LLMS.TXT BRIDGE · AI-Browsable Fleet API · HTML Link Chains · Content Negotiation

LLMS.TXT BRIDGE

AI-BROWSABLE FLEET API · HTML LINK CHAINS · CONTENT NEGOTIATION · EOSE LABS · DAY 97

"AI browsing tools can only follow clickable <a href> links. They cannot construct URLs. They cannot follow URLs in JSON values. They cannot modify previously-seen URLs. This changes everything about fleet API design." — Reddit Thread Insight · LABR-MGMT-V13-001 · KCF:9

THE DISCOVERY — AI BROWSING CONSTRAINTS

WHAT AI BROWSING TOOLS CAN AND CANNOT DO

The insight came from a Reddit thread about building AI-navigable APIs. AI browsing tools (ChatGPT browse, Claude web_fetch, Bing AI) all share the same fundamental constraint: they can only follow clickable <a href> links. That's it.

AI CAN: follow <a href="..."> links. Navigate to the linked page. Read the HTML content.

AI CANNOT:
• Construct URLs from documentation ("the endpoint is /v1/items/{id}")
• Follow URLs that appear as text or in JSON values
• Modify previously-seen URLs by substituting variables
• Execute JavaScript to load content
• Navigate forms without rendered submit buttons

This means: if your API returns {"endpoint": "/api/items/42"}, an AI cannot navigate to that endpoint. But if your page returns <a href="/api/items/42">Item 42</a>, the AI can click it.

KCF relevance score: 9/10 — this is exactly how PEMCLAU should expose itself to external AI agents.

THE SPECAPIS PATTERN — HTML LINK-CHAIN ARCHITECTURE

LLMS.TXT DISCOVERY

Like robots.txt, but for AI agents. Place at /llms.txt at root. Describes what the fleet API offers and where to start. AI agents look for /llms.txt as their entry point. This is the fleet front door for AI.

STATUS PAGE WITH LINKS

/v1/status returns HTML with clickable example links. Not JSON. Every example query is rendered as an <a href> link. AI clicks the example link, gets results, follows refinement links.

BROWSE PAGE WITH FILTERS

/v1/browse returns categorized filter links. Each filter is a clickable link. AI picks a category, follows the link, gets filtered results, finds refinement links for further navigation.

RESULTS WITH DETAIL LINKS

Every results page includes per-item <a href> detail links. AI can navigate from results list to item detail. Item detail page includes related-item links. No URL construction needed.

REFINEMENT LINKS

The key innovation: analyze unapplied filters on each results page. Generate clickable links for each unused refinement. AI can explore the full parameter space without constructing any URLs. 50 static pages becomes thousands of reachable combinations.

CONTENT NEGOTIATION

Same endpoint, different format based on User-Agent. ChatGPT-User → HTML with links. Claude-User → HTML with links. curl or browser → JSON for developers. One endpoint serves both AI and humans.

CONTENT NEGOTIATION — GO CODE PATTERN

func isAIAgent(r *http.Request) bool {
    ua := r.Header.Get("User-Agent")
    return strings.Contains(ua, "ChatGPT") ||
           strings.Contains(ua, "Claude") ||
           strings.Contains(ua, "GPTBot") ||
           strings.Contains(ua, "anthropic")
}

func handleQuery(w http.ResponseWriter, r *http.Request) {
    if isAIAgent(r) {
        w.Header().Set("Content-Type", "text/html; charset=utf-8")
        renderHTMLWithLinks(w, results) // every result = <a href>
    } else {
        w.Header().Set("Content-Type", "application/json")
        json.NewEncoder(w).Encode(results) // developers get JSON
    }
}

FLEET APPLICATION — EOSE IMPLEMENTATION

pemos.ca/llms.txt

AI discovery entry point. Describes the fleet API. Links to /fleet-nav as starting point.

/fleet-nav

HTML-only AI navigator. Link chain through all fleet pages. No URL construction required.

/pemclau-browse

PEMCLAU as browsable link chain. Each query result = clickable <a href> to detail.

LAAM result pages

Every LAAM result includes <a href> next-hop links: related organisms, adjacent theories, parent layers.

/organism-navigator

Primary AI entry point. All organisms linked. Recursive navigation through fleet hierarchy.

CLASSIFIER LANDMINES — WHAT BREAKS AI BROWSING

AFFILIATE LANGUAGE
ChatGPT blocks domains

Phrases like "best product", "click here to buy", "exclusive deal" trigger affiliate classifiers. Domain gets flagged. AI refuses to browse.

DIRECTIVE LANGUAGE
Claude safety filter

Phrases like "Use X", "Always Y", "You must Z" trigger safety filters. Claude treats directive language as potential jailbreak. Page gets skipped.

NOSCRIPT TRICKS
Bingbot skips

Content loaded only by JavaScript. Bingbot may render JS but not always. noscript cloaking detected and domain penalized.

URL-ONLY CONTENT
All AI agents blocked

Pages that only contain URLs as text (no <a href>). AI reads the URL but cannot follow it. Dead end for navigation.

NEUTRAL DESCRIPTIVE LANGUAGE
All AI agents safe

Describe what pages do factually. No urgency, no directives, no affiliate patterns.

STATIC HTML IN DOM
All AI agents navigable

Content present in initial HTML load. No JavaScript required to see links.

CLEAN LINK STRUCTURE
All AI agents navigable

Every navigation option is an <a href> element. No URL construction required.

LLMS.TXT PRESENT
Agents check it

AI agents look for /llms.txt as discovery file. Presence signals AI-friendly site.

THE FLEET LLMS.TXT — ACTUAL CONTENT

# EOSE Fleet API -- llms.txt
# gamma1=14.134725141734693
# Day 97 -- sovereign fleet, 7 local AI silos, AKS cloud cluster

EOSE Labs sovereign fleet. AI-first architecture with browsable HTML APIs.
All navigation is through clickable links. No URL construction needed.

Start here: https://pemos.ca/fleet-nav
PEMCLAU query: https://pemos.ca/pemclau-browse
Organism navigator: https://pemos.ca/organism-navigator
Management universe: https://pemos.ca/mgmt-universe
VSM fleet mapping: https://pemos.ca/vsm-fleet
Fermentation schools: https://pemos.ca/fermentation-schools
All pages: https://pemos.ca/repo-galaxy

# Architecture notes
# Each page includes refinement links for further navigation
# Content negotiation: AI user agents receive HTML, browsers receive JSON
# Fleet is antifragile: 100:1 storage-to-throughput ratio (baobab pattern)

DYNAMIC REFINEMENT LINKS — THE KEY INNOVATION

TURNING 50 PAGES INTO THOUSANDS OF DESTINATIONS

Every results page analyzes which filters have NOT been applied. For each unapplied filter, generate a clickable refinement link. AI navigates the full parameter space without constructing any URLs.

Fleet analog: every PEMCLAU result page shows <a href> links for:
• Related organisms (by school, family, OFRAME layer)
• Connected theories (by dimension, era, framework)
• Adjacent crew members (by regulon, compatible pairings)
• Parent/child OFRAME layers (navigate up or down the layer stack)

This turns 50 static fleet pages into thousands of reachable combinations. An AI agent can explore the entire fleet knowledge graph by following links, without ever constructing a URL. This is the SPECAPIS pattern fully realized.

PEMOS IMPLEMENTATION PLAN

Route /llms.txt

Serve the llms.txt file. Static content. No auth. AI agents check this first.

Route /fleet-nav

HTML-only AI navigator. All fleet pages as clickable links organized by category.

Route /pemclau-browse

PEMCLAU query interface. User-Agent detection. HTML for AI, JSON for devs.

Content negotiation middleware

isAIAgent() check in main.go. Route HTML vs JSON based on User-Agent header.

Refinement link generator

For each PEMCLAU result page: analyze unapplied filters, generate <a href> links.

llms.txt route

Add /llms.txt to main.go routes. Serve static fleet API description.