Current Status

Generated reference. This page is published automatically from the version-controlled repository. Canonical source: docs/ROADMAP.md. Do not edit this wiki page directly.

Nemosyne Roadmap

Canonical implementation-status and execution index. Product/research direction remains governed by Nemosyne Definitive Vision and Roadmap. This file deliberately contains only current state, dependency order, ownership constraints and links to bounded work documents. Historical detail lives in the roadmap historical ledger.

Roadmap hygiene contract

This roadmap is an index, not a diary or design document.

1. Keep docs/ROADMAP.md below 20,000 characters. CI enforces the cap.
2. New non-trivial work gets a focused file under docs/work/ (or an existing scoped programme/RFC document) containing purpose, scope, evidence, exit/STOP criteria and dependencies. The roadmap gets one concise row linking to it.
3. Do not paste PR changelogs, incident narratives, experiment transcripts, implementation detail, completed checklists or historical rationale into this file.
4. When work completes, update its task/spec with final evidence, reduce/remove its live roadmap row, and append a concise completion record to the single historical ledger. Do not create another roadmap-history file.
5. A roadmap status line states only status + next dependency/action + link. Mutable review-finding status remains generated from its ledger.
6. Agents must read this contract before editing the roadmap. If detail does not change sequencing or current status, it belongs in the linked document, not here.
7. Prefer small task documents, normally one concern and well under 150 lines. Split work when a file starts mixing independent experiments or promotion gates.

Generic engineering/authority rules remain in AGENTS.md. The historical ledger preserves the pre-compaction 259k-character roadmap verbatim, so compaction is not information deletion.

Status snapshot - 6 October 2026

Integration base: main@661bd2a8.

• Architecture 2027 decisions are integrated: A27-0 and A27-1 accepted; L0 semantic snapshot, L1 committed perspective/context V2 and L2 Forma admission/KB0 contracts landed in PRs #936-#938. These are enabling contracts, not automatic product-capability promotion.
• P1-TEC / FM0 remains active. Governed replay/consumer-policy work is substantially landed; remaining evidence-identity, calibration/qualification and downstream evidence-reference closure must remain fail-closed. See RFC 0009.
• UXR0-UXR3 are at bounded software exits. UXR4/UXR5 remain physical-evidence open. See UXR programme and Quest validation operations.
• Full Moneta remains IMPLEMENTATION PARTIAL. Bounded FMA-01–13 repairs, static Forma rendering/reverse explanation, V4 replay, FM3/4 composition/resolution fidelity and FM6/FM7 qualification batteries are landed. The FM1/FM2 direct critique-to-alternative product loop and headless projections landed in #1000; remaining work includes investigator-visible FM1/FM2 surface qualification, V4 package capture of direct-loop lineage, measured FM5 target-device value, and integrated human/device evidence. See the 4 October completeness audit, incremental plan and architecture.
• Quest Compute Acceleration is active as a bounded parallel performance lane. QCA0 captured an operator-accepted Quest 3S engineering baseline; its governed INVALID_RUN is unchanged and a corrected physical rerun remains required. QCA1-QCA4 attribution/comparisons remain open. See QCA index.
• Memory Palace authority cleanup is complete: the duplicate mutable MemoryPalaceGraph store was retired under ADR-0012/#964; the product palace remains a deterministic projection of InvestigationAggregate state.
• scriptc is parked for Quest/WebXR. It may be reconsidered only for a real native CLI/sidecar/tool deployment with measured benefit. See scriptc evaluation.

Current order:

P1-TEC / FM0 finite evidence closure
  -> applicable grounded-evidence gates for each downstream claim
owner review of RFC0012 minimum architecture + retained FM-DSE contracts
  -> DSE0 schema/falsifiers -> DSE1 bounded grounded overview -> DSE2/DSE3
  -> smallest complete investigation loop; DSE6 qualification starts at DSE1
conditional later: DSE4 multi-dataset / DSE5 runtime intelligence -> remaining Full Moneta exits
residual before FM8 completion: FM1/FM2 investigator-visible surface qualification + V4 direct-loop lineage capture
remaining FM5: measured advisory value and target-device qualification
parallel when disjoint: UXR4/5 physical qualification + QCA physical attribution/controlled probes
then: P1-WP -> P1-WQ -> PT9/PT10

Forward dependency spine

Work	Status	Next action / dependency	Detail
P1-TEC / FM0 trustworthy evidence	ACTIVE	Finish finite evidence identity/calibration/consumer qualification required by claim-bearing downstream work.	RFC 0009, generated RF table below
FM1 perspective/context	DIRECT-LOOP IMPLEMENTATION LANDED / SURFACE QUALIFICATION OPEN	Context identity, explicit epistemic purpose/history/adoption guards (#957/#960/#961) and direct NIL critique-to-alternative routing (#1000), V4 direct-loop lineage (#1012) and headless surface projections (#1015) are landed. Qualify investigator-visible intent/perspective surfaces through the real product path before claiming the FM1 exit.	Incremental plan
FM2 alternative-aware / Road Not Taken	DIRECT-LOOP IMPLEMENTATION LANDED / SURFACE QUALIFICATION OPEN	Atomic branch/revisit (#957), NIL alternative/intent bindings (#988), and attributable direct critique-to-alternative routing (#1000) and deterministic headless projections (#1015) are landed; direct-loop links/bindings are captured in V4 lineage (#1012). Qualify investigator-visible surfaces before the FM2 exit; Memory Palace remains an aggregate projection (#964).	Incremental plan, ADR-0012
FM3 Forma first vertical slice / L2-FORMA-1	STATIC PRODUCT SLICE LANDED / SEMANTIC QUALIFICATION OPEN	Typed grounded/conjectural admission, production adoption guard, 3D Forma rendering, reverse explanation and V4 static replay are landed (#960/#961). Qualify multi-phenomenon semantic comprehension and product usefulness before the full FM3 exit.	Incremental plan
FM3/4 composition + resolution adaptation / L2-FORMA-2 + L4-RUNTIME-BUDGET	IMPLEMENTATION LANDED / DEVICE QUALIFICATION OPEN	Final-content identity, budget/generation refusal and selected-graph fidelity are landed (#962). Qualify semantic recoverability and information-loss disclosure across materially different Quest/desktop budgets.	Incremental plan
FM5 System-1 assistance	ADVISORY IMPLEMENTATION / PROBES LANDED; VALUE QUALIFICATION OPEN	Consumed advice, disclosure, bypass, recorded replay, ABSTAIN and R5/R7/R8 comparison probes are merged. Finite/reproducible deltas alone do not establish benefit; evaluate measured value and target-device cost.	System-1/System-2 architecture, evidence inventory
FM6 human refinement / Forma knowledge	QUALIFICATION VERIFIED	Qualification battery (FM6-R6) merged in #980: verified holdout evaluation gate fail-closed behavior, System-1 proposal scoring improvement with knowledge case identity binding, frozen research mode, and clean registry rollback.	Incremental plan
FM7 searching / synthesizing Moneta	QUALIFICATION VERIFIED	Qualification battery (FM7-R7) merged in #983: verified novel admissible representation graph construction outside fixed catalogue, System-1 advisory seeding and bypass, Pareto non-domination, MCR7 genome handoff analytical gates, and Research Mode bitwise digest invariance.	Incremental plan
FM8 full Moneta / controlled adaptive representation	IMPLEMENTATION PARTIAL / INTEGRATED QUALIFICATION OPEN	The FMA-01–13 repair set, dual-epistemic core and FM1/FM2 direct loop are landed; do not restore completion claims until FM1/FM2 investigator-visible qualification, measured FM5 value, the RFC 0013 evidence-binding decision and required human/device evidence satisfy the integrated V3.1 exit.	Full Moneta audit + dual-mode addendum
FM-DSE minimum dataset embodiment	DSE6 QUALIFIED / FM-DSE COMPLETED	DSE6 qualification battery merged with CONTINUE verdict across physical Quest bounds, N-independent compute, long-session endurance without leaks, semantic recovery, and dual-mode epistemic disclosure; DSE4/5 remain conditional offline research.	RFC 0012, ADR-0013, DSE6 battery

| UXR4/UXR5 device qualification | EVIDENCE ACQUISITION OPEN | Capture attributable Quest interaction/render/resource/comfort evidence when hardware is available. | UXR programme | | P1-WP → P1-WQ | BLOCKED BY REQUIRED UXR/ASSURANCE CLOSURE | Productionize web surfaces, then ordinary-browser investigator qualification. | Product transition plan | | PT9/PT10 learned/private-preview work | DOWNSTREAM | Requires WQ and finite evidence closure appropriate to the promoted claims. | Product transition plan |

The default execution model remains one owner-controlled forward tranche plus genuinely disjoint parallel-safe ribs. Idle agents are preferable to invented work. High-risk authority/scientific changes require pre-implementation falsifiers and independent post-implementation review under AGENTS.md.

FM-DSE planned tranches

Tranche	Gate / outcome
DSE0 reduction + preflight	Review RFC0012; inventory callers/archives for deletion, derive subject identity with existing contracts where possible, freeze N/cardinality falsifiers and evidence plans.
DSE1 grounded bounded overview	Stable Dataset subject, evidence-bound deterministic composition, two governed structures, bounded cardinality, reverse explanation and attributable critique capture.
DSE2 semantic resolution	Reversible dataset -> structure -> subset -> observation traversal; obligation-preserving device variants; lifecycle identity survives evict/rebuild.
DSE3 dual-mode feedback loop	Grounded/conjectural purpose enforcement, visible disclosure, TechnoCore critique -> Road Not Taken alternative, attributable feedback and V4 lineage.
DSE4 multi-dataset comparison	CONDITIONAL under proposed RFC0012: reopen for a named blocking research task; independent identity and evidence remain required.
DSE5 runtime proposal/search	CONDITIONAL under proposed RFC0012: offline lab supplies governed artifacts; add resident inference/search only for measured incremental task/resource value.
DSE6 qualification	Starts with DSE1; corrected physical controls, semantic-recovery/accessibility studies and STOP/CONTINUE/REVISE. Does not wait for DSE4/5.

Each implementation tranche requires its own reviewed plan and high-risk evidence. The architecture PR does not authorize code changes or promote performance, comprehension, search or learning claims.

RFC0012 proposes removing world/shared-depth model plans, generic modality scaffolding and new subject/working-set stores; merging product-metaphor state and representation orchestration; and moving training/evolutionary campaigns outside the browser product. Hosted collaboration, live streams and native/alternate compute remain conditional. Existing service/readiness obligations change only through their accepted migration, not by this proposal.

Quest Compute Acceleration — individual tasks

This replaces scriptc as the immediate Quest-browser optimization experiment. Rust/WASM remains analytical authority and ordinary browser/WebXR deployment remains the default topology.

ID	Status	Dependency	Task
QCA0	BASELINE CAPTURED / QCA1–QCA4 OPEN	custody-sealed Quest 3S QCA0 baseline	Quest performance baseline
QCA1	READY AFTER QCA0	QCA0 hotspot evidence	Rust/WASM SIMD128
QCA2	READY AFTER QCA0	QCA0 main-thread evidence	Worker/main-thread offload audit
QCA3	BLOCKED_BY QCA2 MEASUREMENTS	prove transfer/synchronization cost matters	Transferable buffers, memory and WASM threads
QCA4	BOUNDED CHANGES LANDED / PHYSICAL COMPARISON OPEN	Low-triangle markers (#979) and operator glyph hygiene (#986) merged; physical frame/legibility/selection benefit remains unqualified	WebXR rendering fast wins
QCA5	EXPERIMENTAL AFTER QCA0	select one measured non-authoritative parallel workload	WebGPU compute spike

QCA1, QCA2 and QCA4 may run concurrently only when their files/ownership do not collide. No Quest performance claim closes from desktop/simulator evidence alone.

Support loops

Loop	Status	Authority / rules
Adversarial Shadow Review (ASR)	READY / advisory	Candidates require independent validation before becoming obligations. Shadow governance
Recursive Falsification Laboratory (RFL)	READY / advisory	Additive falsification only; no production repair/self-adjudication. Iteration contract

Validated critical/high findings pre-empt affected feature work. Candidate findings do not. Agents should inspect their assigned validated findings before selecting ordinary work.

Parked / conditional paths

Path	Status	Re-open only when
scriptc native TypeScript	DEFERRED FOR QUEST/WEBXR	A real native CLI/sidecar/tool target exists and T0-T2 demonstrate semantic parity plus material operational benefit. Plan
Native Quest/Hexagon escape hatch	UNSCHEDULED	Browser/WASM/render optimizations fail a measured requirement that a native topology can plausibly solve. Quest SoC report
World-model representation substrate	UNSCHEDULED	A named evidence-triggered problem cannot be solved adequately by the bounded Forma/System-1/System-2 path.

Completed / historical status

Only completion state belongs here; detailed archaeological material is in the single historical ledger.

Area	Status	Detail
Progressive Disclosure / Stream A	VERIFIED COMPLETE / STOP	Historical ledger
Relationship Graph V1 / Stream B	VERIFIED COMPLETE / STOP FOR FIRST FAMILY	Historical ledger
Visible Investigator C1-C4	IMPLEMENTATION LANDED; physical qualification remains elsewhere	Historical ledger
PT0-PT8	LANDED AT BOUNDED EXITS	Product transition plan
A27-0/A27-1 + L0/L1/L2 foundation contracts	LANDED / ACCEPTED	Architecture 2027 review
Pre-compaction roadmap records, incidents, PR-by-PR history, old stream checklists	ARCHIVED	Historical ledger

Mutable review-finding status

Generated review-finding disposition. Mutable RF status is owned by governance/review-findings.json; edit the ledger, not this table.

Finding	Severity	Status	Current disposition
RF-037	Critical	VERIFIED COMPLETE	Single-replica live admission is verified; multi-replica replay safety remains the explicit RDO-007 deployment obligation.
RF-038	High	VERIFIED COMPLETE	Exact role allow-list and production-path rejection evidence are verified.
RF-039	High	VERIFIED COMPLETE	Production FileLoader -> Atlas -> Rust -> Dataset policy consolidation, pre-read gating, and adversarial live-path tests are verified.
RF-040	High	VERIFIED COMPLETE	Client-wide TelemetryCollector and UXTraceRecorder default-off, revocation halt, and full-erasure contracts are verified; research study helper claims are honestly bounded.
RF-041	Medium	VERIFIED COMPLETE	PR #619 removed the remote Three.js import-map/CSP trust and retained a production hygiene regression.
RF-042	Low	VERIFIED COMPLETE	PR #647 neutralized C0/C1/ESC terminal control sequences and verified the regression on the promoted exact head.
RF-043	High assurance gap	VERIFIED COMPLETE	Systematic hostile-input fuzz and property campaigns verify fail-closed parser and ABI boundary integrity with zero leaks or traps.

Agent update protocol

When an agent starts a task:

1. refresh main and state the exact base SHA in the task/PR;
2. claim one bounded task and identify collision-sensitive files;
3. update the linked work document with implementation evidence, not the roadmap with narrative;
4. change the roadmap only when status, sequencing, dependency or ownership changes;
5. on completion, leave a short final evidence record in the task document and append one concise line to the historical ledger;
6. merge only with exact-head evidence appropriate to risk, then refresh before selecting the next task.

This keeps the roadmap small enough to coordinate humans and agents instead of becoming the project itself.