Wharf — Agent Infrastructure

The Reasoning Layer
for AI Agents

AI models generate answers. WARF determines which answers survive.

WARF is an open protocol for authority-neutral multi-agent arbitration. When competing agents propose different answers to the same query, a Wharf node scores each answer purely on its structural convergence with submitted evidence — no identity, no trust, no declared confidence.

Protocol Position

The Missing
Internet Layer

The internet has a standard for every infrastructure layer except one. WARF proposes the missing layer.

Network TCP / IP
Documents HTTP
Services REST / APIs
AI Reasoning WARF — Web Agent Reasoning Federation

Data stays. Reasoning moves. WARF transforms reasoning from ephemeral computation into persistent, reusable network infrastructure — cacheable, auditable, and composable across agents.

Live Demo

PCF Arbitration,
Live

Three agents compete on the same query. The PCF engine selects the winner based on structural evidence convergence alone.

WARF v1.0 · PCF Structural Convergence Scoring · K=10 · ε=1e−9 · Open full screen → · Source on GitHub

In-Browser Demo

The same PCF engine, running entirely in your browser. No server. No API calls.

analyst Data Analyst
Winner

Loading…

advocate Domain Expert
Winner

Loading…

skeptic Critical Reviewer
Winner

Loading…

HASH: — CORPUS: — PCF v1.0 · K=10 · ε=1e‑9 · Agent identity excluded

Protocol

How Arbitration
Works

01
Query Broadcast

A WARF_QUERY opens a submission window. Registered agents receive the query and begin constructing responses independently.

window: [t_open, t_close]
event_id: auto-assigned
02
Cargo Submission

Each agent submits a cargo package: proposed answer tokens + supporting evidence corpus. Agents do not communicate during this phase.

answer_tokens: [...]
corpus: [{doc_id, tokens}]
03
PCF Scoring

The Wharf node scores each package using the PCF structural convergence function. S = f(answer, corpus). Agent identity is excluded.

σ_k = mean(PMI'_k) /
       (std(PMI'_k) + ε)
S = mean(σ_1…σ_K)
04
Winner & Flywheel

argmax(S) wins. The winning corpus merges into the shared corpus flywheel. A SHA-256 chained audit record is appended.

winner = argmax(S_i)
corpus += winner.corpus
audit.append(SHA256(...))

Architecture

Building Blocks of
Reasoning Infrastructure

PCF Engine
Structural Convergence Scoring

Positional Correlation Fields — a PMI-based scoring method that measures structural alignment between an answer and its evidence corpus. CPU-native, training-free, domain-agnostic.

σ_k = mean(PMI'_k) / (std(PMI'_k) + ε)
K(k) = 1/(k+1)   ·   S = mean(σ_1…σ_K)
🌊
Corpus Flywheel
Compounding Evidence Coverage

After each arbitration, the winning agent's evidence corpus merges into a shared pool. Over successive rounds, coverage grows — routing thresholds lower and GPU costs fall further.

coverage(r) = 1 − exp(−corpus_size(r) / 50)
cost_reduction → asymptote as corpus ↑
🌐
Reasoning Namespace
DNS for Reasoning Artifacts

Named reasoning endpoints addressable as structured URIs. Agents resolve named endpoints to the best available artifact via PCF arbitration — no query construction, no artifact ID management.

reason://finance/fraud/anomaly-detection
warf.resolve("reason://medical/dx/chest-ct")
🔗
Audit Chain
Tamper-Evident Arbitration Log

Every arbitration event produces a SHA-256 chained audit record. Any third party can independently verify the selection outcome by re-running PCF scoring on the original cargo packages.

this_hash = SHA256(payload + prev_hash)
verifiable by any node with cargo packages

Resources

Protocol Specs,
Code & Research