Skip to content
Adjacency Theory markAdjacency Theory

Experiment v0.1 · preregistered bench protocol · locked

We wrote down exactly how the machine could prove us wrong — before we built the machine.

No new mechanism is proposed on this page. It fixes, in advance, what the bench must do: five phases from calibration to autonomous restoration, seven controls including a deliberately cheating one, fifteen primary endpoints with frozen definitions, ten proposed pass thresholds at N = 16, 32 and 64, the instrumentation channels, the rules that void a run outright, and the raw data that must be published for every run — including the failures.

Status

proposed thresholds · not measured results

BENCH PROTOCOL LOCKED · HARDWARE DATA NOT YET COLLECTED · PHYSICAL EVIDENCE NONE

Programmable electronic oscillator network, or hybrid analog oscillators with FPGA-computed coupling, on a bounded-degree sparse substrate. Analogue testbed only.

Phases

5

C0 → trigger → doorway → payload → restoration

Controls

7

A → G, including the cheating positive control

Primary endpoints

15

definitions frozen

Physical evidence

NONE

This protocol is frozen. Any change to a threshold, a window, a control, or an endpoint definition after data collection begins voids the affected series and must be published as an amendment with its own timestamp.

Protocol phases 1 → 5

  1. PHASE 1C0 baseline

    Ordinary local sparse coupling only. No doorway logic enabled anywhere on the substrate.

    propagation latencyspectral responsenoise floorelectrical energyfalse-trigger rate
  2. PHASE 2Trigger

    A single local seed injected above the preregistered amplitude threshold. NO global path command of any kind is issued.

    trigger thresholdhysteresis / operating windowrecruitment curve
  3. PHASE 3C* moving doorway

    Local threshold and recovery law plus compact local address logic may alter ONLY adjacent coupling state, and only as a function of that node's own state and the destination address it already holds. Every controller write and event is timestamped and auditable.

    max active fractiontotal active node-timecontroller writes with origincorridor trace
  4. PHASE 4Payload

    A structured waveform carrying at least four independent components or symbols is injected at A. The destination decoder is blinded to condition wherever practical.

    payload latency from trigger timestamp to decoded arrivalsymbol error rateidentity error
  5. PHASE 5Restoration

    All local coupling states return to C0 autonomously. No global reset is permitted at any point.

    restoration durationresidual coupling deviation from C0autonomous restoration rate

Controls A → G

#ControlPurposePreregistered expectation
AC0 ordinary baselineThe reference latency and energy for the same A/B pair.Defines the denominator of Q_time. Must be beaten.
BGlobal fast-path positive controlAn explicit cheating control that switches a direct path centrally.Allowed to be the fastest condition. Beating it on raw speed is NOT the target and would be suspicious.
CRandom local switchingMatched for number of switch events and energy, but with no address logic.The null the doorway must beat. If it is not beaten, there is no result.
DTrigger below thresholdEstablishes the false-trigger rate and the lower edge of the window.Should almost never recruit a corridor.
EDestination-label shuffleAddresses randomly permuted while everything else is held fixed.Advantage should collapse. If it does not, the address logic is not doing the work.
FAddress disabledLocal threshold/recovery law with no address guidance at all.Isolates how much of any advantage is addressing versus excitability.
GFault injectionEdge failure, node failure, or address corruption injected mid-run.Quantifies robustness and added latency, not headline performance.

Primary endpoints · preregistered definitions

EndpointDefinitionUnitValue
valid_run_rateCounted runs surviving every invalid-run rule, over all attempted runs.fractionNO DATA
trigger_thresholdSeed amplitude at 50% recruitment, plus the hysteresis / operating window edges.V (and window in V)NO DATA
false_trigger_rateCorridor recruitment observed with no trigger applied, or below the threshold window.fractionNO DATA
max_active_fractionPeak simultaneously active nodes divided by N, over the run.fraction of NNO DATA
total_active_node_timeIntegral of active node count over the run duration.node·sNO DATA
payload_latencyTrigger timestamp to decoded arrival at B.sNO DATA
C0_latencyOrdinary-phase latency for the SAME A/B pair on the same hardware.sNO DATA
Q_timeC0_latency / payload_latency. Reported with measurement uncertainty attached.ratioNO DATA
payload_identity_errorComponent-wise reconstruction error and symbol error rate at the decoder.error / SERNO DATA
restoration_durationDecoded arrival to full autonomous return of every local coupling state to C0.sNO DATA
electrical_energySubstrate electrical energy from trigger through restoration.JNO DATA
control_plane_energyController / FPGA energy and total control message count, recorded SEPARATELY.J, countNO DATA
illicit_nonlocal_write_countController writes violating the local-neighbour rule. MUST BE 0.countNO DATA
fault_recovery_successSuccessful reroute rate and added latency under injected faults.fraction, sNO DATA
concurrent_latency_tailp95 and p99 latency under multiple simultaneous payloads.sNO DATA

First-pass criteria · N = 16 / 32 / 64

proposed thresholds · not measured results
#CriterionN = 16N = 32N = 64
1illicit_nonlocal_write_count on every counted run= 0= 0= 0
2false-trigger rate, ≥ 100 baseline/control trials per configuration≤ 1%≤ 1%≤ 1%
3trigger success above the frozen threshold window≥ 95%≥ 95%≥ 95%
4payload identity / symbol success on clean runs≥ 99%≥ 99%≥ 99%
5autonomous restoration without global reset≥ 99%≥ 99%≥ 99%
6max_active_fraction (amendable only BEFORE data)≤ 0.35≤ 0.20≤ 0.12
7Q_time planning target — report exact observed values regardless> 1> 1.5> 2.5
8electrical + control-plane energy reported, favourable or notreportedreportedreported
9successful reroute under one injected edge failure, mature configuration≥ 90%≥ 90%≥ 90%
10qualitative claim of new physics from v0.1NONENONENONE
  • Criterion 7 is a PLANNING TARGET. Any latency advantage must additionally exceed measurement uncertainty AND beat the matched random-switch control C. Beating C0 alone is not sufficient.
  • Criterion 8 permits no energy-pass claim of any kind until total energy is compared against a defined baseline task on the same hardware.
  • All thresholds on this table are PROPOSED and were fixed before any component was purchased. They are not measurements.

Instrumentation channels

ChannelWhat is capturedRate / coverage
OSCPer-node oscillator voltage / quadrature phase≥ 20× node characteristic frequency, all N nodes
STATELocal c_i (corridor) and r_i (recovery) state per nodeevery local-law tick
COUPLEPer-edge coupling gain / switch stateon change, plus periodic full snapshot
WRITEEvery controller write: source node, destination edge, value, timestamp, origin flagappend-only, on event
TRIGTrigger waveform as injected at Afull capture
PAYPayload waveform at A and decoded output at Bfull capture, both ends
PWRSubstrate current/voltage traces and control-plane rail, separated≥ 1 kHz, independent rails
ADDRCompact address state held at each nodesnapshot pre-run and post-run, plus on corruption events
FAULTInjected edge/node/address faults with timestampson event
METATopology and seed, software/FPGA git hash and config, environmental metadataonce per run, hashed

Invalid-run rules · the run is void, not repaired

  • VOIDA nonlocal controller write directly configures the full route.
  • VOIDA destination beacon is broadcast across all nodes.
  • VOIDGlobal fast-path switching is used during a counted run.
  • VOIDA hidden precomputed shortest-path sequence is pushed to all nodes.
  • VOIDAny manual intervention after the trigger.
  • VOIDMissing timestamps or missing log segments.

Required raw-data export · per run

topology and RNG seed
all oscillator voltages / phases over time
local c and r state per node over time
switch and coupling state over time
controller writes with source, destination and timestamp
trigger waveform
payload waveform
decoded output at the destination
power and current traces, substrate and control plane separated
fault events
address state
software and FPGA git hash and configuration
environmental metadata

Raw exports are published for every counted run, including invalid and failed runs, with the invalidating condition named. A run that is not exported did not happen.

Scaling rule

Advance N = 16 → 32 → 64 only if the previous size passes the validity, identity and restoration gates (criteria 1, 4 and 5).

Fit latency, active footprint, energy and control overhead against N.

No O(log N) claim from three points. Any statement about scaling is a FINITE-SIZE TREND over N ∈ {16, 32, 64} and must be written as such.

What would falsify us?

Written before the bench exists. Any one of these conditions ends the v0.1 claim; none of them can be argued away after the fact.

X1Only global controller scheduling produces an advantage.

The architecture is a centralized switch with extra steps. The local claim is dead.

X2The local doorway never beats the matched random/local switching control C.

The address logic contributes nothing; the effect is switching activity, not routing structure.

X3Energy or control overhead scales away the latency effect.

Any speed-up is bought, not earned. Reported as a FAIL, not as a trade-off.

X4Routing or address corruption makes the effect nonrobust.

The corridor is a fragile coincidence of one address assignment.

X5Restoration requires a global reset.

The doorway is not autonomous and the two-phase structure is not realized.

X6Measured behaviour is fully explained by conventional switched-network dynamics with no distinctive predictive value.

We report that we built a switched network. That is the honest outcome and it stays on the site.

For a ten-year-old

Now we stop changing the idea. We write down exactly what the machine must do before we build it, so we cannot change the rules after seeing the answer. If the machine fails, we learn. If it passes, we build a bigger one.