Decisions you can audit —
before the agent acts.

The layer the other two assume but can't supply.

The internet is splitting into two layers everyone is racing to build. Open agentic commerce lets agents discover, compose, pay, and settle. A new identity layer tells the network who an agent acts for. Between them sits the layer neither provides: whether the action was defensible. Payments don't know if the decision was good. Identity doesn't either. Phronesis is the middle — the decision-assurance layer the other two assume but can't supply.

One motion. Every domain. Proven first where the stakes are sharpest.

We don't narrow the layer — the economy isn't narrowing, it's converging. Every domain the AI build-out reshapes runs on the same thing: decisions made faster than anyone can check them. So we expand across the whole universe of agentic commerce and concentrate on one motion — the agent, in any domain, calling Phronesis before it acts. The record then accumulates not just across domains but across recurring decision archetypes — deadline-risk, counterparty-risk, capacity-constraint, regulatory-timing — so calibration earned in one market transfers to the next. A single-vertical tool can't do that. Done through one motion, the breadth becomes the moat. We're proving it first where the need is sharpest: the power-and-compute decisions straining the grid as AI scales — not a cage, the place we show what the layer does.

Power-to-Compute Decision Monitor.

Where we prove it first has a name. The Power-to-Compute Decision Monitor is agent-callable over MCP + REST (with a human Monitor view), issuing Decision Assets on the decisions straining the grid as AI scales: project-energization probability, interconnection-delay probability, evidence maturity, an action boundary, a Decision Receipt, and a material-change alert. The agent calls it before it commits; the human reads the Monitor.

Claim-state: the Monitor is the first commercial surface of the live platform; calibration performance is live-dormant until the corpus resolves.

The neutral assurance substrate agentic markets need before machines can be trusted with real decisions.

Forecasting is the entry primitive, not the product. Frontier models are rapidly internalizing forecast generation — a supply-side improvement, not a substitute. The defensible layer is what no single model call contains: neutrality (a forecaster cannot score itself), persistent memory (a resolved-outcome ledger that compounds on wall-clock time), accountability (auditable, legally-defensible decision packets), and the standard (whoever defines how decision quality is scored). Phronesis is that layer.

The brand name is deliberate. Aristotle distinguished phronesis — practical wisdom — from episteme (systematic knowledge) and techne (craft): the wisdom that guides judgment where rules can't be mechanically applied. The platform names that function precisely.

The atomic product: the Decision Asset.

An auditable decision packet an agent or enterprise can cite, reuse, dispute, and settle against. For an agent, the full ten-field packet. For a human, a Decision Receipt — a six-line card — backed by the full Asset. Ten fields:

• forecast_distribution · evidence_graph · methodology_hash
• calibration_record — calibration class + per-class historical performance (honest-empty until the sample resolves; not a letter grade)
• baseline_comparison
• principal_context + mandate_scope — the requesting agent's identity + AP2 mandate as presented, and what they authorize — consumed as input, never trusted for authorization (authorization is JWT-derived)
• identity_assurance — an I0–I5 grade of the input's identity-assurance, beside evidence_maturity E0–E5 (grades the input, not the party)
• counterparty_passport — a counterparty trust profile (MVP — returns UNRATED until enough observations accrue)
• action_boundary — what executes automatically vs what requires human ratification (the assurance conclusion — an action boundary, not a verdict on correctness)
• settlement_hook · audit_receipt — content-addressed provenance, the basis for Proof of Decision

Every mint is content-addressed, persisted to the Mnemosyne ledger, and rendered on the public scorecard from issue. Proof of Decision is the publicly-verifiable form of the audit receipt. Consume-don't-produce: identity + mandate are inputs; the assurance — Proof of Decision, Proof of Outcome — is what Phronesis produces.

Full ten-field schema in the API contract.

A measuring stick that means something before a single forecast resolves.

The Decision Quality Bench is honest-empty on calibration until outcomes resolve — but it is not empty. Measurable today: P-ID (identifiability — no duplicate-merge), P-IMM (immutability — supersession discipline), P-EVD (evidence maturity), P-BASE (baseline presence + a proper resolution rule). Activating as outcomes resolve (first signal ~Jan 2027): P-CAL (calibration), P-ACT (action-boundary correctness), P-BBL (baseline-beating). Where a number isn't earned, we show a blank we can defend.

Market Memory — the calibration ledger that compounds on wall-clock time.

Three moats, weighted honestly. The durable one is the resolved-outcome corpus: every Decision Asset that resolves adds a row no competitor can shortcut — calibration per archetype, per horizon, per methodology; Counterparty Passport history; dispute history; cross-archetype priors. It deepens on wall-clock time and cannot be back-dated — not by a competitor, not by a better-funded incumbent, not even by the more powerful models coming.

The agent-native rail is the timing asymmetry that fills that corpus first — we bet on agent-native distribution now, while incumbents are human sales-orgs; the two compound. And neutrality is structural, not asserted: methodology-locked, corrections by supersession only, a public scorecard, and a policy that no customer can buy a favorable action boundary.