# Deliberation circle — kido-resonate design (2026-10-14)

Module: kernel-side agent signaling on the resonant mesh. Scheduler tasks are
transmitter-agents; delivery is Lorentzian-gated by each receiver's tuned band
and Q. First deliberation on signaling — no prior records exist (inter-kernel
protocol was deferred to this module's scope by the boot/HAL conformance
reviews; see C5).

## Positions

- **P1 — Mesh purist.** Frequency IS the address. `transmit()` addresses a
  band, not a task id; every tuned task gets its own eta evaluation. This is
  the resonant-mesh spec's publish-subscribe (broadcast bus + tuned receivers),
  not point-to-point wearing a costume. Task-id routing would reintroduce the
  hub-and-spoke the mesh exists to survive.
- **P2 — Kernel pragmatist.** eta is `float`. AArch64 mandates FP/SIMD in the
  base ISA; the hosted path has it; the Lorentzian is smooth math and
  fixed-point would quantize the selectivity curve dishonestly. A future
  FPU-less port declares a fixed-point path then — not now, not faked.
- **P3 — Honesty auditor.** "Receiver state change" must not become fake
  neurons. A task's signal state is a bounded pending queue plus rx/tx/drop
  counters. No activation variables, no thresholds-as-consciousness. The
  neural-dynamics loop (release → route → resonance response → state change →
  reuptake) is honored structurally: transmit = release, eta gate = resonance
  response, enqueue+counters = state change, recv+slot-free = reuptake.
- **P4 — Cooperative realist.** Delivery evaluates synchronously inside
  `transmit()`; consumption happens at the receiver's own pace via `recv()`.
  No interrupts, no async delivery — preemption is still a loud stub, and the
  scheduler's yield points are the natural rhythm without needing hooks into
  the fenced scheduler files.

## Challenges and resolutions

- **C1 — Two tasks share a band.** Both evaluated independently; both may
  receive. That is pub/sub behaving correctly, not a bug.
- **C2 — Queue full on a deliverable signal.** Dropped, counted in
  `rx_dropped` and the call's `dropped_out`, never silent. Resonance passed
  but storage failed — the distinction is visible in the aggregates.
- **C3 — Untuned task (band 0).** Never receives; skipped before any eta math,
  so no division by zero. Band 0 is the untuned marker, not an address.
- **C4 — Broadcast.** `KIDO_BAND_BROADCAST` evaluates every tuned task; each
  still gates on its own eta. A high-Q task off-frequency visibly rejects a
  broadcast — selectivity is per-receiver, not per-send.
- **C5 — Inter-kernel protocol.** The boot/HAL conformance reviews deferred it
  to this module. Verdict: DECLARED, not built. There is no inter-kernel
  transport yet (no shared bus, no doll-layer link driver); inventing one
  would violate loud-stub discipline. The in-kernel draft carries the full
  wire shape (source, dest band, type, strength, frequency) so the future
  transport plugs in without changing the API. Single-hop delivery applies no
  decay (identity) — the decay field is carried for future multi-hop.
- **C6 — Coherence metric.** Out of scope. Tasks are transmitters, not
  neurons; computing Kuramoto over task "phases" would be numerology. The
  kernel provides signaling primitives; coherence is a world-level concern.
- **C7 — Q = 0.** Degenerate but defined: the Lorentzian's (2QΔf/f)² term
  vanishes, so eta = 1.0 for all frequencies — an unselective receiver.
  Mathematically consistent, documented, not a special case hack.

## Convergence

- eta in float; delivery threshold 0.5 — the half-power (−3 dB) point, where
  detuning equals f/(2Q). Principled, not picked.
- `KIDO_BAND_UNTUNED = 0`, `KIDO_BAND_BROADCAST = 0xFFFFFFFF`.
- Per-call aggregates: delivered / rejected (detuned) / dropped (queue full).
  Rejection is a result enum, not an error — detuning is normal physics.
- Static tables only: 16 task slots × 8-deep pending queues. No allocator
  (fenced), no hidden coupling (explicit init, no kido-mem globals).
- Arrival timestamps from HAL `timer_ticks()`; NULL timer tolerated, stamps 0,
  documented — consistent with the scheduler's P5 accounting law.
- Strength in permille (0..1000); >1000 is `KIDO_ERR_EARLY_ENV` per the
  codebase's arg-validation convention. NULL out-pointers: same.

## Preserved dissent

- **D1 — Fixed-point eta.** Overruled (P2). Revisit if N8toriousAi ever ports to
  FPU-less silicon; the formula's quantization behavior would need its own
  deliberation then.
- **D2 — Deferred delivery at yield.** Overruled for v1: evaluating at yield
  would need hooks in the fenced scheduler files. Queued as a future
  refinement with its justification intact.
- **D3 — Threshold 0.25.** Overruled: 0.5 is the half-power point. A
  permissive threshold would admit interference the Q was tuned to reject.
