Skip to content

Latest commit

 

History

30 Commits

Folders and files

NameName
Last commit message
Last commit date
 
 
 
 
 
 
 
 

Repository files navigation

HyperC P34 — The Self-Driving Business AI Model

P34 API

Your AI can talk. P34 lets it do business.

Website · Markets catalogue · Membership · Console · Simulator · Research · Technical report


P34 is a foundational decision model for computable markets. This API answers one question:

Given the trade options in front of you right now — which should you take, at what size, and what profit should you expect?

If your LLM is the voice of your business, P34 is its P&L. Language models reason in words; P34 is trained against realized economic outcomes (Profit-as-Regression Machine Learning, PARML). Ask it about a menu of deals and you don't get an essay — you get a portfolio: selected quantities, predicted profit, and — just as deliberately — the deals it refuses. The refusals are the product.

This repository is the complete user-facing documentation for the P34 API: how it works, how to prepare your data, runnable examples, and a pytest-based integration workflow.

Addresses

Service URL What it is
API https://api.hyperc.com/v1/ The P34 model API (POST /fit, GET /result/...)
Management console https://api.hyperc.com/app/ Account, API keys, plans/billing, session status & cancel
Market simulator https://api.hyperc.com/sim/ Interactive browser simulator — play a synthetic market with P34 predictions

GET https://api.hyperc.com/v1/ is open (liveness + capability listing); all other API calls require an API key from the management console, sent as Authorization: Bearer <key>, on an account with an active subscription (subscribe on the console's plans page — there is no free usage tier).

Is this for you?

  • ✅ You face menu-shaped decisions: inventory purchasing, wholesale lots, loan approvals, load acceptance, contract sizing — many (item, quantity) options per decision moment.
  • ✅ You have history, including the options you declined — or you're willing to start logging it.
  • ✅ You can measure an economic outcome: profit, contribution margin, recovery, yield.
  • ✅ You want an executable answer — sizes and predicted economics — not a dashboard.
  • ✅ You're wiring an AI agent (Claude, ChatGPT, open models, custom code) to real commercial decisions and need the one step it can't do alone.
  • ❌ You need sub-second decisions (fits take minutes; very high-frequency use is out of scope).
  • ❌ You want signals for securities, derivatives or prediction markets (regulated-market uses sit in a separate perimeter — see the terms).

Why not just train a regressor on your history?

Your history is biased: you only observed outcomes for the options your business actually took, and it took them selectively. A model trained naively on that history looks great on business-observed holdouts — then over-buys false positives on the full future menu it was never forced to refuse.

Naive profit regressor P34
Trains on Deals you took (the winners' club) The whole menu, including declined options
Backtest Excellent (0.9266 AUC in our benchmark) Honest
Full future menu* −$417.4k realized +$2,250 on $4,146 deployed
"Do nothing" Not in the vocabulary A first-class, rewarded output
Portfolio Per-row scores that fight for capital Jointly sized book, calibrated as a sum

* Executed slower-market-waves notebook, synthetic market with known ground truth — mechanism demonstration, not evidence of live-market profitability. Methodology and notebooks: p34-technical-report and the research page. In production the model has generated $30M+ in sales for customers with >95% of trades unsupervised (company-reported; not audited by a human licensed auditor).

The mental model

You send two tables and a config, and later receive one predicted menu:

  • Menus — every trade option you faced, historically and right now. One row per (key, quantity option). Historical menus are the model's context: P34 is pre-trained, so fitting on your history doesn't teach it markets from scratch — it calibrates the model to your market before it answers. "Every option" is literal: the deals you declined belong in the context too — the service refuses histories that are all wins.
  • Sales — your realized sales log. Used to ground the history: the service replays your inventory economics (holding costs, write-offs, fees) to reconstruct what every historical option would have earned.
  • market_type — the grounding configuration describing those economics.
  • business_description — your business and, crucially, how its unit economics is computed. Under the recommended business_led grounding mode this text is compiled into the economics used to reconstruct your history, so it is executable input rather than documentation. You can send it per request or save it once in the console.

The task is the menu you want decided now (T = 0, menu = 0). The response fills it in: per key, the selected quantity (qty = 0 = do not trade) and the predicted profit, calibrated as a portfolio sum. The task menu must carry no outcome values — P34 never sees your future.

POST /fit  ──►  validation + grounding  ──►  queued
                                             │   (calculation runs on
                                             ▼    HyperC's compute cluster)
GET /result/{session_id}  ◄──  queued → processing → done | failed

By default, grounding is compiled from your business description and runs after the response, so /fit answers grounding and the same poll loop covers the extra phase:

GET /result/{session_id}  ◄──  grounding → queued → processing → done | failed

Quickstart

Check the service is up (no key needed):

curl https://api.hyperc.com/v1/

Run the complete sample client (fit → poll → portfolio):

pip install pandas requests pyarrow
python examples/client/example_client.py --url https://api.hyperc.com/v1 --key $P34_API_KEY

Or call it directly:

import requests
r = requests.post("https://api.hyperc.com/v1/fit",
                  headers={"Authorization": "Bearer <key>"},
                  json={"menus": [...], "sales": [...], "market_type": {...},
                        # your business + its unit economics (fees, holding
                        # costs, …); or save it once in the console instead
                        # and omit this field entirely
                        "business_description": "..."})
                        # grounding is compiled from that description by
                        # default; send "grounding_mode": "internal" for the
                        # legacy fixed formula
session = r.json()["session_id"]
# poll until done (business-led fits pass through "grounding" first):
requests.get(f"https://api.hyperc.com/v1/result/{session}",
             headers={"Authorization": "Bearer <key>"}).json()

No code? The simulator runs a synthetic market in your browser against this same /v1 API — a good way to build intuition for menus, grounding and portfolio behaviour before wiring your own data.

Documentation map

  1. docs/01-overview.md — what P34 does and the mental model behind the API (menus, sales, the T=0 task).
  2. docs/02-endpoints.md — endpoint reference, auth, result statuses, model versions, confidence correction.
  3. docs/03-data-format.md — the Menus / Sales / market_type input format, rule by rule.
  4. docs/04-errors-and-checks.md — common validation errors and quick self-checks.
  5. docs/05-market-catalog.md — the computable markets catalogue: every market, its tier, its support state, its menu shape and the data it runs on.
  6. docs/06-token-wallet.md — the accumulating token wallet: monthly accruals that carry over (2,000 tokens a month, 4,000 for founding members), transfers between accounts by email, and the full query-able ledger.
  7. examples/ — runnable code:
    • examples/client/ — a complete sample client (fit → poll → portfolio).
    • examples/data/ — sample input as Excel, CSV, and JSON.
    • examples/pytest/ — a minimal pytest workflow you can drop into CI to validate your integration.
    • examples/baseline_comparison/ — a demo that pits P34 against a gradient-boosting profit regressor on a synthetic market with known ground truth.

Access & membership

API access comes with the P34 Membership$2,000/month: a 24/7 virtual machine for your agent — an always-on workspace preloaded with market-access tools, curated data sources and web scraping, so the agent can collect data and operate the business continuously rather than only while you are at the keyboard — plus the API, console and simulator, a weekly compute allowance (shown as % used), access to computable markets — the market you already operate in first, plus supported workflows where we have coverage (Amazon wholesale, US & EU) — agent skills and examples, and a community of operators. Early paid accounts lock the introductory 10% success-fee rate where profit-share pricing applies — assigned by paid-registration order and shown in your account.

Explore membership → · Join at the console → · Enterprise (governed rollout: shadow test → capped pilot → scale with controls): hyperc.com/enterprise.html

Market coverage

→ The computable markets catalogue — 69 markets, six tiers, each with a support state, menu shape and data sources. Read it before the industry cases: it is the overview, they are the deep dives. Machine-readable copy for agents at hyperc.com/markets.json.

Tier What it is Count
Core Institutional scale — wholesale, lending, cards, insurance, treasury, procurement 19
Tier 1 Cleanest telemetry, minimal handling — domain drops, vinyl, retro games, LEGO, TCG, sneakers 10
Tier 2 Strong fit, needs handling or local presence — tools, cameras, salvage, pallets, auctions 18
Tier 3 Digital and intangible — micro-acquisitions, plugins, stock assets, gift cards, points 8
Tier 4 Operational and local — vending routes, rentals, lead arbitrage, work arbitrage 8
Flagged Textbook menu structure, real legal exposure — declined by policy 6

Support states across the catalogue: 1 supported (Amazon wholesale US & EU, in production since 2023), 6 pilot-ready (validated or in enterprise discovery — micro-lending, card credit, bank onboarding, manager underwriting, online arbitrage), 53 research candidates, 9 not currently supported (regulated perimeter or declined by policy).

Synthetic markets are open to every member through the simulator and this API. Listing a market is not a claim of support — check the state before you plan around it, and run the market-fit check on your own market. Proposing a new one: hyperc.com/contact.html?topic=market.

Which market should you choose?

Start from the market you already operate in, or one you know well. Support state records where P34 has already been pointed — it is not a ranking and not a recommendation. What makes P34 work on a market is your data, your constraints and your operating knowledge, so a market you understand beats a market with a pre-built workflow.

In particular, do not default to Amazon wholesale because it is the developed one. It is the founding deployment and the best-understood market in the catalogue, and it is also one of the hardest to enter — and the difficulty is not the model. Amazon account management (ungating, brand and IP complaints, performance metrics, suspension and reinstatement) and wholesale supplier relationships (winning authorised distributor accounts at all, minimums, credit terms) are demanding operating problems, and P34 solves neither of them.

Treat the supported and pilot-ready entries as evidence that the method works, not as a shortlist to pick from.

What P34 is not

  • Not a chatbot. It doesn't converse; your agent does. P34 supplies the economic decision.
  • Not a trading-signal service. Regulated-market uses (securities, derivatives, prediction markets) are excluded from profit-share pricing and gated under the API Terms of Use.
  • Not investment advice. Output is statistical decision support; you own the decisions, the execution, the capital and the results.
  • Not an uncontrolled bot. Recommended deployment runs menu grounding → shadow test → capped pilot → scale, with caps, audit logs and kill switches.
  • Not magic. It requires policy-selection signal (you had more options than you took), tolerates minutes of latency, and refuses work its validation can't stand behind.

Built to pursue profit — not generate pretty answers.

Support


© HyperC (CriticalHop Inc). This repository contains user-facing documentation and examples only; sample data is synthetic. Benchmark results shown are from controlled synthetic markets unless labeled otherwise; production figures are company-reported. Historical results do not guarantee future outcomes.

About

HyperC API documentation

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors