CAT SYNAPSE
Architecture for Deterministic, Predictive, Hallucination-Free
Catastrophe Intelligence.
Every score is reproducible, bounded, replayable. Every prediction carries calibrated confidence and provenance. No generative model invents a fact that drives a real-world response.
Three properties autonomy must have — simultaneously.
When the system contacts 992 policyholders without human review, none of these is optional.
DETERMINISTIC
Identical inputs produce identical outputs, byte-for-byte. Every score and routing decision is a pure function of versioned inputs and versioned model weights. Replay is exact.
PREDICTIVE
The system forecasts severity escalation, boundary expansion, and time-to-impact — emitting these as typed, bounded, confidence-tagged values. Computed by frozen models, not invented by generative ones.
HALLUCINATION-FREE
No generative model produces, modifies, or overrides any value that drives an autonomous action. Generative cognition is confined to language about already-decided facts; every output passes a validator gate.
Six engineering disciplines.
Each materialised in one or more architecture layers — disciplines define the contract, layers deliver it.
Pure-function math
Every score is f(x)→y with no I/O, no state, no RNG. Replay is byte-exact.
Frozen-weight predictive inference
DSTCEs with pinned weight hash, greedy decode, typed I/O, calibrated probability.
Validator gates at every boundary
Type · range · enum · reference · monotonicity · consistency. Failing gate = hard halt.
Generative cognition confined to language
LLMs produce only natural-language strings. No computing. No deciding.
Forecast as signal, not as narrative
Predictions are typed {tier, p, ci95, horizon} — never generated text.
Versioned, replayable everything
Raw events, signals, weights, code, calibration, outcomes — all versioned, append-only.
Pure-Function Math.
Every autonomous-action score is a pure mathematical function.
Replay-exact
Same inputs → same value, byte-for-byte, on any machine, at any time.
Unit-testable
Every function has a fixed input → fixed output test set. CI fails on drift.
Auditable
The function definition is the audit. No post-hoc explanation needed.
Compositional
PH Score = f(CAT × Proximity × Embedding) inherits the guarantee.
No LLM in the calculation path. No 'helper' classification calls inside scoring functions. Generative models cannot make the byte-equality guarantee even at temperature zero.
Proximity Δ
Hist. Embedding
no I/O · no RNG
no global state
The function definition is the audit.
A regulator asking "why was this score 9.57?" receives a function reference and an input vector — not a generated explanation.
Frozen-Weight Predictive Inference.
Predictive is not probabilistic-in-the-wrong-places. A constrained Transformer is a function.
CAT Synapse uses 14 Domain-Specific Transformer-Based Cognitive Engines (DSTCE). Each forecasts — and each is made deterministic by six constraints.
Frozen weights
SHA-256 model hash recorded per inference
Greedy decoding
Temperature 0 · top-k 1 · no sampling
Pinned tokenisation
Tokeniser version hashed alongside weights
Typed inputs
Structured records — not free-form text
Typed outputs
Enum + calibrated probability — never prose
Held-out calibration
Probabilities from validation set, not self-report
- Severity tier ∈ {LOW, MED, HIGH, CRITICAL} + probability
- Likelihood ∈ {Likely, Observed} + probability
- Immediacy ∈ {Immediate, Imminent, Forecast} + probability
- Peril sub-type + probability
A DSTCE under these constraints behaves like a function — input → label, label → probability. Replay with the same inputs against the same model hash and you get the same output. Every time.
Validator Gates at Every Boundary.
Determinism inside a layer is not enough — every value crossing a boundary is checked.
Type conformance
Every field present, every type correct, no nulls in required positions
Range bounds
severity ∈ [0,10] · proximity ≥ 0 · p ∈ [0,1] · window.end > window.start
Enumeration
severity_tier ∈ {LOW,MED,HIGH,CRITICAL} · likelihood ∈ {Likely, Observed}
Reference integrity
County codes resolve in gazetteer · PH IDs in register · model hashes in registry
Monotonicity
event_version strictly increasing · severity escalations require data witness
Consistency
policyholder_scores entries covered by affected_geo.counties
Not a soft warning. Not a fallback. The signal does not propagate. The event is held in quarantine with the gate's diagnostic recorded.
Bad predictions cannot trigger autonomous responses. Model regressions are visible operationally — not weeks later.
L1 → L2 · L2 → L2.5 · L2.5 → L3 · L3 → L4 · L4 → L0 feedback
Every layer boundary runs its own gate.
Generative cognition confined to language.
Three classes of inference, partitioned by determinism level.
Every generated message is validated against the signal it claims to describe. Names a county not in affected_geo, references a severity not in the signal, invents a time window — it is rejected and regenerated. Never delivered.
Forecast as Signal, not as Narrative.
Predictions are typed values produced by deterministic models — never generated text.
Severity forecast
DSTCE-projected severity at t+30 / 60 / 120 / 240 min, each with calibrated probability.
Boundary forecast
Peril-specific evolution model (flood routing, fire spread, hurricane track) projecting affected geography at future timesteps.
Exposure forecast
Joining the boundary forecast against the policyholder register → projected PH count by future timestep.
"forecast": { "model_id": "FLD-EVOL-1.4.2", "model_hash": "sha256:9c8a…", "horizon_minutes": [30, 60, 120, 240], "severity_path": [ { "t_plus": 30, "tier":"HIGH", "p":0.91, "ci95":[0.84,0.96] }, { "t_plus": 60, "tier":"HIGH", "p":0.83, "ci95":[0.72,0.91] }, { "t_plus": 120, "tier":"CRITICAL", "p":0.67, "ci95":[0.51,0.79] }, { "t_plus": 240, "tier":"CRITICAL", "p":0.74, "ci95":[0.55,0.86] } ], "exposure_path": [ /* projected PH count per horizon */ ], "computed_at": "2026-01-17T14:34:08Z" }
L3 may narrate this in stakeholder language. L3 may never produce it.
Versioned, Replayable Everything.
The operational definition of determinism is replay.
"A system is deterministic to the extent it can be replayed." CAT Synapse can replay any past CAT event end-to-end and reproduce every score, every routing decision, every alert. That is the bar.
Raw events
v1 → vN append. Event 252793 traversed v1 → v8 across 8 NWS reissues.
Risk Signals
Each recomputation emits a new signal version. Monotonic counter.
Model weights
SHA-256 content hash recorded in every DSTCE inference call.
Pre-processing
Tokeniser version + normalisation code version pinned per inference.
Decision code
Pure-function math layer ships with git SHA pinned to deployed binary.
Outcomes
Delivery confirmations · acknowledgements · claims correlations append-only.
Architecture overview · L0 → L4.
Six disciplines materialised across five layers + one typed interface.
The Equation of Intelligence.
Three risk metrics — pure-function composite.
CAT Risk Score
Event severity — asset-independent
- DSTCE severity + likelihood
- Alert structured parameters
- Real-time intelligence feeds
tier ∈ {LOW · MED · HIGH · CRITICAL}
Pure function over DSTCE output and feed parameters
Risk Proximity Score
Spatial relationship — asset to event
- PH latitude / longitude
- Event polygon
- Distance to coast · urban class
calculate_proximity_score(lat, lng, geo)
Recalculated continuously as boundary evolves (C-4)
Policyholder Risk Score
Composite priority — the autonomy trigger
- CAT Score
- Proximity Score
- Historical Vector DB embedding
HIGH RISK PRIORITY
Worked example · PH-100017 · REFID-158053
All three metrics are pure functions. Same inputs → same scores, byte-for-byte, on any machine.
The Risk Signal Layer.
The only artefact that crosses the boundary between deterministic math and constrained cognition.
Typed values, not raw text
L3 reads enums and numerics; severity is not interpreted, it is read.
Embedded constraints
Constraints block enforced by the L3 orchestrator and conformance gate.
Provenance attached
Engine version + model hash + code SHA + input hash on every signal.
Versioned propagation
Recomputation emits a new signal version; old versions retained for replay.
Conformance gate downstream
Every L3 output validated against the signal before leaving the layer.
{
"event_id": "REFID-158053",
"event_version": 3,
"peril_class": "WILDFIRE",
"classifications": {
"severity_tier": "HIGH", "severity_p": 0.94,
"likelihood": "Observed", "likelihood_p":0.97,
"immediacy": "Imminent", "immediacy_p": 0.88
},
"scores": {
"cat_score": 7.8,
"affected_geo": {
"state":"TX",
"counties":["Brewster","Pecos","Terrell","Val Verde"]
}
},
"forecast": { /* typed severity & exposure paths */ },
"policyholder_scores": [ /* 992 entries */ ],
"constraints": {
"l3_must_not_recompute_scores": true,
"l3_must_not_modify_classifications":true,
"l3_must_not_invent_geography": true
},
"provenance": {
"engine_version":"L2-DPE-1.4.2", "code_sha":"git:9a2f4e8",
"input_hash":"sha256:c1b3…", "replayable": true
}
}
Determinism scorecard.
How each layer satisfies the three core properties.
Every numeric that triggers an autonomous action is replayable from versioned inputs. Every forecast carries typed horizon, probability, and confidence. Every generated message has been validated against the signal it describes.
Patent claim mapping · C-1 through C-4.
Each claim anchored to the architecture and to a specific inventive step.
Layered Autonomous CAT Risk Pipeline
End-to-end L1→L4 pipeline; no human arbitration; frozen-weight DSTCE inference; deterministic decision weighting.
Full autonomy survives audit because every node is deterministic; replay is exact.
Composite PH Risk Scoring Engine
L2 pure-function composite — CAT × Proximity × Historical Embedding → PH Score.
A single autonomy-triggering score that is reproducible byte-for-byte from versioned inputs.
Synaptic Activation & Self-Organizing Optimization
L4 dynamic topology activated per event trajectory; pathways auto-prune on subsidence.
The architecture itself is automated; not a static workflow with a runtime.
Event Impact Evolution & Autonomous Retraining
Versioned signal propagation; closed-loop Vector DB enrichment; out-of-band model promotion preserves determinism.
Continuous recalibration without sacrificing the replay guarantee — retraining is versioned, not in-place.
Live system proofs.
The architecture in operation against real NWS-issued events.
WILDFIRE
Red Flag Warning
REFID-158053
FLOOD
Mississippi River Warning
Event 252793
CAT SYNAPSE
Deterministic because every score is a pure function of versioned inputs.
Predictive because frozen learned models produce typed, calibrated forecasts inside a deterministic envelope.
Hallucination-free because generative cognition is structurally confined to language about already-decided facts, and every generated artefact is validated against the signal it claims to describe.
Patent Pitch v9 · FIG.1–FIG.8 · Claims C-1 through C-4
Strictly Confidential — Attorney-Client Privileged