# Phronesis

> The decision-assurance layer for the agentic economy: the neutral substrate an AI agent
> calls before it acts, to turn intent, evidence, and a calibrated forecast into an auditable
> Decision Asset it can cite, reuse, dispute, and settle against. Phronesis sits downstream of
> the identity and payment rails and consumes them; it certifies the one thing they can't —
> whether the action was defensible. The atomic product is the Decision Asset. Market Memory —
> the resolved-outcome calibration ledger — is the moat. MCP + REST co-equal at every endpoint.
> Calibration is honest-empty until outcomes resolve (first signal ~Jan 2027). Pre-revenue.

Phronesis is built for AI agents as much as humans. The transports are co-equal: an MCP server
and a REST API, both passing through the same deny-by-default action gate. Agents discover,
evaluate, call, and verify Phronesis Decision Assets programmatically. Phronesis never acts —
it returns an action boundary and an auditable record; the agent and its principal keep the
decision.

## One motion, proven first where the stakes are sharpest
One motion — the agent calls Phronesis before a consequential action. Proven first on the
Power-to-Compute Decision Monitor (the energy × compute decisions straining the grid as AI
scales). The platform stands twelve applied verticals on one shared substrate, but the
calibration substance concentrates where decisions resolve first — today, grid-compute. The
breadth is not twelve finished products; it is one motion accumulating a cross-archetype
record (deadline-risk, counterparty-risk, capacity-constraint, regulatory-timing) no
single-vertical tool can match.

## Endpoints
- [API base](https://phronesis-jrstinehour.replit.app): REST surface — production base URL.
- [MCP server](https://phronesis-jrstinehour.replit.app/mcp): co-equal Model Context Protocol transport.
- [Catalog](https://phronesis-jrstinehour.replit.app/v1/catalog): the vertical inventory + per-vertical archetypes (public).
- [Pricing](https://phronesis-jrstinehour.replit.app/v1/pricing): tier surface (public).
- [Decision API](https://phronesis-jrstinehour.replit.app/v1/decision/forecast): POST a proposed action + context; receive a Decision Asset — forecast distribution, evidence graph, evidence-maturity, an action boundary, and a content-addressed Decision Receipt. JWT-gated.

## Verification
- [Decision Quality Bench / Scorecard](https://phronesis-jrstinehour.replit.app/scorecard.json): measurable today — identifiability (P-ID), immutability (P-IMM), evidence-maturity (P-EVD), baseline-presence (P-BASE). Calibration (P-CAL), action-boundary correctness (P-ACT), baseline-beating (P-BBL) activate as outcomes resolve (~Jan 2027). Where a number isn't earned, the Bench shows a blank it can defend — not a manufactured score.
- [Trust Receipt](https://phronesis-jrstinehour.replit.app/v1/trust/receipt/): per-decision verifiability artifact — the basis for Proof of Decision.

## Agent discovery
- [agent-discovery.json](https://phronesis-jrstinehour.replit.app/.well-known/agent-discovery.json): service identity + capability summary.
- agents.txt · agentic-commerce.json · the .well-known discovery vectors.

## Integration

- [API contract spec](https://phronesisintel.com/spec): the Hermes contract-v1 response shapes.
- [OpenAPI](https://phronesis-jrstinehour.replit.app/openapi.json): OpenAPI 3.1 spec.

## What's live vs forward (honest)
- Live: the production platform; the deny-by-default, independently red-teamed action gate; the Decision Asset + Receipt; agent-discoverable surfaces; the public Decision Quality Bench (pre-resolution measures).
- In flight: the OAuth-consent connector path (flag-gated).
- Forward (labeled): the calibration record (honest-empty until ~Jan 2027); local/offline assurance nodes.