DPSH Temporal Motifs

DPSH Temporal Motifs — A Conservative Concept and Validation Plan

LEGACY—FALSIFIED. This branch is a closed research artifact retained for reproduction. It must not be used as the active CCDL observation engine or as a blueprint for constructing a DPSH network. The active revision is CCDL v0.2 in Docs/cognia-causal-discovery-loop/cognia-causal-discovery-loop.md.

Document status

Final experiment status: FALSIFIED / CLOSED. The complex DPSH model did produce a reproducible temporal skeleton, but it added no held-out predictive value. Its activity mostly tracked spike density, an activity-preserving surrogate produced an even richer DAG, the skeleton was not specific to the edge-of-chaos regime, and the current sequential representation failed an isomorphic population-relabeling control. The end-to-end analytical path was also 11.50x slower than the equivalent task without the DAG. DPSH temporal DAGs are therefore a research dead end under this formulation. The retained implementation is falsification evidence and reusable experimental apparatus, not a recommended runtime module or a basis for an internal consciousness network.

This document describes an experimental analytical extension of DPSH. It is not new neuron physics, a new learning rule, or a completed theory of metastability.

Its purpose is to test a limited question:

Does the precise DPSH spike stream contain recurring temporal motifs that carry useful information beyond firing rates, state trajectories, and simple temporal statistics?

Only a positive result would justify studying the composition of motifs into higher levels. Inspiration from Levelized-CDCT means only the gradual composition of structures that have already been validated. It does not mean transferring the CDCT computational model into DPSH or assuming that the discovered structure must itself be a neural network.

1. What DPSH already provides

DPSH is an asynchronous event-driven spiking runtime. The following existing properties are relevant to this experiment:

  • model time is an integer ModelTime expressed in nanoseconds,
  • event ordering is deterministic,
  • spikes propagate through explicit synapses,
  • every synapse has a stable EdgeId, polarity, weight, and positive delay,
  • a synaptic event carries its source, target, time, and edge identity,
  • the runtime provides a canonical spike log,
  • the recorder can store spikes, state samples, and limited provenance,
  • snapshots, forks, replay, frozen randomness, ablations, and perturbations already exist,
  • trajectory, macrostate, dwell-time, transition, and surrogate analyses exist,
  • an observer is not allowed to alter network dynamics.

Temporal motif analysis is therefore not a new way to simulate DPSH. It is a candidate observer-side representation over data that the runtime already knows to a large extent.

2. What DPSH already analyzes and what would be new

Current macrostate analysis primarily operates on sampled state trajectories. It measures state separability, dwell time, transition matrices, transition entropy, and perturbation stability, among other properties.

The proposed analysis would additionally examine the organization of events in time:

A@10 ms → B@14 ms → C@21 ms

This may distinguish dynamics that use similar active neurons and firing rates but differ in event order or relative timing. That is a genuinely different analytical view. It must not be assumed in advance to outperform existing state analysis or a simple spike-sequence model.

3. The time-unfolded graph

The physical synaptic network may contain cycles:

A → B → C
↑       ↓
└───────┘

Individual events are nevertheless separated in time:

A@t0 → B@t1 → C@t2 → A@t3

If every edge points to a later time, the time-unfolded representation is a DAG even though the physical network is recurrent. This is compatible with DPSH, but it does not by itself establish the existence of recurring or functionally meaningful motifs.

4. Three different meanings of an edge

The word “causal” must be used carefully.

4.1 Transmission edge

A spike of neuron A actually created an event delivered to neuron B through a specific synapse. DPSH knows this relation exactly from source, target, EdgeId, spike time, and synaptic delay.

spike A@t ──edge 17, delay 4 ms──> delivery B@(t+4 ms)

4.2 Candidate contribution edge

A delivery to B physically and temporally preceded a later spike of B and may have contributed to it. This does not mean that it was the sole or necessary cause. The spike may depend on several inputs, previous state, adaptation, refractoriness, oscillation, predictive modulation, and stochastic firing.

4.3 Intervention-supported causal edge

Removing or perturbing the candidate input from the same snapshot measurably changes the later spike or trajectory of B. Only such a relation has stronger causal support.

The first experiment may construct transmission and candidate-contribution graphs. Without intervention-based validation it must not call the result an exact causal DAG. The working name is Temporal Influence DAG.

5. Data that is currently missing

The production windowed path preserves the exact synaptic event during fan-out, but represents a committed firing result only as (time, neuron). The recorder therefore records the spike without an incomingCause on this path.

Possible temporal relations can be reconstructed from the spike log and topology. It is not currently possible to identify exactly which delivered events formed the context of a particular firing. A more precise experimental mode will probably need to observe:

  • delivered event identity and EdgeId,
  • delivery source, target, time, and contribution,
  • the set of recent inputs preceding a firing,
  • firing type and relevant neuron state,
  • the relation between the firing, its window, and a possible thinning candidate.

Instrumentation must be optional and bounded, and it must satisfy invariant I-8: a run with and without the analyzer must produce an identical event log.

6. Working hypotheses

H1 — additional temporal information

Temporal motif features distinguish at least some dynamic regimes better than firing rates, population vectors, and simple pairwise temporal statistics.

H2 — relation to macrostates

Some state-space macrostates have a reproducible distribution of temporal motifs. It is not assumed that every macrostate corresponds to a single motif.

H3 — dynamic regimes without state separation

Two dynamics may have similar distributions of state values but different organizations of spike events.

H4 — value of hierarchy

Composing motifs into a higher level improves held-out prediction or compression over Level 1. This is a separate hypothesis and may fail even if H1 succeeds.

7. Minimal representation

Level 0 — events

The basic record is not necessarily limited to (neuron, time). Depending on available instrumentation it may contain:

TemporalEvent {
    event_id,
    neuron,
    time,
    kind,
    incoming_edge?,
    source_event?,
    contribution?
}

Level 1 — short temporal motifs

A motif is a small directed graph with temporal constraints:

A --[3..5 ms]--> B --[5..9 ms]--> C

Its minimal identity includes:

  • node types or roles,
  • edge directions and types,
  • relative-delay intervals,
  • optionally synapse polarity and group,
  • a motif-canonicalization version.

Absolute time is not part of motif identity.

Higher levels

A higher-level node may refer to several occurrences of lower-level motifs. The analytical expression must be lowered into a physical network according to its dependency DAG, rather than being attached afterward as a flat readout. For an edge u -> v, the construction must satisfy

level(v) >= level(u) + 1

and each level becomes a separately testable neuron block. Residual input may be carried forward explicitly when a later expression still depends on it. This is the mandatory construction form for the experiment, not evidence that a deep hierarchy is useful. The hypotheses are whether the derived nodes, blocks, and additional depth improve held-out prediction, compression, extrapolation, or intervention response enough to justify their cost.

8. Equivalence and tolerance

Exact timestamp equality would be too brittle. The first version should use explicit time bins or intervals:

Δt(A,B) ∈ [3 ms, 5 ms]

Invariance to small jitter, global time translation, mild temporal scaling, one missing spike, and one inserted unrelated spike should be tested separately.

These invariances must not all be enabled automatically. Each increases the number of accidental matches. Tolerances must be declared by the experiment and penalized in model cost. The first implementation does not need general graph isomorphism; short bounded motifs can use canonical node ordering and hashes over quantized delays.

9. Motif metrics

A single weighted score with freely selected coefficients would make it too easy to obtain a desired result. The first phase should report separately:

  • occurrences per unit of model time,
  • support across independent seeds and runs,
  • precision and recall on synthetic ground truth,
  • predictive gain on held-out spike-stream continuations,
  • compression gain after accounting for motif-dictionary cost,
  • stability under declared jitter,
  • selectivity for an experimental condition,
  • survival or disappearance under a control intervention.

Any later aggregate score must be pre-registered.

10. Required baselines

The Temporal DAG must not be compared only against firing rate. The minimal control set is:

  1. firing rate per neuron or group,
  2. set of active neurons,
  3. population or state vector,
  4. pairwise-delay histogram or cross-correlation,
  5. spike n-gram or another simple sequence model,
  6. Level-1 temporal motifs,
  7. optionally Level-2 composition.

If Level 1 merely reproduces an n-gram or pairwise-delay result at a higher cost, the hypothesis that a graph representation adds value is not supported.

11. Smallest falsification experiment

Prepare two controlled regimes with the same neuron set, similar firing rates, similar total activity, and similar marginal distributions of state variables, but different propagation order:

regime A: A → B → C → D
regime B: A → C → B → D

Procedure:

  1. Generate several seeds and separate train, validation, and test runs.
  2. Verify that simple rate baselines genuinely fail to distinguish the regimes.
  3. Derive Level-1 motifs only from the training partition.
  4. Measure classification and prediction on the test partition.
  5. Repeat after time shuffle, edge rewiring, and phase-scramble controls.
  6. Remove a candidate input from a shared snapshot and measure causal impact.

Success requires reproducible held-out gain over simpler baselines and loss of that gain under a control that destroys temporal structure. One visually impressive DAG from one run is not a result.

12. Relation to metastability

Current DPSH defines macrostates over trajectories of state features. Temporal motifs must not replace that definition before the experiment. Useful questions include:

  • Does the motif distribution change at transitions between existing macrostates?
  • Does a motif predict a future transition better than the state vector alone?
  • Do stable motif regimes exist within one state-space macrostate?
  • Is motif stability reproducible across seeds and matched controls?

Possible outcomes include redundancy with state analysis, refinement of within-macrostate transitions, unstable random motifs, or earlier prediction of a transition. The statement “metastability is a Temporal DAG” would be premature. These are competing or complementary representations that must be measured.

13. Memory and computational cost

For S spikes and average fan-out d, an explicit transmission graph contains approximately O(S) nodes and O(S·d) synaptic transmissions. Candidate-contribution edges additionally grow with attribution-window width. General subgraph discovery may be combinatorial, so the experiment must be bounded by:

  • a finite time window,
  • maximum motif node count and depth,
  • minimum support,
  • fixed delay bins,
  • candidate and memory limits,
  • streaming counters where possible,
  • explicit observer-overhead measurements.

The first prototype should search for short paths and small branching motifs, not arbitrary subgraphs.

14. Recommended phases

T-01 — observational data contract

Define the transmission-event record and verify I-8. Do not implement pattern mining yet.

T-02 — synthetic ground truth

Find known motifs in manually constructed spike streams and reject noise, broken provenance, and motifs outside the declared tolerance.

T-03 — Level-1 motif baseline

Implement bounded canonicalization of short motifs and compare it with rate, pairwise-delay, and n-gram baselines.

T-04 — matched DPSH experiment

Run two regimes with similar activity and different timing across several seeds, snapshot forks, and surrogate controls.

T-05 — intervention-based validation

Validate candidate-contribution edges by ablation or perturbation from a shared snapshot. Stronger causal claims become possible only here.

T-06 — compile the analytical DAG into a levelized network

Lower every accepted analytical dependency into a positive-delay DPSH edge, preserve provenance and topological roles, and test the physical network end to end. Keep additional depth only when it improves held-out prediction or MDL compression after accounting for its cost.

15. Conditions for rejecting the direction

The direction should not be considered useful if:

  • motifs are not reproducible across seeds,
  • small physically reasonable jitter destroys them completely,
  • a cheaper n-gram or pairwise statistic gives the same result,
  • motifs add neither held-out prediction nor compression,
  • motif stability is unrelated to any measurable DPSH regime,
  • the result survives time-shuffled or edge-rewired surrogates,
  • attributed “causal” edges do not survive intervention,
  • cost is unbounded or the observer changes the event log or RNG consumption,
  • Level 2 adds no value over Level 1.

A negative result is a valid experimental result.

16. What not to do yet

  • Do not change neuron dynamics, STDP, or the learning gate.
  • Do not feed discovered motifs back into the network.
  • Do not control the workspace from analyzer output.
  • Do not present a possible temporal relation as proven causality.
  • Do not implement unbounded graph-mining algorithms.
  • Do not add unbounded Level N without a budget, provenance, and held-out ablation.
  • Do not connect the result to claims about consciousness or qualia.

17. Technical conclusion

Final verdict — compatible as an apparatus, functionally falsified in the tested formulation.

The reason is not the mere existence of a DAG; unfolding a recurrent network in time is an expected representation. The interesting and unverified possibility is that small temporal motifs carry information about dynamic regimes that is not captured by current state analysis or simpler sequence baselines.

Current DPSH provides precise time, synaptic delays, determinism, replay, interventions, and a read-only observer boundary. For precise firing attribution, however, the production windowed path does not yet retain the full relation between delivered events and the resulting spike. The first phase must therefore operate on transmission and candidate-contribution edges, validating stronger causal claims through counterfactual runs.

A hierarchical levelized plan is the construction form in which an analytical DAG must run: the network follows the expression dependencies instead of receiving a flat post-hoc readout. The tested hierarchy produced a small, reproducible train-only temporal skeleton, but that skeleton did not improve held-out prediction, failed key surrogate and relabeling controls, was not specific to the target dynamic regime, and cost substantially more than simpler representations. Section 18 records the evidence that closes this branch.

18. Implementation and experimental record

This section is the translated consolidated record of the experiments performed after the original proposal. A successful engineering test is not counted as support for the scientific hypothesis: construction invariants and known-truth recovery establish that the apparatus works, while held-out experiments decide whether the representation is useful.

18.1 Known-truth motif tests

The Level-1 implementation used a training-only motif dictionary and the same nearest-centroid classifier for every representation. In the controlled timing assignment task, the motif representation scored 1.0 while the required simple baselines scored 0.5. Swapping neuron order, however, gave both the spike n-gram and the motif representation 1.0; the DAG representation was therefore redundant on that task.

The runtime known-truth experiment forked both conditions from the same snapshot. They had equal spike counts and equal marginal delay distributions, but differed in which delay was assigned to which transition. The motif representation scored 1.0, the n-gram 0.5, and a time shuffle reduced the motif result to chance. This established that the implementation could recover information present only in temporal assignment. An autonomous phase-scramble control then scored 0.833 for motifs and 1.0 for the n-gram, so it did not show incremental value. A degree- and biophysics-preserving edge-rewire test also left both motif and n-gram accuracy at 1.0 after per-neuron rate matching.

A bounded partial-order representation was added to avoid reducing simultaneous or near-simultaneous events to one arbitrary order. It scored 1.0 against 0.5 baselines on its dedicated known truth, but matched the n-gram at 1.0 on the production phase-scramble and rewire tasks. It was correct but still redundant.

18.2 Pre-registration and first complex pilot

The complex experiment pre-registered independent train, validation, and test seeds, 4,000 steps of 5 ms, a washout, three train-fitted macrostates, 100 ms windows, 1 ms bins, motifs of length three, at most eight skipped events, and a 50 ms span. Balanced accuracy was the primary classification metric. Success required a margin of at least 0.05 over the best simple baseline on both validation and untouched test data, together with emergence and surrogate gates.

A four-seed pilot produced 45/18/24 balanced train/validation/test windows. Test accuracy was 0.708 for firing rate, pairwise delay, spike n-gram, and contiguous motifs; active set scored 0.333, and the partial-order model scored 0.667. Its 9,454 features for only 45 training windows were also a clear overfitting warning.

The same apparatus was then run on four independent seeds of the 256-neuron Modules4 model: two train, one validation, and one test. Every run had 800 steps of 5 ms, with the last 200 autonomous. Classification used 45/18/24 balanced 100 ms windows and a stricter partial-order space: length three, 1 ms bins, at most two skipped spikes, and 20 ms span. Validation/test accuracy was:

Representation Validation Test
firing rate 0.778 0.708
spike n-gram 0.778 0.708
Level-1 motif 0.722 0.708
partial-order motif 0.722 0.667

The classification verdict was redundant.

18.3 Continuous endpoint and physical lowering

A continuous known truth used targets 0.2 and 0.8 while holding firing rate, active set, neuron order, and marginal delay histograms equal. Only delay assignment to transitions differed. Level-1 motif RMSE was numerically zero (8.1e-17) on validation and test, versus 0.3 for firing rate, pairwise delay, and spike n-gram. It passed the pre-registered 0.1 improvement margin.

The end-to-end pipeline was then implemented as:

train-only discovery -> occurrence DAG -> levelization -> role-safe
consolidation -> physical DPSH lowering -> held-out regression

The known truth discovered eight motifs and 54 occurrence nodes. Its DAG had 120 observed edges; 36 zero-lag edges were rejected as non-physical rather than shifted artificially. Consolidation yielded 18 modules and 28 edges, and the physical DPSH graph contained 18 neurons and 28 gated synapses. The untouched test retained motif RMSE 8.1e-17 against the 0.3 sequence baseline. This closed the engineering loop, but it remained a synthetic result.

On the complex continuous endpoint, 12/6/6 evenly distributed windows produced 80 motifs, 1,462 occurrence nodes, and 61,029 edges; 3,848 zero-lag edges were rejected. Role-safe levelization yielded 713 neurons and 41,138 gated synapses. Validation/test RMSE was:

Representation Validation Test
Level-1 motif 2.261 2.618
spike n-gram 2.366 2.214
firing rate 3.566 3.308
pairwise delay 2.472 2.356
active set 1.296 2.332

The validation advantage over the n-gram did not transfer to the untouched test. The continuous verdict was not supported.

The pilot also exposed two scaling defects that were fixed without changing the scientific result. The bounded-lag DAG builder had continued scanning beyond its lag horizon; it now stops at the run or horizon boundary. Multiple spikes from one population in the same 1 ms bin are now canonicalized to one (population, bin) event for DAG construction. Direct levelization and consolidation avoids materializing the full occurrence plan. These changes reduced intermediate work but did not remove semantic expansion: 80 motifs still became 713 modules.

18.4 Recurrent compaction and evidence filtering

compactTemporalDagRecurrently was implemented to merge occurrences by motif identity while preserving cycles explicitly. Tarjan strongly connected components become recurrent blocks; only the condensation graph is levelized. Internal edges carry a feedback flag, remain gated, and require positive physical delay. Provenance and total occurrence support are preserved.

On the complex pilot this reduced the physical network from 713 to 80 neurons and from 41,138 to 2,704 synapses, and reduced test time from about 86 to 16 seconds. All 2,704 edges were nevertheless feedback edges and all 80 motifs belonged to one SCC. The permissive definition “any successor within 20 ms” therefore produced one dense recurrent block, not a hierarchy.

A train-only evidence filter was added. For every direction A -> B it measures prevalence across independent training runs and directional lift P(B in the lag window | A) / P(B in the lag window | any motif). A success is counted at most once per source occurrence, and overlapping windows from one seed share an evidence group rather than pretending to be independent replications.

With two training seeds, prevalence 1.0, and lift at least 1.25, the filter rejected 56,218 of 57,181 physically possible occurrence edges. The remainder consolidated to 141 physical edges among 80 motifs, with 60 feedback edges and five dependency levels. A train-only sweep gave:

  • prevalence 1.0, lift 1.0/1.25/1.5/2.0: 391/141/88/47 edges, 343/60/16/2 feedback edges, and 3/5/8/5 levels;
  • prevalence 0.5: 1,678/1,168/953/725 edges and 2/3/4/3 levels.

This showed that the dense SCC was partly an edge-definition artifact, but did not establish predictive usefulness.

18.5 Six-seed replication

Training was expanded to seeds 101, 103, 113, 127, 131, 137, while validation 107 and test 109 remained untouched. Three evenly spaced windows per training seed gave 18 discovery windows and six independent evidence groups.

At prevalence 6/6, lift 1.0/1.25/1.5/2.0 yielded 114/9/2/0 edges, 86/0/0/0 feedback edges, and 3/3/2/0 levels. The two-seed claim of a robust hierarchy over the whole threshold range was therefore falsified. The working pre-registration, prevalence 6/6 and lift 1.25, found 107 motifs but only ten incident motif types. Excluding 97 isolated motifs yielded a genuine sparse levelized network with ten neurons, nine synapses, three dependency levels, and no feedback.

Structural reproducibility still did not add prediction. Level-1 motifs and the spike n-gram both had validation/test RMSE 1.854/1.745; pairwise delay scored 1.777/2.185.

The nine discovered edges were then used directly as a train-only evidence-dag-edge representation. A first validation seed 107 and test seed 109 looked positive: RMSE 1.132/1.406 versus the best baselines 1.777/1.745. A stricter replication left training and all thresholds fixed and expanded validation to 107,139,149 and test to 109,151,157. It falsified the apparent gain: DAG-edge RMSE was 2.671/2.261, while pairwise delay achieved 1.266/1.206.

The same frozen DAG was finally tested as a transition skeleton. Edge lag dynamics supplied 18 features, level flow three, path completion six, and the combined representation 27. Validation selected path completion, but its validation/test RMSE was 2.684/2.151 against pairwise delay 1.266/1.206. Lag dynamics scored 3.021/2.168, level flow 3.092/2.495, and the combination 2.885/2.689. Exact edge timing, level flow, and path completion therefore did not rescue the hypothesis.

18.6 Explanation and falsification controls

Frozen-DAG activity correlated mainly with spike count (r = 0.730), less with 1 ms synchrony (r = 0.322), and still less with instantaneous stimulus state (r = 0.240). Much of the signal was therefore overall activity rather than an independent description of emergence.

Shuffling neuron identities while preserving global spike times and identity counts created a richer, not poorer, graph: the original 10 incident modules, 186 accepted occurrence edges, and three levels became 50/3572/8. The structure survived an activity-preserving surrogate.

More seriously, a global bijective relabeling of population identities changed 10/186/3 into 6/45/3. Simultaneous spikes had been ordered by numeric ID, making sequential motifs dependent on arbitrary labels. The current miner is therefore not invariant under an isomorphic renaming of its nodes.

An activity-matched regime control used recurrent gain 0.5/1.0/1.5. Mean spikes per window were 159.5/162.2/162.8, only 1.68%/0.38% away from the middle condition and safely inside the pre-registered 10% tolerance. Yet the three regimes produced similar structures: 19/542/5, 16/429/5, and 17/393/6 incident modules / accepted occurrence edges / levels. The edge-of-chaos condition had no unique maximum in graph size, edge count, or depth.

18.7 Cost and final decision

In a warmed process, the same 36-window workload was measured three times. The non-DAG variant fit and evaluated the five existing representations. The DAG variant performed the same work plus discovery, causal and evidence filtering, recurrent compaction, levelization, physical lowering, edge readout, and four transition readouts.

Median time was 127.7 ms without the DAG and 1,467.9 ms with it: an 11.50x slowdown. Net overhead was 1,340.2 ms, approximately 37.2 ms per 100 ms window, compared with 3.55 ms per window for the baseline path. This is an end-to-end CPU analysis cost, not a CUDA neuron-runtime benchmark, but it is the relevant cost for the tested held-out task.

Final scientific status: FALSIFIED / CLOSED. A small train-reproducible temporal skeleton exists at one pre-registered threshold, and the implementation can recover synthetic timing-only truths. It does not robustly predict the emergent macrostate, is beaten by simpler held-out baselines, survives a structure-destroying surrogate, fails relabeling invariance, is not unique to the target dynamic regime, and makes the equivalent analysis approximately eleven times slower. DPSH-DAG remains a legacy reproducibility artifact, not an active network-construction path. Reopening the direction would require a new pre-registered target, new untouched seeds, and a representation of simultaneous events that is invariant to node naming.