Skip to content

Roadmap

Shipped in v0.1.0

  • Sprint 9 — Probe orchestrator pilot (2 active probes, 12 TDD tests)
  • Sprint 10 Phase α Batch 1 — 3 rectifiers (HIGH confidence)
  • Sprint 10 Phase B — generalized specification bridge (1696 items)
  • Sprint 10 Phase C — probe orchestrator 3-mode rectification (observe / suggest / act)
  • Sprint 10 Phase D — persistent trace coefficient (faculty_persistent_traces)
  • Sprint 10 Phase E — coupling refactor (canonical schema + unified factory)
  • Sprint 10 Phase G — Publication polish

Shipped in v0.2.x

  • Container-first install (Docker Compose)
  • Multi-LLM provider support (LiteLLM)
  • Specification corpus shipped via package
  • Daemon idempotence audit + trajectory tooling
  • Benchmark v2 with bilingual judges, 1200 invocations

Sprint 11 — Probe orchestrator extension Batch 2

  • Transpose 7 additional rectifiers (MEDIUM confidence) from primary sources.
  • Extend probe orchestrator to cover the 7 new rectifiers.
  • Scope : approximately 49 new specification items.

Sprint 11+ — Batch 3 LOW

  • 3 remaining rectifiers. Some require philological work on alternate editions of primary sources.

Sprint 12 — Source extension

  • Transpose additional source sections.
  • Integrate with existing corpus via see_also links.

Sprint 13 — Pre-extension foundational corpus

  • Important for structural context.

Sprint 14 — Cross-source coherence

  • Tests for convergences and divergences across primary sources.

Adversarial counterpart (deferred)

The design review for an adversarial counterpart to the probe orchestrator found no direct attestation in the primary sources. Possible alternative scopes to be decided :

  • Channel failure detector (4 modes : closed / misoriented / insufficient receiver / exposed back) — derivation from authoritative reading.
  • Integration with existing adversarial tests (malakhim/adversarial/) without a new engine.
  • Skip entirely ; the existing qualification testing framework already covers adversarial validation.

Longer term

  • Complete implementation of all 10 cognitive faculty modules with 4-level qualification tests.
  • Synthesis bridge (cross-faculty insight ↔ causal reasoning).
  • Full daemon integration for all hourly and daily rectification cycles.
  • Academic paper describing the probe orchestrator auto-rectification mechanism.

v1.0 milestones

  • arXiv paper "A Cognitive Operating System for Structurally-Evolving LLM Agents"
  • Cognitive OS Evaluation Suite (8 metrics) outperforms baselines (LLM seul, LangChain, AutoGen) on 6/8 with p<0.05, reproducible across 4 LLMs
  • Repo public stable, pip + docker
  • Discord 100+ members, 3 external contributors, 1 third-party reproduction of the benchmark

Sprint cadence

Sprints are not time-boxed. Each Sprint ends when its DoD (Definition of Done) is met and all non-regression tests are green. Historical cadence has been one Sprint per 1-2 weeks of active work.