Ecphory

Event-first memory experiments for testing whether repeated imperfect event observations consolidate into stable reusable temporal motifs.

The current prototype asks a narrow question:

Can an event-first hierarchy preserve reusable temporal structure more reliably than exact event recovery alone?

Core Result

The strongest current metric is Representation Gain:

Representation Gain =
  canonical length-3 motif support fraction
  - exact canonical motif recovery

Positive gain means recurrence across episodes concentrates support around canonical temporal motifs beyond what exact event recovery predicts.

In the useful regimes, Representation Gain is positive. In the over-entropic union detector regime, it is negative. Null controls collapse or sharply reduce the gain.

That supports the bounded claim:

too little recall             -> undercapture
too much candidate entropy    -> negative Representation Gain
stable imperfect recurrence   -> positive Representation Gain

Pipeline

raw signals
  -> event detector
  -> discrete event decoder
  -> event transitions
  -> episodes
  -> motifs
  -> static projections only afterward

The detector and event identity are intentionally separated. Detectors produce candidate event envelopes. Decoders canonicalize event identity. Motifs are mined afterward as recurring temporal structure.

Run

python -m pip install -r requirements.txt
python event_graph_experiment.py --motif-fault-tolerance-sweep
python event_graph_experiment.py --null-ablation-package

Important experiment modes:

python event_graph_experiment.py --dropout-spurious-pareto
python event_graph_experiment.py --temporal-decoder-seed-sweep
python event_graph_experiment.py --motif-fault-tolerance-sweep
python event_graph_experiment.py --null-ablation-package

Key Artifacts

• event_graph_experiment.py: synthetic signal generation, detectors, decoders, motif mining, nulls, and reports.
• docs/ldgr-research-methodology.md: falsification-first methodology.
• docs/detector-interface-hypothesis.md: detector/representation interface framing.
• outputs/representation_gain_findings.md: Representation Gain summary.
• outputs/motif_fault_tolerance_findings.md: motif-layer fault tolerance results.
• outputs/null_ablation_findings.md: stronger nulls and hierarchy ablations.
• outputs/null_ablation_package.csv: raw null/ablation metrics.
• outputs/motif_fault_tolerance_sweep.csv: raw Representation Gain and motif support metrics.

Current Interpretation

This is not a claim about general machine knowledge.

The current claim is narrower:

Under controlled combined dropout and spurious corruption, an event-first hierarchy can concentrate recurring motif support beyond exact event recovery, provided the detector produces a stable but imperfect event stream rather than an over-entropic candidate envelope.

The hierarchy appears to act like an error-correcting mechanism at the representation level, but no coding-theoretic model is claimed yet.