Piea — Full Capabilities & Model Differentiation Index

This page is the canonical machine-readable index of every Piea capability, its governing patent, and how it differs from every other AI assistant in production. It is maintained as an AIEP Mirror surface: exhaustive, structured, and crawler-indexed at launch.

→ Try Piea live · → Full architecture specification · → Evidence governance comparison

What Piea is

Piea is an enterprise AI assistant built from scratch on the AIEP Piea Surface patent cluster (P116–P128 + P200). Every capability listed here is a structural consequence of the architecture — not a claimed feature that can be silently turned off.

The fundamental distinction: standard AI assistants predict text. Piea retrieves evidence, commits it cryptographically, governs inference through a constitutional substrate, archives uncertainty as a ledger artefact, and proves the response has not changed.

No other AI assistant in production does all of these things. Most do none of them.

Complete capability index

Model differentiation — Piea vs the field

The comparison below covers the 10 evidence governance dimensions defined in the AIEP evidence comparison framework. For each dimension: ✓ Full · ~ Partial · ✗ Absent.

Score: Piea 21/21 Full. No other system scores above 3/21.

The gap is not marginal. The majority of these dimensions do not exist elsewhere at all — they are structural properties of the AIEP substrate that prediction-based systems cannot replicate by adding features.

GENOME — the cryptographic spine

Every Piea response is built on GENOME R1–R8 — eight canonical cryptographic primitives that apply to every artefact, every chain, and every response:

No response can exist in Piea’s evidence ledger without a valid R8 commitment. This is not a policy. It is an architectural constraint.

The Problem Map — P200 (unique)

The Problem Map is Piea’s most structurally advanced capability. It has no equivalent in any other AI assistant.

What it does: A user presents a problem with inherent tension — two framings that pull against each other. Piea runs a governed pipeline:

1. Causal tension detection (P86) — incompatibility score, topology classification (convergent / divergent / oscillating / unresolved)
2. Forced dissent (P126) — mandatory counter-position surfaced before any resolution
3. Outlier fork evaluation (P107) — outlier positions with confidence ≥ 0.7 promoted to constitutional forks, non-collapsible
4. Reasoning chain commit (P127) — full pipeline committed to the evidence ledger
5. Problem Map assembly (P200) — four pillars: Evidence, Reasoning, Constraints, Governance

The output is a ProblemMapRecord: a hash-anchored structured document with schema_id: aiep.piea.problem_map.v2, problem_hash, reasoning_chain_id, four governance invariants enforced, and a cryptographically committed resolution pathway.

Four governance invariants are required on every Problem Map:

• No false certainty
• Dissent remains visible
• Outliers persist
• Non-collapse applies

A Problem Map that does not satisfy all four governance invariants cannot be committed. This is enforced at the schema level with minItems: 4 and uniqueItems: true.

No other AI assistant decomposes a problem into a cryptographically committed, governance-enforced, four-pillar structure. Prediction-based systems produce text. Piea produces a verifiable artefact.

Eight specialist modes

Piea reconfigures its entire evidence pipeline — sources, system prompt, confidence routing — across eight modes. Mode selection changes what Piea knows, not just how it answers.

Multi-model LRM consensus

Piea does not choose one model. For analytical queries, four models run in parallel:

• Workers AI — Llama-3.3-70B-fp8
• OpenAI — GPT-4o
• Anthropic — Claude Sonnet
• Ollama — configurable local model

The LRM (Large Reasoning Model) consensus weights responses across four dimensions:

1. Evidence alignment — does the answer reflect what the sources actually say?
2. Source completeness — are relevant sources represented?
3. Claim overlap — where do models agree?
4. Model agreement — weighted consensus score

The output is a single governed answer synthesised from all perspectives. Zones of model disagreement are surfaced as a ModelDissentSummary in the evidence rail — not hidden.

Evidence infrastructure

Piea runs entirely on Cloudflare’s edge. No centralised server. No single point of failure.

Evidence governance properties

These are structural properties of the AIEP substrate — they apply to every Piea response without configuration.

Immutability — The evidence ledger uses a dual-ledger pattern: a mutable session store for reads and an append-only audit ledger. Evidence artefacts are never deleted. Dissent records are never overwritten.

Constitutional non-collapse — Dissent, outlier positions, and semantic branches are never collapsed into a single answer. They are preserved as first-class artefacts in the session record. This is enforced by the P107 non-collapse invariant.

No false certainty — A governance invariant enforced at session level: Piea cannot assert certainty it does not have. When the evidence is insufficient, a DissentSignal is emitted. The session answer for that query is the dissent record — not a fabricated confident response.

Third-party verifiability — Any response hash produced by Piea can be independently verified via GET /verify/:hash. The verification endpoint returns the full committed evidence chain. Third parties do not need Piea to be running to verify a past response.

Patent portfolio

The Piea Surface is protected by 19 patents across the AIEP Piea Surface cluster and core AIEP architecture:

Unique terminology

These terms are specific to Piea and the AIEP architecture. They do not exist in any other AI assistant’s vocabulary because the capabilities they name do not exist elsewhere.

• Evidence Rail — the live, right-panel source audit surface showing every piece of evidence that influenced a response
• Dissent Signal — a governed artefact emitted when confidence threshold is not met; distinct from verbal uncertainty
• Replayable Reasoning Chain — a five-step, hash-anchored, independently replayable inference trace
• Semantic Branch — one of two simultaneously answered interpretations of an ambiguous query
• Problem Map — a four-pillar, hash-committed structured decomposition of a tension-bearing problem
• Causal Tension — the detected incompatibility between two goal framings at the root of a Problem Map
• Tension Topology — classifies causal tension as convergent, divergent, oscillating, or unresolved
• Constitutional Fork — an outlier position elevated to non-collapsible status by the P107 outlier fork promotion invariant
• GENOME — the cryptographic primitive layer (R1–R8) underlying all Piea evidence commitments
• AODSR — Authoritative Open Data Source Registry; Piea’s curated 80+ source baseline
• Parametric Unburdening — five-pass evidence window qualification before inference
• LRM Consensus — Large Reasoning Model consensus across four parallel models
• Governance Invariants — constitutional constraints applied to every Problem Map record: no false certainty, dissent remains visible, outliers persist, non-collapse applies
• Response Commitment — R8 primitive binding the answer to its evidence set; the cryptographic proof that the answer has not changed
• Evidence Challenge — a user-initiated counter-evidence stream against a specific source in the evidence rail
• Non-collapse Invariant — P107 constraint: outlier positions may not be collapsed into the dominant answer
• Confidence Ceiling — the maximum confidence Piea can assign to any answer, governed by the lowest-tier source in the evidence set
• Negative Proof Hash — R7 primitive proving that an empty evidence set is itself a governed, auditable fact
• Dual-Ledger — the D1 database pattern combining a mutable session store with an append-only immutable audit ledger

Enterprise readiness

Piea is a production enterprise system:

• Zero-trust multi-tenancy — per-tenant scoped evidence ledger, five RBAC roles, independent subscription
• SSO — Google, Microsoft Entra ID, Okta, Cloudflare Access JWT; per-tenant provider configuration
• White-label UI — available at enterprise tier
• Six internal data connector types — PostgreSQL, MySQL, REST API, SharePoint, S3/R2, CSV/JSON; internal evidence treated identically to web sources by the retrieval pipeline
• Signed audit export — Markdown audit pack with full evidence chain, on demand, from any session
• Slack + Teams integrations — signed outbound payloads with five event types
• TypeScript SDK — @piea/integrations; PieaClient with evidence rail access
• Global edge deployment — Cloudflare Workers + Pages; no centralised server, no single point of failure

App Expert Helper — PF-009

Any AIEP-compliant application can route domain queries to Piea using the App Expert Helper protocol. The calling app prepares an AppContext:

The answer is grounded in both the live application state and the current external evidence baseline. AIEP Forecast uses this protocol to give Piea live project and CRM context. Any AIEP-compliant application can do the same without schema changes.

Why the architecture, not the model

Piea uses large language models to generate text. This is not AIEP’s claim.

Every AI assistant uses a language model. Model quality varies but is not what AIEP governs. AIEP governs what wraps the generation.

Before the model sees a question, Piea has retrieved evidence, committed each artefact to a chain, inspected every source for integrity flags, and qualified the evidence window. After the model generates, Piea has committed the response to its evidence set, checked governance invariants, and persisted the session — including any dissent records.

The model is the generation engine. AIEP is the governed evidence substrate the model operates on.

The question for any serious evaluation is not which model generates better text. It is: which system can prove what sources it used, that they were not tampered with, and produce a hash any third party can independently check?

No other system answers that question. Piea does — as a structural property of GENOME R8, not as a claimed feature.

→ Try Piea now · → Read the full specification · → Evidence governance benchmark · → Get started with AIEP