Implementation Tracker — 2026-06-02

Generated from the specification corpus in PR #1. Every implementable unit of work in the spec is filed below as a GitHub issue; each is independently shippable, and cross-issue dependencies are noted inline and in each issue body. Issue numbers link to the tracker.

Totals: 94 issues (77 implementation + 17 tracking). Milestones: v0.1 (40), v0.2 (28), v0.3 (5), v1.0 (4).

Current Evidence Track

The original generated tracker has been substantially implemented. The closed claim-MVP epic #190 established the end-to-end federated LeWorldModel smoke path with published HF datasets, checkpoint, and report artifacts.

The scaled empirical evidence stream closed in #200 and PHASE2.md. It covers GPU-backed HF Jobs, larger participant silos, downstream eval, baselines/ablations, performance curves, and the model-card evidence bundle. Those runs predate the outer-update direction correction and are historical execution/artifact evidence, not validation of the corrected runtime. Issue #335 tracks the clean replacement run.

The historical operational consortium-training evidence stream is #220 and PHASE3.md. It covers governed multi-party training with separate participant agents, a coordinator service, secure-sum and DP reporting cross-checks, downstream evaluation, lifecycle reporting, and a final evidence bundle. It explicitly excludes the provenance ledger and RFC-0006 cryptographic-proof implementation. Closing that evidence epic did not complete the end-to-end privacy contracts.

Privacy integration status (live checkout, 2026-07-24)

Surface	Tested/live component	End-to-end blocker
Participant DP transform	Participant.local_round() clips and adds Gaussian noise before constructing the released update; mechanism tests cover clipping and noise scale	Noise is derived from the shared, manifest-recorded root seed plus round and participant id. The coordinator can reconstruct and subtract it. Claim-grade mode needs participant-secret CSPRNG entropy and must keep deterministic replay explicitly non-private.
DP accounting	RDP/PRV accountants are stateful primitives with direct tests	The Phase 3 report constructs a fresh accountant after each commit, accounts one step (rounds_accounted=1), and discards it. No persistent pre-release check or cumulative hard stop is wired.
Secure aggregation	Pairwise-mask/dropout primitives, fixed-point sum harnesses, and the software TEE path have direct tests	The coordinator accepts unmasked per-participant deltas and commits from them before the service computes a secure-sum equivalence report. The reported sum is not optimizer input; masking selection remains an explicit fallback.
Historical Phase 3 privacy rows	Ten secure_sum rows, ten DP-accounted rows, and per-round epsilon near 5.30 were recorded by the evidence run	Read them as post-commit secure-sum cross-checks and independent one-round accountant snapshots, not as coordinator confidentiality, a composed ten-round epsilon, or run-budget enforcement.

These blockers remain implementation work even though the historical evidence epic is closed. The accepted RFCs preserve the target protocol; they do not describe the current service as satisfying it.

The Status column below is the 2026-06-02 generation snapshot, not a live field. The evidence-track epics above (#190, #200, #220) closed their historical evidence streams; that does not close the privacy integration blockers listed above. Authoritative per-issue state lives on the GitHub tracker, not in this column.

Milestone: v0.1 (Stage A — single-site upper bound + scaffolding)

#	Title	Area	Priority	Effort	RFC	Status
#31	artifacts: safetensors checkpoint format with JSON header	artifacts	p0	m	0010	Open
#32	artifacts: canonical-byte SHA-256 hashing and integrity checks	artifacts	p0	s	0010	Open
#66	ci: GitHub Actions pipeline (lint, types, unit, coverage)	ci	p0	m	0009	Open
#65	ci: pytest + hypothesis harness and CPU fixtures	ci	p0	m	0008	Open
#34	config: Hydra structured config schema and validation	config	p0	m	0009	Open
#35	config: deterministic seeding scheme	config	p0	s	0009	Open
#7	contracts: implement ActionSpec and its validation	contracts	p0	s	0007	Open
#6	contracts: implement LatentState and the WMCP conformance check	contracts	p0	m	0007	Open
#3	core: implement the error taxonomy in lensemble/errors.py	core	p0	s	0001	Open
#2	core: scaffold the lensemble package and module layout	core	p0	m	0001	Open
#21	data: episode/transition schema and windowed loader	data	p0	m	0004	Open
#24	data: public probe set, landmark targets, and hash pinning	data	p0	m	0004	Open
#23	data: residency guard (INV-RESIDENCY, fail-closed)	data	p0	m	0004	Open
#52	eval: evaluation harness	eval	p0	m	0005	Open
#51	eval: latent MPC planner	eval	p0	l	0005	Open
#10	model: video-ViT encoder with V-JEPA 2 warm-start	model	p0	l	0008	Open
#11	model: action-conditioned latent predictor	model	p0	m	0008	Open
#12	model: SIGReg implementation (shared sketch + Epps-Pulley statistic)	model	p0	l	0008	Open
#59	observability: redaction guard (INV-RESIDENCY, fail-closed)	observability	p0	s	0015, 0004	Open
#71	packaging: pyproject.toml and dependency pins	packaging	p0	s	0009	Open
#5	cli: Typer app skeleton with config loading and RunManifest emission	cli	p1	m	0001	Open
#37	config: reproducibility guarantee test	config	p1	s	0009	Open
#36	config: RunManifest schema and emission	config	p1	s	0009	Open
#8	contracts: define the ActionHead interface and cond_dim contract	contracts	p1	s	0007	Open
#4	core: enforce module dependency layering (no cycles)	core	p1	s	0001	Open
#22	data: lance / hdf5 / lerobot data adapters	data	p1	l	0004	Open
#70	ci: documentation link-check gate	docs	p1	s	0001	Open
#73	docs: CONTRIBUTING and the RFC process	docs	p1	s	0009	Open
#74	docs: SECURITY.md and threat-model alignment	docs	p1	s	0006	Open
#53	eval: metrics (success, planning cost, effective dim, comms)	eval	p1	m	0005	Open
#14	model: numerical contract (dtype/device/determinism)	model	p1	m	0008	Open
#13	model: the composite training objective	model	p1	m	0008	Open
#58	observability: metric taxonomy emission	observability	p1	s	0015	Open
#57	observability: structured JSON logging	observability	p1	s	0015	Open
#72	packaging: license files (code/docs/data)	packaging	p1	s	0009	Open
#27	provenance: canonical episode hashing	provenance	p1	s	0014	Open
#28	provenance: Merkle tree, root, and inclusion proofs	provenance	p1	m	0014	Open
#26	data: data-quality metadata and WMCP precondition declaration	data	p2	s	0004, 0007	Open
#76	docs: changelog discipline (Keep a Changelog)	docs	p2	s	0009	Open
#75	docs: CODE_OF_CONDUCT.md	docs	p2	s	0009	Open

Milestone: v0.2 (Stage B — simulated federation + the gauge + paper experiments)

#	Title	Area	Priority	Effort	RFC	Status
#39	federation: deterministic Nesterov outer step	federation	p0	m	0003	Open
#38	federation: PseudoGradient contract	federation	p0	s	0003	Open
#42	runtime: Coordinator orchestration	federation	p0	l	0013	Open
#43	runtime: Participant local round	federation	p0	m	0013	Open
#41	runtime: round state machine	federation	p0	m	0013	Open
#16	gauge: Variant A landmark frame-anchor loss	gauge	p0	m	0002	Open
#19	gauge: frame-drift diagnostic (the headline measurement)	gauge	p0	m	0002, 0005	Open
#15	gauge: closed-form Procrustes alignment with degeneracy handling	gauge	p0	m	0002	Open
#49	privacy: per-participant clip+noise mechanism	privacy	p0	s	0012	Open
#61	verify: aggregation determinism self-check	verify	p0	s	0006, 0003	Open
#46	aggregation: in-process simulated secure aggregation (Stage B)	aggregation	p1	m	0011	Open
#68	ci: aggregation determinism gate	ci	p1	s	0006, 0003	Open
#67	ci: ML-specific property tests (gauge invariance, anchor, SIGReg)	ci	p1	m	0008, 0002	Open
#9	contracts: wmcp_version federation join gate and conformance errors	contracts	p1	s	0007	Open
#55	eval: ablation ladder runner	eval	p1	m	0005	Open
#54	eval: baseline configurations	eval	p1	m	0005	Open
#17	gauge: Variant B rotational-drift anchor penalty	gauge	p1	m	0002	Open
#18	gauge: Layer-3 Procrustes re-alignment backstop at aggregation	gauge	p1	m	0002	Open
#60	observability: frame-drift diagnostic emission contract	observability	p1	s	0015, 0002	Open
#50	privacy: (eps,delta) accountant	privacy	p1	m	0012	Open
#29	provenance: DatasetCommitment and binding Delta_c to R_c	provenance	p1	s	0014	Open
#62	verify: public recomputation of frame alignment	verify	p1	s	0006, 0002	Open
#33	artifacts: schema versioning and migration	artifacts	p2	s	0010	Open
#69	ci: performance smoke budget	ci	p2	s	0001	Open
#25	data: probe versioning and re-anchoring event	data	p2	s	0004	Open
#56	eval: non-IID severity and scale sweeps	eval	p2	l	0005	Open
#20	gauge: Layer-4 function-space distillation fallback	gauge	p2	l	0002	Open
#30	provenance: append-only contribution ledger	provenance	p2	s	0014, 0004	Open

Milestone: v0.3 (Stage C — two real sovereign nodes)

#	Title	Area	Priority	Effort	RFC	Status
#47	aggregation: pairwise-mask secure aggregation with dropout robustness	aggregation	p0	l	0011	Open
#44	runtime: elasticity, churn, and rejoiner recovery	federation	p1	l	0013	Open
#45	runtime: networked control-plane transport	federation	p1	l	0013	Open
#48	aggregation: TEE-based aggregator backend	aggregation	p2	l	0011	Open
#40	federation: optional int8 pseudo-gradient quantization	federation	p2	s	0003	Open

Milestone: v1.0 (hardening, public-API freeze, Fork A, verified proof-readiness)

#	Title	Area	Priority	Effort	RFC	Status
#63	verify: proof-readiness audit	verify	p1	m	0006	Open
#77	docs: documentation site build and API docs	docs	p2	m	0009	Open
#78	packaging: release automation and artifact release	packaging	p2	m	0009, 0010	Open
#64	verify: Phase-2 prover interface stub	verify	p2	s	0006	Open

Tracking issues

• #79 [Tracking] Architecture & foundation — 4 children
• #80 [Tracking] WMCP latent contract & embodiment adapters — 4 children
• #81 [Tracking] Model, objective & numerical contracts — 5 children
• #82 [Tracking] The latent gauge & frame-anchored aggregation — 6 children
• #83 [Tracking] Data, sovereignty & provenance (data layer) — 6 children
• #84 [Tracking] Provenance commitments & Merkle scheme — 4 children
• #85 [Tracking] Checkpoint & artifact format — 3 children
• #86 [Tracking] Configuration, run manifest & reproducibility — 4 children
• #87 [Tracking] Federated training protocol — 3 children
• #88 [Tracking] Coordinator & participant runtime — 5 children
• #89 [Tracking] Secure aggregation protocol — 3 children
• #90 [Tracking] Differential privacy accounting — 2 children
• #91 [Tracking] Evaluation & benchmark protocol — 6 children
• #92 [Tracking] Observability, diagnostics & telemetry — 4 children
• #93 [Tracking] Verifiable contribution (proof-readiness) — 4 children
• #94 [Tracking] Testing strategy & CI — 6 children
• #95 [Tracking] Packaging, docs & release — 8 children

Cross-cutting dependencies

Selected cross-subsystem edges (the full graph is encoded in each issue's Dependencies section). A -> B means A blocks B.

• 2 (core: scaffold the lensemble package and module layout) blocks #5 (cli: Typer app skeleton with config loading and RunManifest emission)

• 2 (core: scaffold the lensemble package and module layout) blocks #21 (data: episode/transition schema and windowed loader)

• 2 (core: scaffold the lensemble package and module layout) blocks #31 (artifacts: safetensors checkpoint format with JSON header)

• 2 (core: scaffold the lensemble package and module layout) blocks #57 (observability: structured JSON logging)

• 2 (core: scaffold the lensemble package and module layout) blocks #65 (ci: pytest + hypothesis harness and CPU fixtures)

• 2 (core: scaffold the lensemble package and module layout) blocks #71 (packaging: pyproject.toml and dependency pins)

• 2 (core: scaffold the lensemble package and module layout) blocks #72 (packaging: license files (code/docs/data))

• 2 (core: scaffold the lensemble package and module layout) blocks #73 (docs: CONTRIBUTING and the RFC process)

• 2 (core: scaffold the lensemble package and module layout) blocks #74 (docs: SECURITY.md and threat-model alignment)

• 2 (core: scaffold the lensemble package and module layout) blocks #75 (docs: CODE_OF_CONDUCT.md)

• 2 (core: scaffold the lensemble package and module layout) blocks #76 (docs: changelog discipline (Keep a Changelog))

• 3 (core: implement the error taxonomy in lensemble/errors.py) blocks #6 (contracts: implement LatentState and the WMCP conformance check)

• 3 (core: implement the error taxonomy in lensemble/errors.py) blocks #7 (contracts: implement ActionSpec and its validation)

• 3 (core: implement the error taxonomy in lensemble/errors.py) blocks #23 (data: residency guard (INV-RESIDENCY, fail-closed))

• 3 (core: implement the error taxonomy in lensemble/errors.py) blocks #34 (config: Hydra structured config schema and validation)

• 3 (core: implement the error taxonomy in lensemble/errors.py) blocks #59 (observability: redaction guard (INV-RESIDENCY, fail-closed))

• 6 (contracts: implement LatentState and the WMCP conformance check) blocks #10 (model: video-ViT encoder with V-JEPA 2 warm-start)

• 6 (contracts: implement LatentState and the WMCP conformance check) blocks #15 (gauge: closed-form Procrustes alignment with degeneracy handling)

• 7 (contracts: implement ActionSpec and its validation) blocks #26 (data: data-quality metadata and WMCP precondition declaration)

• 8 (contracts: define the ActionHead interface and cond_dim contract) blocks #11 (model: action-conditioned latent predictor)

• 10 (model: video-ViT encoder with V-JEPA 2 warm-start) blocks #16 (gauge: Variant A landmark frame-anchor loss)

• 10 (model: video-ViT encoder with V-JEPA 2 warm-start) blocks #24 (data: public probe set, landmark targets, and hash pinning)

• 10 (model: video-ViT encoder with V-JEPA 2 warm-start) blocks #51 (eval: latent MPC planner)

• 11 (model: action-conditioned latent predictor) blocks #51 (eval: latent MPC planner)

• … and 28 further cross-subsystem edges (see issue bodies).