Product explanation

How P34 works

P34 is a foundational decision model for computable markets. This page translates the technical thesis into plain language — without flattening it.

The loop

Evaluate → reject → size → execute

Everything P34 does is one loop, run with discipline.

01ObjectiveDeclare the economic target and hard constraints — capital, inventory, timing, compliance, risk.
02Data & contextThe business menu plus operational telemetry: what you could do, and what happened when you acted before.
03Evaluate & rankEconomic scoring with false-positive control. Most candidates are rejected — including, when warranted, all of them.
04Execute with controlsHumans or agents act on the portfolio under caps, approvals, audit logs and kill switches.
Core concepts

The vocabulary, in plain English

ConceptWhat it means
Business menuThe set of actions your business could take right now — deals to buy, loans to approve, loads to accept, offers to make — each with quantities, timing, counterparties, prices and constraints.
ContextWhat the market and your business look like around that menu: history, inventory, capital, seasonality, fees.
Decision modelP34 evaluates the menu as an economic decision problem, not only as a prediction task. The output is a decision, with predicted economics attached.
No-tradeDoing nothing is a valid output. When no candidate meets the model's criteria, P34 declines — a first-class, rewarded decision, not a failure mode.
Portfolio outputP34 selects actions and sizes them as one coordinated portfolio or workflow, so positions don't fight each other for capital or capacity.
Partial observabilityThe true market state is never fully visible. Operational telemetry — your realized outcomes — carries the signal that missing price discovery doesn't.
Controlled autonomyThe system can move from recommendation to execution — but only within user-defined limits and rollout stages.
Positioning

AI about money, not about words

What LLMs do

Language models reason, summarize, and advise. Ask one about a deal and you get an articulate essay — trained on text, scored on plausibility, accountable to nothing. Seven years of prompts won't change what the objective function rewards.

What P34 does

P34 is trained against realized economic outcomes — profit as the regression target (PARML). Ask it about a deal and you get a decision: take it at this size, or don't take it at all, with predicted economics it can be held to.

The two compose: your LLM agent talks to customers, reads contracts and orchestrates workflows — and calls P34 whenever a decision involves money. Seven years, not seven prompts.

The hard part

Why rejection is the product

Business history is biased: you only observe outcomes for the deals a previous policy already accepted. Models trained naively on that history become systematically optimistic — they inherit yesterday's enthusiasm and none of its regret. P34's model class treats declining to act as a rewarded output and controls false positives explicitly. In live operation, most of what P34 sees, it refuses.

Universe

Every candidate action is embedded with its context — including candidates historical policies never touched.

False-positive control

The model is penalized for confident acceptance of deals that would have lost money — not just rewarded for winners.

Portfolio / no-trade

Survivors are sized jointly under constraints. If nothing survives, the answer is nothing — and that answer is respected.

Autonomy

The same architecture, at your scale

Four stages from decision support to self-driving business. Toggle the view — the loop doesn't change; the stakes and the sign-off do.

Available now

1 · Decision Support

You: read ranked menus, decide, act.
P34: scores and explains.
Controls: none needed — nothing executes.

Available now

2 · AI Copilot

You: approve each prepared action.
Your agent: assembles orders via the API.
Controls: per-action approval, budget cap.

Rolling out

3 · Supervised Autonomy

You: review daily digests.
P34 + agent: execute approved action classes.
Controls: caps, drawdown limits, instant pause.

Planned

4 · Self-Driving Business

You: set the envelope, watch the P&L.
System: full loop within limits.
Controls: audit logs, drift monitoring, kill switch.

Available now

1 · Decision Support

Team: analysts consume ranked menus in existing tools.
P34: shadow-scores live menus.
Controls: read-only integration, reconciliation.

Available now

2 · AI Copilot

Team: operators approve queued actions.
P34: proposes sized portfolios per unit.
Controls: review thresholds, four-eyes sign-off.

Pilot stage

3 · Supervised Autonomy

Team: exception handling only.
P34: executes capped, reviewed portfolios.
Controls: portfolio caps, data-quality gates, audit logs.

Planned

4 · Self-Driving Business

Team: governance and strategy.
System: end-to-end within the governed envelope.
Controls: drift monitoring, allowed-market lists, kill switches.

Stages labeled “Rolling out” / “Planned” are roadmap: planned, not guaranteed, and gated by market, controls readiness and legal review.