STP — Semantic Transfer Protocol

01 — THE PROBLEM

Current workarounds
are inadequate.

The modern web was designed for human eyes. When an AI agent needs to read a webpage, it reverses engineers meaning from a presentation layer never intended for machines. The waste is structural.

HTML PARSING

Brittle by design

Breaks on redesigns. Can't infer semantic relationships — only structure. One CSS change breaks the agent.

LLM COMPREHENSION

Token-heavy

A 4,000-word article becomes ~40 semantic facts — after burning the full token budget to get there.

JSON-LD / SCHEMA.ORG

Not relational

Describes what a page is for search snippets. Not what it means relationally. No confidence. No provenance.

RAG PIPELINES

Structural waste

Every agent independently rebuilds the same semantic graph from the same source. The waste compounds at scale.

"The web was built for humans. Search engines retrofitted machine readability on top. STP asks a different question: what if we designed the data layer for agents first, and let humans have a translation on request?"

02 — HOW IT WORKS

Same URL. Two audiences.
No new infrastructure.

STP embeds a structured semantic block inside any webpage via a script tag. Browsers skip it. Agents parse it and skip the DOM entirely. Where HTML communicates presentation and JSON communicates data, STP communicates meaning.

// Any webpage. Browsers ignore this. Agents read only this. <script type="application/stp+json"> { "stp": "0.1", "concepts": [ { "id": "stp:ai.ml.006", "ref": "large_language_model", "weight": 1.0 }, { "id": "stp:ai.ml.009", "ref": "training_data", "weight": 0.8 } ], "relations": [ { "from": "stp:ai.ml.006", "to": "stp:ai.ml.009", "type": "requires", "confidence": 0.85, "provenance": "https://arxiv.org/abs/2005.14165" } ] } </script>

8 relation types. Canonical concept IDs. Confidence scores with provenance chains. Designed to degrade gracefully — pages without STP fall back to HTML parsing. Pages with STP are just faster, cheaper, and structurally richer to consume.

03 — ARCHITECTURE

Three layers.
Independently adoptable.

Each layer is safe to ship without the next. The reading layer is inherently safe — nothing executes. The action layer requires a complete security specification first. The A2A protocol can only follow.

L1
Reading Layer
Agents read structured semantic data from webpages. Concept Registry (23 concepts, 6 domains), 8 typed relation types, confidence propagation with hop decay and cross-domain penalties, deterministic conflict resolution engine. Inherently safe.
IN DEVELOPMENT
L2
Action Layer
Agents execute structured operations via action manifests declared in STP blocks. Direct API calls — no browser, no DOM, no selector breaks. 5-step security pipeline: signature verify → injection scan → domain allowlist → scope check → human gate.
SECURITY SPEC COMPLETE
L3
Agent-to-Agent Protocol
Agents communicate directly via typed STP packets — zero natural language. 8 message types: QUERY, ASSERT, CHALLENGE, RESOLVE, DELEGATE, ACK, REJECT, COMPLETE. 2.7× compression vs natural language coordination.
PROTOTYPE

04 — THE NUMBERS

Measured honestly.

The RAG comparison is the most practically important — it's what agents actually use today for web reading. The 39× number is robust even with minimal STP blocks because typed structure is categorically richer than prose fragments.

COMPARISON	SAVINGS	NOTE
STP vs raw HTML	161×	Real but not the fair comparison
STP vs stripped text	48×	Assumes well-authored STP block
STP vs RAG (5×512 chunks)	39× — and more structured	Use this number.

MODEL	CONVENTIONAL (PAGES / CTX)	STP (PAGES / CTX)
Claude Sonnet 200K	53 pages	2,545 pages
Gemini 1.5 Pro 1M	264 pages	12,723 pages

BENCHMARK	CONVENTIONAL	STP	IMPROVEMENT
Task completion time	9.27s	0.42s	21.8× faster
Bytes processed	102KB HTML	892 bytes	116× less
LLM calls required	1	0	Eliminated
Selector breaks	1 (CSS → XPath retry)	0	None possible
Crawler compression	288KB HTML	2,620 bytes STP	112.5×

Honest caveat: 48× assumes a well-authored STP block. Real-world numbers in early adoption are probably 20–35× vs extracted text, climbing toward 48× as tooling matures. $9/day saved at 1,000 pages/day at current Claude Sonnet pricing. $928/day at crawler scale (100K pages).

05 — THE TEMPORAL LAYER

Static knowledge is easy.
STP tracks how it changes.

An agent reading an ML paper from 2023 needs to know that the field's confidence in that claim has since dropped. A static knowledge graph gives it the claim. STP's temporal layer gives it the claim, its current standing, and the event that caused the revision.

Jan 2022

Pre-ChatGPT Era

Transformer 97 · Attention 96 · LLM 71 · Agent 38 · Tool Use 21. Architecture is understood. Applications unclear.

Dec 2022

ChatGPT Changes Everything

LLM confidence spikes. Everything connected follows. The graph restructures around a new center of gravity. Agent begins its long climb.

Jul 2023

Emergent Behavior Gets Challenged

Papers question whether emergence is real or an evaluation artifact. Emergent Behavior confidence drops −12. The LLM causes EmergentBehavior relation weakens and changes type to relates_to.

Apr 2024

Reasoning Models Emerge

The field starts treating reasoning as a separable capability, not just an emergent property of scale. Reasoning gets its own node and begins climbing toward 99.

Oct 2024

Agentic Frameworks Mature

Tool Use climbs from 21 → 96. Agent follows. Edges between Agent, Tool Use, and Reasoning tighten. The graph reorganizes for the second time.

Mar 2026

STP Era Begins

Reasoning 99 · Agent 97 · Tool Use 96 · Emergent Behavior 41. The field moved on. The graph recorded it.

06 — PROTOTYPES

12 working prototypes.
All open source.

Every layer of STP is interactive and runnable. Not slides. Not mockups. Working code that demonstrates what the protocol actually does.

01

Conflict Resolution Engine

5-criteria deterministic pipeline for contradictory semantic claims. Confidence delta → domain authority → recency → source type → corroboration. UNRESOLVED when all five tie.

4 LIVE TEST CASES

02

Security Specification

8 identified threats (3 CRITICAL, avg CVSS 8.5). 9 mitigations. Phase-gated roadmap. Action layer cannot be built until Phase 0 mitigations are implemented.

CVSS MAX 9.8 · 7 UNMITIGATED

03

Action Layer

Live execution pipeline. Select an action, fill parameters, watch all 5 security checks run in sequence. Payment actions halt for human confirmation. No DOM. No browser.

5-STEP SECURITY PIPELINE

04

Agent-to-Agent Protocol

Zero natural language. Two agents negotiate via typed STP packets. QUERY → ASSERT → CHALLENGE → RESOLVE → COMPLETE. Watch the exchange happen in real time.

2.7× COMPRESSION VS NL

05

Unified End-to-End Demo

One agent. One task. Every layer firing in sequence. Registry → Confidence → Conflict → Security → Action → Human Gate → A2A → Complete. Human gate holds at payment until approved.

8 STAGES · 3.6× COMPRESSION

06

Crawler Simulator

8 pages. One crawler. Zero HTML parsed. Watch the knowledge graph assemble in real time as each page's STP block is read. 13 nodes. 22 edges. 2,620 bytes vs 288KB.

112.5× COMPRESSION

07

Benchmark

Side-by-side race. Conventional agent (browser + DOM + LLM call) vs STP agent (semantic block + direct API). Same task. The conventional agent hit a CSS selector break mid-run.

21.8× FASTER · 0 SELECTOR BREAKS

08

STP Block Generator

Paste any URL or article text. Claude extracts concepts, infers typed relations, assigns confidence scores calibrated to source type, outputs a deploy-ready STP block.

30 SECONDS · BRING YOUR API KEY

09

Diff Engine

Git for semantic graphs. Paste two versions of an STP block. Get a structured diff: concepts added/removed/modified, relations changed, confidence drift, agent cache invalidation recommendation.

3 EXAMPLE SCENARIOS

10

Validator

Schema errors, injection scan, signature check, registry compliance, confidence range warnings, relation consistency. Everything a developer needs before deploying an STP block to production.

6 VALIDATION CHECKS

11

Temporal Graph

The AI/ML knowledge graph animated month by month, 2022–2026. Watch concepts gain and lose confidence as papers publish. Watch relations appear, strengthen, weaken, and get overturned. Nobody built this before.

48 MONTHS · 16 CONCEPTS · PLAY →

12

Concept Registry

23 canonical concepts across 6 domains. Canonical IDs, aliases, domain weights, pre-defined relations. The shared vocabulary that makes STP blocks interoperable across sources.

v1.0 · 6 DOMAINS

VIEW ALL ON GITHUB →

07 — DEVELOPMENT LOG