Match
Vertex re-ranker over federated supply. Returns top-k tail candidates with confidence and reasoning codes.
§01 · INTELLIGENCE
Six models.
Six models.
One closed loop.
We don't sell a black box. The matching, pricing, fit, and routing models that run the marketplace are documented here — what they do, what they consume, and how the loop closes back on itself every night.
§02 · CLOSED LEARNING LOOP
Every event, signal.
RFQ, quote, accept, decline, divert, wheels-up — every event in the marketplace becomes a row in the warehouse. Six Vertex models retrain on it nightly. The same diagram appears on the home page; the same component renders it here.
- 01Observe. Every action — RFQ posted, quote returned, principal accepts or declines, sector flown, divert called — written to the warehouse with the reasoning codes that produced it.
- 02Retrain. Six models retrain at 02:00Z on the trailing 90-day window. Validation against held-out 7-day fold. Promotion gate at p95 lift over previous epoch.
- 03Replace. New model promoted only on positive lift. Failed promotion is logged, kept in the warehouse, and the previous epoch stays in production.
- 04Return. Every match returns the seven reasoning codes that produced it. Brokers and operators read the same codes the model used.
§03 · REASONING SURFACE
No black box. Read the codes.
Every match returns a confidence score and the seven reasoning codes that produced it. Same surface, same codes, same plain language for principals, brokers, and operators.
RANK01
CONFIDENCE0.94
REASON CODES7
Codes are stable across model epochs. New codes are added with a SemVer bump and a six-week deprecation window for any retired ones.
§04 · SIX MODELS
What runs. What it consumes.
Six models in production. Each documented with input, output, and retrain cadence. None of them are LLMs — they are gradient-boosted trees and learned re-rankers, served on Vertex AI with versioned promotion gates.
Price
Quote floor / list / settle. Honest pricing — floor never crossed, list informed by clearing data, settle reconciled to wallet.
Fit
Principal–operator compatibility. Cabin class, service tier, prior decline patterns, jurisdictional posture.
Route
Re-route on operator decline or unavailability. Spine fallback when a quoted operator can't lift the leg.
Forecast
Empty-leg sell-through forecast. Posts legs only above operator's price floor; confidence-weighted. (Temporal-Fusion Transformer.)
Risk
Slot, weather, and operational risk delta. Reads telemetry and slot grids; flags exposure for diversion or float.
§05 · BENCHMARKS
Honest numbers, updated.
Three benchmarks we publish. Latency we measure end-to-end. Sell-through we measure against the federated category baseline. Precision we measure against the rank-1 confidence of accepted matches. Round 2 figures shown are sample; Round 3 wires these to live BigQuery exports.
p50p95p99
oneways federated supplycategory baseline
rank-1 confidence
Cohort: trailing 28 days · 12 weeks · 12 weeks respectively. Excludes test traffic and synthetic load. Weekly cohort starts Monday 00:00Z.