozen evidence-backed producer/role contract including canonical temp identity, GCC actor/family, create/write-capable open, delete and causal-chain correlation. Broad temp/cache/source whitelisting remains forbidden.

V3 accepted-variance contract

The research contract kept invariant, observed-variable, explicitly accepted-variable and unseen effects distinct. Frequency/recurrence never grants authorization.

V4 poisoning/falsification

The original adversarial suite closed P3_V4_FALSIFICATION_PASS_BOUNDED with 10/10 attacks blocked as designed.

Original V5 negative evidence

The first real-workload campaign remains permanently P3_INCOMPLETE_EVIDENCE; its failed samples were never replaced or relabeled. Shared-FD incompleteness was subsequently isolated to an evidence-retention/certification boundary.

C1–C6R observer-correctness bridge

The research path established bounded internal clone/fd-table certification without changing the public/default raw-v2 contract. Direct default integration was explicitly rejected as C6_COMPATIBILITY_REGRESSION. The accepted certificate-aware behavior remained research-only under C6R_ISOLATED_RESEARCH_GUARD_PASS_BOUNDED.

Corrected R2-C / R3-C

A fresh six-run learning campaign with normalization matched to public CLI default semantics produced complete, warning-free, certificate-positive evidence.

R3-C metrics:

  • raw variable candidates: 114
  • targeted GCC variable candidates: 12
  • non-target variable candidates: 102
  • projected variable candidates: 102
  • lingering targeted random identities: 0
  • explicitly accepted variable effects: 0
  • non-target support/provenance mismatches: 0

Decision: P3_V5_REAL_WORKLOAD_VALUE_REQUALIFIED_BOUNDED.

The earlier zero-target result produced under the wrong normalization profile remains retained as invalid-comparability negative evidence and is not used for the final product decision.

R4 six-check requalification + sentinel

Exactly six fresh unchanged checks were admitted with no replacement.

Every check reproduced four eligible GCC ephemeral diff findings, and the bounded projection removed the randomized GCC identity component while preserving non-target effects.

The explicit accepted-variable set stayed empty. The multi-run/explicit-variance layer did not reduce residual findings below the ephemeral-only projection on any check.

The preregistered meaningful-drift cat sentinel remained unseen/unaccepted and survived the GCC projection.

Decision: P3_R4_CHECK_SET_COMPLETE_SENTINEL_PROTECTED.

R5 independent falsification

Accepted R5:

  • workflow 36578564586
  • source ed5ca6bc07273ade4864eba821f0dc55e10f76a8
  • artifact 11038625288
  • artifact SHA-256 sha256:fe76d24e44ec6324f8aeceecd67e04f8798ee0670722b9c3828b1a5717487c2c

Results:

  • immutable evidence verification PASS
  • intentional corruption rejection PASS
  • R4 sentinel replay PASS
  • original V4 falsifier replay 10/10 PASS
  • new current-boundary R5 falsifier 4/4 PASS
  • public M11 fail-closed reproof 6/6 PASS

Decision: P3_R5_FALSIFICATION_PASS_BOUNDED.

Product interpretation

Established

The evidence supports retaining the narrow architectural mechanism:

evidence-backed ephemeral identity classification -> bounded canonical projection

for cases where the producer/role/grammar/provenance contract is independently established and falsified.

This mechanism can reduce identity-only nondeterministic noise without silently authorizing unrelated variable behavior in the declared test scope.

Not established

P3 does not support promoting any of the following as product semantics:

  • frequency-based authorization
  • automatic acceptance of recurrent variable behavior
  • broad $TMP, cache, module, stdlib, workspace or source-tree suppression
  • the 102 observed non-target variable candidates
  • generic multi-run variance as a PASS-producing mechanism
  • public certificate-aware ptrace completion under raw v2

The multi-run analyzer and accepted-variance contract remain useful research/falsification tools, but they are not promoted to automatic product authorization from P3.

Integration boundary

P3 closes at the research/product-decision level only.

A later release/integration gate may evaluate a narrow implementation of the bounded ephemeral classification/projection together with the required evidence contract, compatibility/migration analysis and public fail-closed regression suite.

Until that separate gate closes:

  • v0.1.0-alpha.4 remains unchanged;
  • stable @v0.1 remains unchanged;
  • main receives no P3 runtime integration;
  • certificate-aware completion remains research-only;
  • P3 research failures remain part of the permanent evidence record.

Next program gate

Per #100, P3 closure authorizes movement to:

P4 — Backend Adapter / Proposition-Authority Architecture

P4 must strengthen proposition-specific authority without rebuilding a generic host-security collector and without assuming backend interchangeability before proposition-by-proposition equivalence is demonstrated.