kadubon/collective-phase-control-fabric: CPCF 0.7.0: bounded endogenous capability frontier expansion
CPCF 0.7.0 adds Bounded Endogenous Capability Frontier Expansion: exact finite planning in which declared outcomes can enable previously latent actions in an immutable finite super-catalogue. Planning and checking The active frontier, activation depth and lineage are part of canonical state and policy digests. Exact producer/successor rules bind immutable actions, capabilities, schemas, lifecycle and policy objects. A research action can enable a verifier or another research action; a later activation can make funded growth continuation feasible. All branches retain exact rational resource, physical-time, queue and repair-debt accounting. Independent checking reconstructs activations and rejects latent use, tampered digests or witnesses, expiry, missing objects, invalid prerequisites and depth overflow. Duplicate/self-activation and zero-progress cycles earn no objective credit. Interaction-restricted comparators are reoptimized over every legally reachable remaining activation path. Search exhaustion remains unknown. Fresh evidence reassessment reuses existing admission mechanisms. Replanning produces unsigned proposals and preserves signed history, admitted lineage, original depth and prerequisite constraints. The original growth objective is unchanged. Model activation is useful only through downstream policy feasibility or attainment; frontier size is not an intelligence score. Witnesses describe selected-policy use and do not assert causal marginal contribution. Compatibility and examples Three additive closed kinds provide frontier declarations, plans and assessments. No-frontier contracts preserve 0.6.1 behavior, preceding 52 schema identities and existing plan encodings. Package interfaces report 0.7.0; protocol identity remains cpcf.io/v0.6. Eight deterministic synthetic scenarios cover research unlock, verifier investment, failed discovery, useless frontier growth, a depth-two chain, zero-progress cycles, comparator alternatives and safe fallback without guaranteed entry. For example: pip install collective-phase-control-fabric==0.7.0 cpcf self-check --json cpcf growth example --scenario frontier-chain --json cpcf growth example --scenario frontier-verifier --json The chain selects prepare → reuse → continue with activation depth two. The verifier scenario selects prepare → verifier → reuse → continue, accounting for queue pressure and verification cost before model-supported entry. Software validation Local validation: 853 tests passed and 3 existing PostgreSQL service tests were skipped without their disposable database URLs. The focused critical suite passed 538 tests. With branch measurement enabled, combined statement/branch coverage is 91.13% overall, 95.60% across the critical suite and 97.83% for endogenous-frontier code; all 14 critical subsystem gates pass the unchanged 95% floor. The activation kernel has 100% statement and branch coverage. The complete native Linux mutation run detected 10,749/12,284 mutants (87.50%), above the unchanged 85% threshold. It left 1,530 surviving mutants, 4 timeouts and 1 untested mutant, all counted as failures; none were incomplete. The new frontier modules detected 1,159/1,225 mutants (94.61%). These are software test results, not an assurance of real-world outcomes. CI and release distribute the unchanged catalogue over five disjoint shards, independently reject any missing or overlapping assignment against the reviewed catalogue fingerprint, and apply the original 85% gate to the complete union. All shard jobs must succeed before publication. Ruff, strict mypy, Bandit, OSV pip-audit, 215 schemas, 9 fixtures, generated references, 13 synthetic examples, publication hygiene, Gitleaks, distribution metadata and Twine checks passed locally. Fresh base-wheel installation verifies dependencies, all five import versions, self-check and the preparation/depth-two frontier examples. Portable CLI reference generation is also checked on Linux Python 3.12.3 and Windows Python 3.14.6. Required remote platform, PostgreSQL, security and mutation gates must pass for the release source before publication. The existing release workflow builds the distributions, produces the SBOM and checksums, attests provenance and publishes through PyPI OIDC trusted publishing. Public-index installation, all five package versions, CLI behavior, asset hashes and provenance are verified after that workflow completes. The release source is merged commit 0afced595da264c500b527533321aa50d73fc4bb. PR #10 was merged with the owner's explicitly authorized administrator exception only after all 13 required checks passed. The exact merged commit also passed main CI and security analysis. The complete PR and main mutation gates each scored 87.50% across all 12,284 declared mutants. Documentation Frontier semantics and scientific boundaries Validation record Eight synthetic results Paper-to-code mapping Wiki frontier guide README, specification, formal model, agent skill, evidence/science guidance, audit findings and generated schema/CLI/error references are updated. The Wiki includes the frontier guide and updated Home, finite planning and release pages. Beta scientific status No external empirical acceleration experiment was performed. The endogenous capability frontier is a finite model and software-control mechanism. It does not establish real-world capability reproduction, collective intelligence, AGI, ASI, causal acceleration or indefinite growth. Model enablement does not admit a capability, trust an executable, generate a receipt or authorize execution. No external operational-assurance evidence is fabricated or inferred from successful package publication. The next major formal extension is fixed-parameter, set-valued model uncertainty with observation-driven model-set contraction and prior-free value-of-information planning. It is documented, not implemented in 0.7.0.