Skip to content

roadmap.md

Latest commit

 

History

History
98 lines (85 loc) · 7.02 KB

roadmap.md

File metadata and controls

98 lines (85 loc) · 7.02 KB
Raw

Roadmap — Townlet · Updated 2026-10-01

This records product intent, not a delivery schedule. Workstream numbers are an inventory, not an execution order. One bounded recovery unit is selected at a time; forecasts and program sequencing require their own evidence. See current state, metrics and PDR-0154.

Now — continuing recovery

Strangler rewrite behind the compiled-universe contract remains the committed bet (PDR-0006). Main consolidation is publication inside the bet, not its exit. WS-1 and WS-7 are closed; WS-0/2/3/4/5/6 remain open. Tracker dependency path remains WS-6 → WS-2 → WS-3 → WS-4. The documentation-replacement program remains separately gated under PDR-0125; this factual checkpoint and unconditional merge README review do not complete it.

Delivered within the bet:

  • Token observation engineering, all five milestones, including compact replay, declared attention, recurrent learning machinery, shipped-pack migration and historical M4 qualification (PDR-0133 through PDR-0144). M4's four passing cells are one versioned engineering cohort, not a fresh general convergence claim.
  • Declaration-store Cut A: files are transport, declaration discovery/merge/provenance and one authored clock authority (PDR-0149). Cut B: canonical variables, explicit lifecycle, supported symbols and typed token scopes (PDR-0151). Their original planning is complete.
  • Static epistemic access: required finite roles, checked runtime/publication, immutable policy, attempted-write authorization and policy-bearing persistence/identity (PDR-0153). Permission authoring is no longer absent. Custom action writes are declared and exercised by the committed ADVANCE witness. No old profile/overlay authoring reader is preserved.

Exit review required. PDR-0058's trigger fired as DIV-013/014/015 were added without an entry becoming terminal across successive checkpoints. Keep the oracle and the current non-exit status; adjudicate the exit/instrument framing before treating the next unit as progress toward a settled retirement criterion. Qualification of a registered divergence is not a terminal disposition. Do not widen allowances or change frozen inputs to conceal it.

Recommended next unit: truthful episode lanes. Reproduce current main starting with hamlet-d6fc84d147, then scope a lifecycle/accounting PRD. Acceptance should share active-on-entry and newly-terminal events across runtime, replay, exploration and output sinks: one terminal transition/finalization, frozen post-death outcomes, exact live-transition versus vector-tick units. This is a recommendation under the compile → render → converge objective; no new implementation, deadline or training campaign is authorized by the checkpoint.

Next — shaped, separately scoped

  • Reset isolation: active effects must not survive a new episode (hamlet-d76684f549). Use actual mutation/reset witnesses and explicit callback/persistence semantics. Keep this out of the episode-lane package unless an evidenced dependency requires it.
  • Item fidelity: initial appearances must survive cold/cache artifacts; carried items, type identity, locality, capacities and per-world mutable state need coherent contracts. Held-item issue hamlet-4b931faaf4 remains open. Item-bearing scenarios must qualify these seams rather than relying on static role authorization alone.
  • Experiment-control and viewer truth: replay warmup, evaluation cadence, checkpoint publication/retention, RNG isolation, resume boundaries and error propagation must match declarations. The browser must identify waiting/baseline/loaded/stale/failed models and actual evaluation evidence; a spent budget is not convergence.
  • A genuinely different domain witness: author one scoped task through the normal compiler/BAC path, without privileged Python. Predeclare task success, attainable scripted behavior, baselines, seeds, budgets and checkpoint-selection rule; render the exact checkpoint that passed independent evaluation. Reconcile default scenario descriptions and intended brains rather than treating their old names as distinct-world evidence.
  • Close remaining authoring gaps (WS-4): use an updated capability/consumer inventory, not the old approximately-two-of-seven fraction. Canonical variables, static permissions, custom writes and token configuration are delivered slices, while reward/lifecycle, spatial edge semantics, curriculum and broader world authoring still require inspection. A low-level class, parsed setting or rejected feature is not a completed authored capability.
  • Dynamic epistemic propagation (PDR-0120): shaped and unbuilt. Define observer/source visibility, ownership, brightness/floors and world axes separately from delivered static engine/agent roles. hamlet-83a043a9b9 remains the broader privacy boundary; static permission acceptance does not automatically schedule this extension.
  • BAC Layers 1 and 3: topology, personality, goals, panic/compliance, composable faculties and think-loop graph remain unbuilt. Supported Q-network configuration and recurrent BPTT are useful Layer-2 machinery, not a complete declarative mind.

Later — directional, no dates or implied authorization

  • Model plus observation/action contract export into another engine or pipeline (hamlet-0cdb8a6d1a); first-class world/brain bundles, explicit transfer versus exact resume.
  • Tech-demo suite and locked showcase experiments, preserving Townlet Town as a first-class demonstration and the historical six-dimensional substrate witness (PDR-0003/0025/0026). Distributable showcases and releases need their separate owner boundary.
  • Tick-level governance proof, exact continuation, run lineage and evidence custody.
  • Authoring workbench: templates, scaffolding, diagnostic navigation and live preview, shaped for the intended researcher/prototyping audience rather than an invented user base.
  • Recording/replay: preserve intended capture, playback and export capability before any deletion (hamlet-16ae192d42, PDR-0007); decide its supported/replacement path explicitly.
  • Murk as the intended CPU/RAM substrate below Townlet. Qualify state/action/observation, reward/reset/identity and PyTorch boundary parity; CPU execution is deliberate. Migration is not a prerequisite for proving the current narrow Townlet slice.
  • Wardline adoption as a real declared-boundary hygiene gate; external adoption readiness after authoring evidence and documentation truth. No outward release is implied.

Retained decision history

Earlier detailed unit histories remain in their dated PDRs and metrics. Presentation/name-blind observation units (PDR-0066/0069/0075/0076), per-level brain/lineage ruling (PDR-0027), token milestones (PDR-0132 through PDR-0144), fifth main merge (PDR-0146), A/B (PDR-0149/0151) and static access (PDR-0153) are accepted history, not current dispatch instructions. The retired authoring-trial corpus publishes no revived success rate (PDR-0111/0142).