Skip to content

ROADMAP.md

Latest commit

 

History

History
108 lines (76 loc) · 6.01 KB

ROADMAP.md

File metadata and controls

108 lines (76 loc) · 6.01 KB
Raw

Roadmap

Concrete milestones for future releases of ARC-Neuron LLMBuilder. Items are ordered by impact and by how much they preserve the governance doctrine.

Current release: v1.0.0-governed (2026-04-22)

Status: Shipped.

  • Three governed promotions on record (v4, v5, v6_conversation) at v1.0.0
  • Four post-audit governed promotions on record (v7, v8, v9, v10) at v2.0.0 — overall score 0.6836→0.9237
  • All four Gate v2 decision states fired lawfully
  • Conversation-harvested corpus proven to grow the brain
  • Omnibinary ledger measured: 6,600 ev/s append, 8,900 lookups/sec O(1)
  • Arc-RAR bundles restorable end-to-end
  • 115/116 tests passing (v2.0.0), 9/9 demo proof workflow green

v1.1.0 — Expanded Native Lane

Target: widen the native ARC-Neuron lane without relying on external backends.

Planned

  • ARC-Neuron Base tier — 4 layers, 256 block, 128 embd (~1M params). GPU-accelerated training script with CUDA + MPS detection.
  • Real tokenizer beyond byte-level — SentencePiece or BPE built from the accumulated corpus via arc_neuron_tokenizer/builder.py. Native models pick it up via TransformerConfig.vocab_size.
  • Distillation wave v2 — programmatic conversation-harvest driver that runs run_proof_workflow.py N times with varied prompts and auto-builds training packs.
  • Benchmark breadth expansion — new lanes for safety_refusal, numerical_reasoning, structured_json, tool_calling (+50 tasks).
  • Scorer v3 — per-capability weighting in overall_weighted_score rather than equal weights.
  • CLI frontend — arc command dispatching to all operator scripts (arc train, arc promote, arc gate, arc explain).

Acceptance

  • v7_native promoted through Gate v2 with cleanly higher overall score than v6_conversation using only the new tokenizer and ARC-Neuron Base tier.
  • Distillation wave v2 produces 100+ SFT pairs in one run.
  • CLI is documented in USAGE.md as the canonical surface.

v1.2.0 — External Backend Integration

Target: prove the governance machinery holds against a frontier-grade external model.

Planned

  • Reference integration docs for Qwen3-32B-Instruct, Llama-4, DeepSeek-Coder via llama_cpp_http.
  • External-model benchmark preset — 142-task suite pre-run against a canonical open-weights model as a reference ceiling.
  • Per-adapter scoreboard namespacing — prevent external-model scores from displacing native-model incumbents accidentally (keep separate tracks under one scoreboard).
  • Command adapter timeout tuning — expose first-output / idle / overall timeouts as --timeout-* flags to promote_candidate.py.
  • Reflection loop v2 — skip the critique stage when confidence is already bounded; detect overclaim patterns more precisely.

Acceptance

  • Gate v2 promotes a fine-tuned Qwen3-7B candidate through the identical pipeline.
  • Reference benchmark numbers for Qwen3-32B published in reports/external_reference/.

v1.3.0 — Multi-Repo Integration

Target: tighten the cross-repo contracts across the seven ARC repos.

Planned

  • OmniBinary ↔ LLMBuilder federation — a daemon in the OmniBinary repo that subscribes to this repo's Omnibinary ledger and mirrors into the canonical binary-runtime event graph.
  • ARC-Core event attestation — every promotion receipt co-signed with an ARC-Core signing key from the bootstrap flow.
  • Arc-RAR → Cleanroom replay — a one-command restore from any Arc-RAR bundle into a Cleanroom-runtime-governed execution slot.
  • Language Module canonicalization — sync the LLMBuilder terminology store with the Language Module's governed truth surface, replacing the current local-only JSON store.

Acceptance

  • Cross-repo integration test: one conversation turn produces a receipt that is co-signed by ARC-Core, mirrored into OmniBinary, preserved in Arc-RAR, and restorable into Cleanroom.

v2.0.0 — Production Governance

Target: institutional-grade readiness for enterprise adoption.

Planned

  • Formal governance spec — a versioned specification document (beyond the current Markdown doctrine) with machine-checkable invariants.
  • Sandboxed execution — Gate v2 runs inside an isolated environment with enforced resource limits.
  • Audit trail export — one-command export of every receipt, bundle, and decision for a date range into a tamper-evident format.
  • Per-organization scoreboards — multi-tenant scoreboards with role-based access to promotion decisions.
  • Compliance tooling — optional integration hooks for SOC 2 / ISO 27001 auditing workflows.

Acceptance

  • External audit validates the ten governance invariants.
  • One external organization runs the governed loop in production under their own compliance framework.

Ongoing

These are not tied to a specific release:

  • Documentation quality — FAQ, GLOSSARY, EXAMPLES, MODEL_CARDs kept current as the system evolves.
  • Benchmark quality — task diversity increases, rubric sophistication grows, paraphrase-pressure coverage expands.
  • Performance — Omnibinary throughput, Arc-RAR pack time, benchmark runtime.
  • Cross-repo hygiene — each of the seven ARC repos stays role-pure.

How to influence the roadmap

What won't be on the roadmap

  • Alignment or safety filtering. The gate is a capability + regression gate. Safety is an orthogonal concern best handled by the operator's policy layer.
  • Hosted cloud service. This is a local-first project. Hosted offerings are out of scope.
  • Closed-source components. MIT all the way down.
  • Role inversion. The seven-repo contract is permanent. New capabilities go into their correct home repo, not into whichever one is most convenient.