Skip to content

CHANGELOG.md

Latest commit

 

History

History
72 lines (62 loc) · 5.29 KB

CHANGELOG.md

File metadata and controls

72 lines (62 loc) · 5.29 KB
Raw

Changelog

[unreleased] — overnight CPU-first build

Added — foundation (chief engineer / Opus Max)

  • Boundary block (BOUNDARY_BLOCK) and verifier (verify_boundary, check_fusion_intent).
  • Canonical-JSON serialiser + content-addressable sha256: IDs.
  • UniversalLayerEnvelope + DeviceResponseObject + CrossModelDisagreementRecord
    • EarlyWarningSignal + SourceManifest + ReasonerTuple Pydantic-v2 schemas.
  • AuditWriter (JSONL + DuckDB) with mandatory boundary check + provenance fields.
  • KGStore (JSONL + NetworkX MultiDiGraph) with the PRD's 14 node types and 12 edge types, plus GraphML export.
  • L6 control plane: EnergyConfig (ENERGY_* env flags), AdapterRegistry seeded with 17 adapters across both sub-verticals, falsifier router with 4 generic falsifiers.
  • FastAPI REST stubs for L1-L5 endpoints (electrochem + fusion) with config-flag cutover scaffolding for Runpod backends.
  • Typer CLI: health, boundary, registry, smoke, serve-rest, falsification-wave, electrochem-e2e, fusion-phase0.
  • Foundation contract tests: 61 tests covering schemas, audit, KG, L6, REST stubs, plug-replaceability, canonical JSON.

Added — adapter stacks (parallel subagents)

Electrochem L1-L5 (Sonnet subagent — DONE)

  • L1: ElectronicStructureAdapter (PySCF 2.13.0 H2 RHF/STO-3G real CPU; Marcus / optical-spectrum analytic fixtures).
  • L2: MLIPManifestAdapter (manifest-only with license gate); deterministic synthetic MSD.
  • L3: parser-first VTK/HDF5 manifest with synthetic tortuosity / effective diffusivity.
  • L4: PyBaMMBatteryAdapter (DFN P2D Chen2020 real CPU), SolcorePvAdapter (Solcore failed pip on 3.13 → analytic Shockley-Queisser fallback), CanteraSofcAdapter (GRI30 real CPU), PemAdapter (analytic Butler-Volmer; AlphaPEM GPL-3 isolated), ThermoelectricAdapter (analytic).
  • L5: PyPSALcoeAdapter (PyPSA 1.2.0 + HiGHS real CPU + 200-sample MC LCOE), PvlibYieldAdapter (pvlib 0.15.1 Ineichen Sandton clear-sky), PySAMLcoeAdapter (analytic).
  • 29 tests green in 5.8s.

Fusion L1-L5 + reasoning bench (Opus subagent — STALLED mid-test-write; orchestrator finished tests)

  • L1: OpenMcManifestAdapter (CPU-fixture transport + nuclear library manifests for ENDF/B-VIII.1, JEFF-4.0, JENDL-5, FENDL-3.2c, TENDL, IRDFF-II with sha256 placeholders).
  • L2: TglfReducedAdapter (analytic gyroBohm reduced transport), CgyroNonlinearAdapter (REST-stub Runpod-parked), GyroSwinSurrogateAdapter (REST-stub).
  • L3: FreeGS4eAdapter (freegs 0.8.2 diverted equilibrium real CPU), JorekDryRunAdapter, BoutDryRunAdapter (parser-only).
  • L4: imas_fixture.write_fixture/read_fixture (netCDF4 IDS-shaped equilibrium + core_profiles fixture, COCOS=11, DD=3.41.0), ImasPythonAdapter, OmasConverterAdapter, ReducedTransportCpuAdapter (0D ITER H98(y,2) scaling — emits envelope + DRO), DuqtoolsConfigAdapter.
  • L5: ParamakGeometryAdapter (Paramak when installable; CSG fallback), OpenmcCsgFixedSourceAdapter, OpenmcR2sAdapter (analytic activation stub).
  • FusionReasoningBench: 50 tasks across 5 categories (imas_ids, q_profile, disruption, blanket_tbr, forbidden); rules-based scorer; refusal_recall=1.0 on the 10 forbidden tasks.
  • 20 tests green in 15.7s.

MCP server suite (Sonnet subagent — DONE)

  • mcp 1.27.0 FastMCP installed; 9 servers, 9 tools.
  • Servers: pybamm_mcp, pvlib_mcp, solcore_mcp, cantera_mcp, pypsa_mcp, pysam_mcp, openmc_mcp, imas_codex_mcp (read-only), aiida_mcp.
  • Boundary block in every tool description; fusion intent gate on fusion tools.
  • 33 tests green in 3.6s.

TDA + 12-test falsification wave (Sonnet subagent — DONE)

  • ripser+persim CPU detector with Takens embedding; persistence entropy + max H0/H1 lifetimes.
  • Cross-domain pre-canned configs for battery thermal runaway, fuel-cell membrane breakdown, electrolyser stack degradation, SOFC delamination, plasma disruption.
  • No-leakage guard (no future leakage, pulse-level split, normalisation fitted on train only).
  • 12-of-12 falsification wave passes (boundary mutation, license promotion, stub validity, units, COCOS, T<0, n<0, fill-factor>1, above-Carnot, SoC out-of-range, missing IDS, cross-model disagreement).
  • 63 tests green in 7.0s.

Source manifests + reasoner curator + decision log (Sonnet subagent — DONE)

  • 41 source manifests covering every tool referenced in the PRD (PySCF through JEFF-4.0).
  • 41 license findings with per-tool class A-E + SPDX + verdict + promotion status.
  • SourceLog, license_gate, reasoner_curator helpers.
  • 4 decision docs: 000-overview, 001-license-policy, 004-data-sovereignty, 005-deviations-from-prd.
  • 63 tests green in 1.2s.

Added — handoff

  • BOUNDARY.md — verbatim boundary block + fusion specialisation rule.
  • scripts/full_check.sh, scripts/clean_runtime.sh, scripts/quick_demo.py.

Notes

  • Python 3.13.12 venv (system 3.9 supported but PyBaMM/Cantera prefer 3.11+).
  • All artifacts carry the boundary block verbatim. Mutation -> boundary fails closed.
  • No bulk datasets vendored. All external sources are manifest-only with sha256 placeholders until a separate verification wave fills real checksums.
  • Class C/D/E backends gated: scientific mode requires kg://license-grant/* or https:// or file:// evidence URI.
  • Stubs cannot set scientific_valid=True (validator-enforced).