Skip to main content
The Artifact

The cascade. Tier by tier.

Pick a tier in the rail. See what it does, the joule range it spans, the model it runs, and the receipt it signs. The whole cascade ships behind one API.

I/O
Tier
Ingress
0.01–0.05 J
per call

Query intake + canonicalization

Receives the query, normalizes whitespace, applies the user's policy bundle, attaches the caller's joule budget. Doesn't think — just shapes the work.

Model
Rule-based (deterministic)
Sample receipt
JWP ReceiptPayload
kind "txai.ingress.normalize"
caller "client_a3f1c91a"
joule_budget 100
policy_id "realestate-prod"
joules 0.01
cite "verity-cascade ingress v0.6"
sig "ed25519:0x4f...c1a"
Anatomy — operational specs
Latency p99
1.8 ms
Throughput
120k qps / core
Failure mode
Reject + cite policy violation
Tier
Memoization
0.02–0.4 J
per call

Cache lookup by query fingerprint

Hashes the canonical query (blake3), checks a vector-indexed cache. Hit → returns prior answer + provenance, skips upstream tiers. Miss → forwards. Hit rate above 35% in production for repeated query classes.

Model
blake3 + ANN index (HNSW)
Sample receipt
JWP ReceiptPayload
kind "txai.memoize.lookup"
query_hash "blake3:c0ff…ee01"
hit true
distance 0.018
saved_joules 2310
joules 0.04
cite "verity-cascade memoize v0.6 · ttl=14d"
sig "ed25519:0x4f...c1a"
Anatomy — operational specs
Hit rate (prod)
35–62% by query class
TTL default
14 days
Eviction
LFU + cost-weighted
Ω
Tier
Imagine
10–800 J
per call

Perception — spatial / visual / audio grounding

Custom non-transformer architecture for perception tasks. Grounds spatial intent ("30 minutes to the lab"), parses visual queries, embeds audio. The most expensive tier when called — typically skipped on cache hits.

Model
Imagine v2 (custom, ~12B params)
Sample receipt
JWP ReceiptPayload
kind "txai.imagine.ground"
modality "text+geo"
tokens_in 84
tokens_out 316
datacenter "us-west-1"
joules 240
cite "Imagine v2 · datacenter:us-west-1"
sig "ed25519:0x4f...c1a"
Anatomy — operational specs
Cold latency
1.4–4.2s
Cache hit miss-rate
55% triggers a full tier run
Determinism
Seeded + reproducible at receipt level
λ
Tier
Semantic
20–400 J
per call

Language — entity linking, constraint expansion

Parses the constraint set, runs entity linking against the InformationOS graph, expands implicit constraints (e.g., "solar-ready" → roof orientation + HOA tier + interconnect status). Output: a fully-explicit constraint vector.

Model
Semantic v2 (custom, ~8B params)
Sample receipt
JWP ReceiptPayload
kind "txai.semantic.parse"
constraints_in 5
constraints_out 17
linked_entities 9
joules 32
cite "Semantic v2 · datacenter:us-west-1"
sig "ed25519:0x4f...c1a"
Anatomy — operational specs
Cold latency
0.8–2.4s
Output schema
InformationOS Constraint v3
Determinism
Seeded + reproducible
Tier
Math-Ground
1–60 J
per call

Deterministic constraint solver

Pure-function reasoning. Takes the explicit constraint vector, queries the current state, returns the solution set. No randomness, no temperature. Cheap by design — and always re-run because cached deterministic answers expire as the world changes.

Model
Math-Ground v3 (custom solver, not a neural net)
Sample receipt
JWP ReceiptPayload
kind "txai.math.solve"
constraint "price ≤ 640000 ∧ bedrooms ≥ 3 ∧ solar"
candidates 318
solutions 14
reproducible true
joules 7.2
cite "Math-Ground v3 · pure-function"
sig "ed25519:0x4f...c1a"
Anatomy — operational specs
Latency p99
120 ms / 1000 candidates
Determinism
Bit-reproducible
Fallback
Returns ∅ rather than guess
Tier
Composer
0.4–8 J
per call

Merge tier outputs into ranked result

Joins memoized perception + memoized language + fresh deterministic results. Retains provenance: each output field points back at the tier that produced it. The result is a typed object, not free text.

Model
Rule-based + scoring
Sample receipt
JWP ReceiptPayload
kind "txai.compose.merge"
tier_outputs 3
result_rows 14
provenance "all"
joules 1.06
cite "verity-cascade composer v0.6"
sig "ed25519:0x4f...c1a"
Anatomy — operational specs
Latency p99
8 ms
Output type
Typed (JSON-Schema validated)
Provenance
Per-field tier attribution
Tier
Receipt Signer
0.05–0.2 J
per call

Sign + emit the JWP ReceiptPayload

Ed25519 signature over the canonical JSON: joules total per tier, memoize stats, datacenter, model versions, the result hash. The signed receipt IS the audit trail. Compliance, billing, and Insights consume it directly.

Model
Ed25519 + canonicalization
Sample receipt
JWP ReceiptPayload
kind "txai.cascade.query"
tier "imagine→semantic→math"
memoize { hit: true, saved_joules: 2310 }
result_hash "blake3:abcd…1234"
joules_total 8.4
joules 0.07
cite "verity-cascade v0.6 · datacenter:us-west-1"
sig "ed25519:0x4f...c1a"
Anatomy — operational specs
Sign latency
0.4 ms
Format
JWP ReceiptPayload v1
Verified by
ComplianceOS, Insights, downstream pillars

Why a cascade and not a model

One model is one shape. A cascade is a budget.

The cheapest tier that solves the query wins. Memoization eats repeat work. Math-Ground does the deterministic reasoning no model can be trusted with. The bill is dominated by deterministic compute, not generation — and the receipt proves it.