External review, 2026-09-15
An external review dated 2026-09-15 examined the spec and the first
implementation plan against the dataset, the Shiu et al. code, and the
cflib source. Its findings and their dispositions follow. The section
numbers in the disposition column refer to the design specification,
docs/superpowers/specs/2026-09-15-flatline-imago-design.md, not to
chapters of this book.
| Finding | Disposition |
|---|---|
| Signed LIF with histaminergic photoreceptors silences the lamina | Accepted; section 8.6 policy and probe |
| LIF differs from Shiu et al. (g reset, refractory freeze, strict threshold) | Accepted after checking the released code; section 7.1 |
| Hover curriculum solvable by a zero decoder | Accepted; section 12.3 |
| No optic-flow diagnostic | Accepted; section 8.6 probe |
| 10 Hz camera far below motion-vision bandwidth | Accepted as risk and measurement; section 8.5 |
| Haltere round-robin discards type structure | Accepted; grouping by type; section 8.4 |
flight-core mask removes 0.2% of neurons | Accepted; replaced by no-optic-lobe, masks are not a real-time strategy; section 7.6 |
| Training compute unbudgeted | Accepted; section 19 |
| Safety envelope inadequate for an open workshop | Accepted; section 10.5 |
| Missing success criteria and open questions | Accepted; section 17 |
| Connectome contract errors (unclear label, eight scales, 1,314 descending, side rule, hex target types) | Accepted; sections 6.3, 6.5, 6.6 |
| Decision log lacks alternatives and reasons | Accepted; section 4 |
| Spec and plan file names diverge | Accepted; section 13.1 |
| Command-axis signs half-specified | Accepted after checking cflib; section 10.4 |
| Clock, threads, determinism underspecified | Accepted; sections 7.8 and 10.1 |
| Shuffle does not preserve in-degree | Accepted; the shuffle permutes partners within synapse-count classes, section 6.3 |
| Retina normalization scaled every photoreceptor edge | Accepted; section 8.3 |
| Photoreceptor column should be the strongest single target | Rejected; strongest column by summed weight retained with the reason in section 4 |
| Input neurons not bypassed and edges onto them kept | Accepted; sections 6.3 and 7.2 |
| Full edge table loaded whole for placement | Accepted; plan filters by photoreceptor before materializing |
| Event propagator loops per brain on Metal | Accepted; packed batch; section 7.3 |
| Config and checkpoint schemas missing | Deferred to the sub-project plans that create them |
The full review is filed at
docs/reviews/2026-09-15-grok-review.md
in the repository.