Evidence-gated delivery milestones

Status: proposed for review, 2026-10-16. This is the delivery plan for the roadmap, not a new scientific protocol, calibrated model or source of experimental verdicts. Historical locks, scored results, w_syn=0.192 and th_jump=0 are unchanged. Do not infer biological fidelity from GPU parity, anatomical rendering, geometry, wiring shuffles or a synthetic closed loop.

Baseline and conventions

• main at this consolidation: f5e3709 (merged #230). #227 and #230 land the diagnostic bounded skeleton renderer. The 16 MiB motion campaign is a veto, not a headset cadence admission, and does not meet the exit of #177 or #24. The planning baseline was fd0ab7a (merged #221). #40 is closed after its worker baseline; unfinished adversarial Series S hardware cases are tracked separately in #224. #226 and #228 add the dimensional unit check and do not meet the exit of #148.
• Milestone IDs below are planning identifiers, not claims that GitHub's native Milestone objects have been created or that target dates have been committed. Issue descriptions and this index provide traceability until native objects are explicitly provisioned.
• Each existing open issue has one primary milestone (index below). It may block other milestones as a dependency without being duplicated or closed early. Closed issues and merged PRs are historical evidence, not outstanding work.
• Entry means prerequisites that must exist before a confirmatory run. Exit requires an immutable artifact, an independently specified checker, explicit pass/fail/indeterminate/blocked verdict, and an issue-specific acceptance decision. blocked is honest progress, not completion.
• A no-go or failed model result is publishable and does not authorize changing a frozen lock, repurposing a held-out dataset for tuning or asserting validated animal behavior. Propose a new preregistered version only with new evidence.
• Dates are intentionally unset. External biological evidence and physical device access are not predictable deadlines. Review status when evidence is acquired and at each merged milestone-related PR.

Critical path (science)

S0 numerical reliability → S1 sourced sensory physiology → S2 evidenced sensorimotor interfaces → S3 independent held-out prediction → S4 external replication and claims audit.

S1 and S2 can conduct independent source audits in parallel. Their scientific acceptance cannot bypass S0's model-specific numerical gates. S3 requires a locked, testable prediction and eligible untouched data; S4 may reproduce historical outcomes earlier, but its final claims sign-off depends on S3.

S0 — Numerical reliability and identifiability (P0)

Owner issues: #148. Related native semantics: #39.

Evidence already available: convergence v1 passes its preregistered battery criterion; weak-pathway sugar → MN9 first-spike latency does not converge (35.3 ms at 0.1 ms versus 77.4 ms at 0.025 ms). Sensitivity-grid identifiability remains post-hoc, with no biological parameter distributions or credible intervals. Seven positive battery cases exceed 199 degree-preserving rewires, a topology-specific result only.

Work: flymsg.units parses every evidence-registry unit and reads the live sim._constants and sim.graded_arrays values (decays, gv, kick, graded count). That check does not satisfy the exit below. Path-metrics v1 already froze latency, jitter, recruitment and rate and archived the 0.1 ms verdicts. Uncertainty propagation stays refused until a measured distribution exists. Shuffle baselines v1 are archived. Numerical-mode coverage stays with #39, and convergence v1 stays as archived.

Exit: reproduce and independently audit the archived verdicts, record all nonconvergent or unidentifiable outputs and provide a defensible numerical eligibility decision per prospective biological observable. An unqualified observable stays excluded from S1/S3 scoring. No invented intervals.

S1 — Sourced sensory and circuit physiology (P0)

Owner issues: #34, #84, #85, #97, #140.

Evidence already available: T4/T5 v4 remains robust: neither, giant-fiber looming v1/v2 and APL-claw v1 are uninformative, song v1 stopped before held-out data. Optional gap/graded kernels are implementations, not measured in-vivo conductances or release curves.

Work: build sourced observation/transfer contracts for at least one eligible modality (e.g. visual motion, looming, odor/APL), including cell identity, sex, stimulus geometry, units, gain/kinetics, measurement type, trial variability, and calibration-versus-holdout separation. Validate alternatives against simpler stimulus-following models and negative controls. Audit what cannot be mapped; do not convert calcium signals into Hz by fiat.

Exit: at least one independently defensible sensory/neural input/output comparison is eligible for prospective scoring with a frozen protocol. If source/registration/observation semantics remain absent, mark S1 blocked; an engineered Poisson encoder does not pass.

JO progress (2026-10-17): protocol 6 now has a source/identity/input design and accepted ADR. The 146 selected MaleCNS bodies are an anatomical population, not 146 measured transfer functions. Absolute source calibration and sex transfer remain blocked; runtime implementation awaits ADR acceptance. This advances S1 design, not its scientific exit.

S2 — Evidenced body coupling and causal geometry (P0)

Owner issues: #19, #31, #35, #176, #179, #180.

Evidence already available: versioned 6-DoF frames, column optics, FlyGym/NeuroMechFly adapter, five-arm synthetic causal loop and deterministic checkpoint/replay. The MaleCNS eye-to-body registration and the neuron → muscle → joint/actuator/force/timing chain remain unqualified. Every unsupported validation/motor_map.yaml path stays unmapped.

Work: independently qualify optical projection and its error bounds under fixed/moving poses, then obtain sourced coordinate registration and motor mapping/actuator evidence. Verify later sensor observations depend causally on preceding body motion against intact, open-loop, disconnected, shuffled and replayed controls. Archive real units, provenance, restart/failure traces and positive or null outcomes.

Exit: qualified world → eye → sensory → neuron → muscle → actuator → body → later world/eye path for the specifically mapped channel, with no engineered link mislabeled as physiological. Without both sourced sensory and motor links, retain the synthetic loop and mark biological S2 blocked.

S3 — Prospective independent physiological and behavioral tests (P0)

Owner issues: #36, #149, #150.

Entry: an S0-eligible observable, S1/S2 interfaces needed by the selected task, independent accessible measurements, blinded or frozen calibration/ holdout partition, exact negative-control and simpler-model baselines.

Work: preregister one falsifiable neural/behavioral target before data inspection. Record exact provenance, trial variation, eligibility tests, statistical thresholds, effect sizes and adverse/null outcomes. Compare model variants only by separate prospective protocols; do not rewrite or reuse held-out files from earlier no-go experiments.

Exit: publish an independently scored preregistered test with a reproducible pass/fail/indeterminate/no-go and all controls. Only a pass against untouched animal observations satisfies the project's stated behavioral-completion claim. A properly documented failed or blocked run is valuable research but does not close that claim.

Dated progress (2026-10-17): the Baker audit completes anatomical routes for 12 of 28 classes, including the still-excluded dimorphic AVLP_pr36; 14 remain partial and two absent. The M4 v2 proposal records a scoped basis for noncommercial local analysis and publication review. This advances candidate identification, not S3 qualification: sex compatibility, qualified measured JO input, observable qualification and the prospective partition/lock remain required. No response data were opened or benchmark scored.

S4 — External reproduction and claim audit (P1; mandatory for research release)

Owner issue: #151.

Work: package exact lock/code/data/acquisition/license and hashes, parameter evidence, independent checker, golden positive, negative and no-go examples, artifact-integrity verifier and documented environment. Arrange a clean-clone execution by an operator/machine distinct from the authoring workspace. Classify each public statement as software, anatomical, physiological, causal or independent-animal evidence.

Exit: independently signed results and a machine-readable exception ledger, with failures/unavailable licensed inputs explicitly recorded. Local CI or clean re-reads on the authoring machine do not close S4. External reproduction of a failed biological result is still valid replication; it does not turn that result into a successful animal prediction.

Parallel engineering milestones (not scientific prerequisites)

H1 — Qualified native execution and physical console recovery

Owner issues: #39 and #224. Closed #40 retains the worker baseline.

Available: full-connectome GPU/CPU bit parity in measured runs; 12 version-1 plus 60 supported version-2 fixture cases bit-identical on physical Series S, with 12 declared dynamic-clamp refusals. The 2,048-neuron ordinary both lifecycle qualifies cancel, suspend and resume; a noncooperative process kill recovers bit for bit with and without a checkpoint. An API-induced device removal is handled and recovered bit for bit; a hardware GPU fault is not qualified. A native publish whose temporary file does not grow is recorded as failed and is not replaced. The graded million-step publish that stayed in checkpointing with an empty temporary file was not rerun and does not meet the exit of #224. The sustained v1 full-connectome workload passes cold/warm setup and four back-to-back 500,000-step jobs per stimulus bit for bit, at median advance rates of 2,649 steps/s looming and 3,693 steps/s bitter without throughput decline.

Exit: adversarial numeric and resource boundaries; independently verified noncooperative termination and GPU/device-fault recovery; sustained representative load, transfer/OS/GPU memory and physical input evidence. If device fault cannot safely be injected, document a separate unqualified gate. No conflation of synthetic throughput with full-CNS real-time or neuromechanical parity.

H2 — Full-fidelity 3D viewer and physical WebXR qualification

Owner issues: #24 and #177.

Available: the merged #221 records source-attested Wi-Fi Quest A/B/B/A trials, rejects the 4 MiB paged approach on three valid dense blocks, and includes an exploratory bounded-error original-skeleton-path prototype (not production enabled). Merged #227 and #230 keep that renderer diagnostic. The recorded 16 MiB anatomical-motion campaign vetoes the candidate (median p95 ratio 1.409) and does not meet the exit of #177 or #24. The earlier complete-content partial 90 Hz session failed timing.

Exit: original-quality correctness plus strict visual/error budgets on real selected anatomy, physical phone cold/warm/memory baseline, repeatable eligible headset cadence and stereo/session/context-loss/ sustained thermal tests. Separately qualify any alternative renderer or bounded-error representation under its own ADR. A fast empty/incomplete scene, background tab or changed quality is not a pass.

Open-issue primary assignment (2026-10-16)

The following 20 issues are assigned only in this plan:

• S0: #148.
• S1: #34, #84, #85, #97, #140.
• S2: #19, #31, #35, #176, #179, #180.
• S3: #36, #149, #150.
• S4: #151.
• H1: #39, #224.
• H2: #24, #177.

Each issue retains its present open/closed state, historical result and evidence blockers. The set must be refreshed when issues change, rather than pretending this snapshot is a live GitHub query.

Execution and change-control protocol

1. For each milestone, nominate one issue as the execution index; keep implementation work and evidence requests on their existing issues.
2. Audit exact external data eligibility and model/source identity before drafting a new experimental lock. Publish the lock and independent gates on main before starting the scored run from a clean commit.
3. Separate each PR by one coherent boundary: protocols versus runs versus implementation; native/Quest qualification versus scientific outputs.
4. Require Ruff/pytest, relevant JS/native tests, documentation formatting, and reproducible checks appropriate to modified files. A passing GitHub status is not evidence of physical hardware that CI cannot exercise.
5. Attach raw checksummed evidence, negative results, observed units/variance, and a source-specific limits statement; leave failed/blocked items open.
6. Report at review: delivered / verified / failed / blocked / next evidence request. No percentage-complete estimates for unsolved biological inference.