01
Query
Intake
0.01 J
A user asks the cascade for a house.
"3-bed, solar-ready, under $640k, 30 minutes to the Salk Institute." The query lands at the cascade ingress. Routing decisions start with a hash, not a model call.
The Scenario
One query. Eight steps.
2,310 joules saved by the memoization layer.
Watch the cascade route a real-estate search through Imagine, Semantic, and Math-Ground. A near-identical prior query has already paid the upstream cost. The bill drops to 8.4 J.
02
Hash
Memoization
0.04 J
The query is hashed against the memoization layer.
blake3 fingerprint of the canonical query form. The cascade asks: have we answered this — or something almost identical — before? Lookup is cheap by design.
JWP ReceiptPayload
kind "txai.memoize.lookup"
query_hash "blake3:c0ff…ee01"
lookup_ms 1.4
joules 0.04
cite "verity-cascade memoize v0.6"
sig "ed25519:0x4f...c1a"
03
Hit
Memoization
0.02 J
Hit. A near-identical query 14 days ago covers most of the work.
Cosine distance 0.018 between the new query and a cached answer from 2026-04-23. TTL still valid. The cache provides the Imagine and Semantic tier outputs — the most expensive work was already paid for.
JWP ReceiptPayload
kind "txai.memoize.hit"
match_hash "blake3:7f24…a90c"
distance 0.018
saved_joules 2310
joules 0.02
cite "verity-cascade memoize · ttl=14d"
sig "ed25519:0x4f...c1a"
04
Imagine
Tier 1
0.00 J
Imagine tier — bypassed. 240 J saved.
Perception layer. Normally tokenizes spatial intent ("solar-ready", "commute"), grounds it against a geo-graph. Today: cached output reused; tier is skipped entirely.
05
Semantic
Tier 2
0.00 J
Semantic tier — bypassed. 1,840 J saved.
Language tier. Normally parses the constraint set, runs entity linking, expands the implicit ones (e.g., "solar-ready" → roof orientation + HOA tier + interconnect status). Today: cached output reused.
06
Math
Tier 3
7.20 J
Math-Ground tier — refreshes the price filter. 7.2 J spent.
The deterministic reasoner. Today's listings are different from 14 days ago, and the price ceiling shifted (user's pre-approval went up). Math-Ground re-solves the constraint, fresh — no cache shortcut for deterministic work.
JWP ReceiptPayload
kind "txai.math.solve"
constraint "price ≤ 640000 ∧ bedrooms ≥ 3 ∧ solar ∧ commute ≤ 30"
solutions 14
reproducible true
joules 7.2
cite "Math-Ground v3 · pure-function"
sig "ed25519:0x4f...c1a"
07
Merge
Composer
1.06 J
Merge cached perception + cached language + fresh math.
The composer joins the three tier outputs into a single ranked list. Per-listing provenance retained: each result links to which tier contributed which field.
08
Sign
Receipt
0.07 J
Sign the receipt. Return it with the answer.
Ed25519 over the canonical JSON. Joules total, tiers exercised, memoize stats, datacenter, model versions. The receipt IS the audit trail; rewriting it means forging a signature.
JWP ReceiptPayload
kind "txai.cascade.query"
tier "imagine→semantic→math"
memoize { hit: true, saved_joules: 2310 }
result_hash "blake3:abcd…1234"
joules 0.07
cite "verity-cascade v0.6 · datacenter:us-west-1"
sig "ed25519:0x4f...c1a"
The arithmetic
8.40 J spent. 2310 J saved.
A cold cascade query would have cost ~2,318 J. The memoization layer turned it into 8.4 J. Run the same query class 10,000 times this month and the bill is dominated by the deterministic tier — which is the cheapest, by physics. The system gets cheaper as it learns.