| Transcript level ≠ functional channel density |
Use expression as a prior only, with wide variance; regulation sets the actual density |
| Transcriptome false negatives exclude real channels |
Channel presence is a spike-and-slab prior with detection-based inclusion probabilities; hard exclusion needs reporter or electrophysiological evidence (L2); truncation checked with π_min = 0 on the worm |
| Many invertebrate channel/receptor genes lack measured kinetics |
Homology priors + inference; flag sensitive unknowns for targeted electrophysiology |
| Per-type regulatory programs Φ are not identifiable from realistic data (fewer parameters, but still sloppy in the directions that matter) |
M3 identifiability gate runs before M4. If Φ fails to identify, change what is fitted (e.g. share regulatory gains across families of types, or fit set points only) or add the experiments the gate says are missing |
| Regulation rule misspecified (the mechanistic rule form is wrong even if H* is right) |
Fit a learned local rule family alongside the mechanistic one (L7); H* is judged on locality, not on one functional form |
| H* and R1 (genetic hardwiring) can't be separated with acute data |
Commission chronic-manipulation and timed-depletion datasets (roadmap); F3, F7a |
| Falsification tests are underpowered, so a negative result is uninterpretable |
Power computed by simulation at M3 on the real design; data requests quote required sizes; underpowered tests can only be inconclusive (02-hypotheses.md) |
| A lost prediction contest is over-read as biological evidence |
Decision rule separates predictive selection from mechanistic evidence; negative mechanistic claims require excluding missing mechanisms, weak interventions, insufficient measurement and optimization failure |
| Commissioned datasets arrive late or not at all |
Tracks A–C produce results from public data first; the decision rule names which verdicts are possible without F3/F7 (R2 vs. rest) |
| Platform-first trap: years of infrastructure before any science |
Adopt existing tools first; extract platform components from working code (07-platform) |
| Connectome-free data-driven models match CAKE, so anatomy and physics add nothing measurable |
Baselines are scored from M0; CAKE's claimed advantage is generalization to unseen perturbations (E5), which is tested explicitly. A tie there is a publishable negative result about the value of mechanism |
| Leakage between priors and validation data inflates results |
Leakage audit at M0 and on every benchmark release |
| Regulation machinery can't reproduce even a known case, making H* tests uninformative |
AFD positive control must pass before M4 tests H* |
| Fixed synapses can't reproduce learned behavior |
Associative plasticity in L7 for well-characterized circuits; E4 includes learned behavior |
| Results depend on an assumed rearing environment |
Default to the data-generating lab conditions; report sensitivity to alternative rearing distributions (L7) |
| H* is false (regulation does not determine θ well) |
F1–F7 are run to outcomes in P0. If H* is rejected, fall back to per-type direct fitting with the same infrastructure. If it stays undecided, P1 proceeds with the predictive selection. The Necessity Map and H2/H3 still stand |
| Chaotic dynamics break gradients |
Multiple shooting, teacher forcing, shadowing gradients for time averages, compare statistics rather than trajectories for E3 |
| Too many positive Lyapunov exponents make shadowing gradients expensive |
Measure the Lyapunov spectrum on the worm in M2; fall back to ensemble likelihood-ratio estimators or shorter statistic windows |
| Fixed-point burn-in misses slow regulation dynamics (e.g. hysteresis, multiple fixed points) |
The developmental outcome is defined by the dynamics, with an explicit selection rule (L7); endpoint mode only under validity conditions V1–V8 and with certification (residual, stability, basin check); otherwise transient mode; validated against direct burn-in on the worm |
| Root-finder converges to an unstable or unreached equilibrium |
Stability check on the loop gain and basin check against transient integration from the initial state (06-compute.md §5.2); pseudo-transient continuation follows the flow |
| Warm starts or cache hits select a different branch |
Warm starts only with certification; cache keys include initial-state and schedule samples (06-compute.md §8) |
| Constant-gain mechanistic rules cannot produce lasting developmental history effects (F7b) |
Known in advance (L7); the learned family includes state-dependent and higher-order rules; F7b estimates which mechanism carries history rather than assuming one |
| Regulation rates unidentifiable from equilibrium data |
Known gauge (L7); rates fitted only from transient data (F3/F7 time courses), which the commissioned datasets provide |
| Regulation-level waveform relaxation fails to converge (strong regulation–network feedback) |
Estimate the spectral radius of the outer iteration during the run and switch to the joint solver above 0.9 (06-compute.md §5.3); measure how often this happens on the worm |
| Reduced model drifts out of its validated regime |
N1 reductions certified over a declared parameter domain, rebuilt on exit; fidelity probes with automatic refinement (06-compute.md §3, §9) |
| Local reduction error underestimates error in network or behavioral results |
Goal-oriented (adjoint-weighted) error propagation and coupled paired-replica comparisons with feedback on, including closed-loop behavior (06-compute.md §9) |
| Lumped synapses misrepresent stochastic release or receptor saturation |
Lumping exact only under stated conditions; otherwise a validated approximation (S0/S1) with distributional CI tests against per-site models (06-compute.md §3) |
| Results differ across hardware or batching despite identical RNG streams |
Bitwise reproducibility only within a reproducibility class; statistical equivalence across classes, tested in CI (06-compute.md §4.6) |
| Benchmark metrics accept predictions with wrong amplitude or timing |
Primary losses score amplitude and timing (trace MSE, effect-size and trajectory losses); correlation only as a diagnostic (03-evaluation.md) |
| H3 residuals attributed to latent edges when they come from other model errors |
Structured decomposition with competing error sources, coherence diagnostics, and a synthetic specificity check before F6 (02-hypotheses.md) |
| H2 headroom mistaken for validated prediction |
B1 reports headroom only; F5 needs a named dataset with sufficient bandwidth, transmission-mode labels and an observation model |
| Brain states are not Markovian at a usable lag time, so stratified averaging is biased |
Implied-timescale tests; fall back to longer time-averaging windows (higher cost) |
| Content-addressed cache serves stale results after a silent code change |
Cache keys include code version, reproducibility class and full input hashes; periodic random re-computation audits |
| Spec documents drift out of sync with each other or with the code |
Each document has an owning working group; cross-document links checked in CI; decisions recorded in the decision log |
| EM morphology errors (diameters, shrinkage) |
Treat as latent with priors; test sensitivity |
| Individual variability vs. single-reference connectome |
Use multiple connectomes where available (worm developmental series, fly female/male/BANC); report predictions as distributions |
| Developmental connectome series mixes age with individual variation |
Probabilistic developmental wiring model (age trend + individual deviation); burn-in on sampled histories; report branch-switching fraction (L7) |
| Compute estimates prove optimistic |
All budget numbers are feasibility hypotheses with a testing milestone and a re-plan criterion; end-to-end workload benchmarks (06-compute.md §10) |
| Information-budget tolerances are too loose because local Fisher estimates miss nonlocal errors (branch switches, bifurcations) |
Coupled comparisons measure the total Δ on validation runs and calibrate local estimates; certification checks 2–3 are never relaxed by Fisher weighting (12-information-budget.md §4.3, §8) |
| Budgets certified on one benchmark version silently fail on a larger one |
Cached artifacts record their Δ and benchmark version; configurations are re-certified per version (12-information-budget.md §1) |
| Imaging cannot supply information as fast as chaotic dynamics create it, so trajectory fitting is ill posed |
Data-rate gate before fitting; failing tasks are scored as distributions (12-information-budget.md §6) |
| Low-precision early solver iterations switch equilibrium branches |
Precision schedule in the solver's own metric, kept below the stability margin; basin check unchanged (12-information-budget.md §4.3) |
| Milestones depend on components scheduled later |
Minimal regulation simulator and reduced closed loop (M2b) precede M3 and M4; A0 demonstrator precedes both (09-roadmap.md) |
| Compute cost |
Reference fidelity only for small systems and validation; reduced models gated by the Necessity Map |
| Data access and licensing |
Ingest only open datasets first; record licenses in provenance metadata |