Phase 80 · model-validation milestone · toy computational model
The architecture narrative changes. Earlier phases assumed the routing information — the destination field — had to be constructed for every trip. In the latest toy simulations a destination field produced only by local neighbour relaxation persists across trips, deforms locally when the destination moves, and often survives node damage without any repair sweeps at all. That is a statement about a graph model in code. It is not physical evidence, and the scaling question it opens is still unanswered.
Validation status · permanent
Active research frontier · next wall
NOT YET VALIDATEDDoes a bounded local disturbance require only a sublinear — ideally vanishing — fraction of the network to update?
N_repair ≪ N, and ideally N_repair / N → 0
Every result above measures T_repair: how many sweeps a fix takes. That is the wrong ruler on its own. A repair that finishes quickly but touches the whole network is still a global operation wearing a local costume. The decisive measurement is N_repair — how many distinct nodes change state at all — as a function of N under a fixed, bounded disturbance. If N_repair / N does not fall, the persistent field is a global object and this milestone reduces to a caching trick.
What changed
Old model
OLD MODEL — build global routing information, or a fresh destination field, for every trip. The substrate pays construction on every journey, which is exactly the accounting that failed the one-off lifecycle gate in Phase 69.
Current model
CURRENT MODEL — maintain a persistent local field, pay cold start rarely, then use incremental local deformation plus transient front activity. Construction becomes infrastructure, not a per-trip charge.
Latest simulation findings · toy computational model
F1 · NOT PHYSICAL EVIDENCE
A destination field generated only by local neighbour relaxation — no global solve, no broadcast, no oracle — persisted across trips in the tested toy model. Routing performed after the field had matured completed 100% of sampled trips, and median stretch improved as the field continued to relax.
The tail is not clean. p95 stretch sits near 1.5 and is UNRESOLVED. It stays visible as an open issue and is not averaged away.
F2 · NOT PHYSICAL EVIDENCE
For nearby destination moves of roughly 2–3 graph hops, incremental field update required typically 4–8 local relaxation sweeps in the tested samples, against a ~8-sweep cold start in that same experiment.
Interpretation, held loosely: nearby destination changes can often reuse the existing distributed field rather than rebuild it. Far moves were not characterised.
F3 · NOT PHYSICAL EVIDENCE
After a mature persistent field was established, removing 5%, 10% or 20% of nodes often left routing inside the provisional gate with zero repair sweeps for the sampled N = 512–2048 runs. At N = 4096 representative samples needed 8 repair sweeps at 5% damage, 4 at 10% and 1 at 20%, against a 32-sweep cold start.
These are random-sample toy results from a small number of draws. They MUST NOT be read as established scaling laws. Larger replication across seeds, sizes and damage models is required before any of it is trusted.
Results table · simulation samples, not hardware data
| N | Cold start | Destination update (2–3 hops) | Repair · 5% loss | Repair · 10% loss | Repair · 20% loss |
|---|---|---|---|---|---|
| 512 | ~8 sweeps | 4–8 sweeps | 0 sweeps* | 0 sweeps* | 0 sweeps* |
| 1024 | ~8 sweeps | 4–8 sweeps | 0 sweeps* | 0 sweeps* | 0 sweeps* |
| 2048 | ~8 sweeps | 4–8 sweeps | 0 sweeps* | 0 sweeps* | 0 sweeps* |
| 4096 | 32 sweeps | not characterised | 8 sweeps | 4 sweeps | 1 sweep |
* Routing remained inside the provisional gate without any repair sweeps in the sampled runs at these sizes. Every cell is a SIMULATION SAMPLE from a toy graph model, not verified hardware data, and not a fitted scaling law. Single-sample cells at N = 4096 are representative draws only.
Lifecycle model update
C_total = M · (C_field-build + C_trip)
C_total = C_cold-start + Σ C_local-update + Σ C_trip
The field may exist continuously, deform locally after a change, and guide a transient C* front. Cold start is paid rarely rather than per trip. These are normalized counts of modelled sweeps and operations — not joules, not seconds, not physical energy.
Current candidate control stack
Persistent distributed field
Maintained by local neighbour relaxation only. Survives between trips.
Local differential sensing
Each node reads gradients against its own neighbours. No global state read.
Sequential-confidence decision rule
Accumulates local evidence before committing a hop, instead of a single-shot threshold.
Transient C* recruitment
A moving front wakes only the nodes it needs, for as long as it needs them.
Local recovery / self-restoration
Disturbed nodes relax back toward the standing field without a controller.
Why this matters for hardware
If local-repair-cone scaling holds, a physical substrate would not need a global controller to recompute everything after every fault or destination move. A node failure, or a destination that shifts a few hops, would propagate only as far as it needs to — a bounded neighbourhood of update rather than a whole-array rewrite. That would remove the single largest control-plane objection to building this at all: the assumption that any reconfiguration is a global operation.
In plain language
We learned the map may not need to be rebuilt every time. The network can keep a gentle direction field alive and only fix the nearby part when something changes. Now we have to prove the repair stays small as the network grows.
