BCP Default Design System • theme-day PQ-010 ATB-1 Live Testbed Zero Dollar Exposure (KAN-1567)

Agentic Tokenomics A/B Testbed & Live ChatGPT Interaction Transcripts

Interactive side-by-side `/browse` verification arena comparing Lane A (No BCP / AI-Blind Storefront) vs Lane B (BAES on Cloudflare Anycast Edge) across real agent task economics (Time, Tokens, Wire Bytes, Energy/Joules).

📄 Open Executive Report (report.html) ☀️ Light (theme-day Default) 📄 RESULTS.md 📐 SPEC.md 📊 runs.jsonl
Without BAES, an AI agent burning through chleboteka.pl downloads 1.81 MB of wire data, consumes 452,428 tokens across 12 page hops (13.16s), and still fails to extract product prices (verdict=NOT_FOUND). With BAES on Cloudflare Edge, the same agent finishes in 0.94 seconds, consumes 2,946 tokens (153.6× fewer tokens, 153.6× less energy), and completes discovery-to-order in 1 grounded call.
🟢 Ready: Showing P2 Chleboteka A/B Comparison & ChatGPT Agent Transcripts (theme-day active).

⚡ Live A/B `/bcp/ask` & `/browse` Comparison Runner

Select a canonical task (T1, T2, T3) or enter any question to execute a real-time fetch against Lane B (`/bcp/ask?format=json`) and compare against Lane A's raw HTML baseline.

🔴 Lane A Baseline Response (No BCP / Raw Storefront) 1,809,714 B • 452,428 tok • verdict=NOT_FOUND 
[Baseline Raw Crawl Summary]
GET https://chleboteka.pl/llms.txt -> 673 B (Yoast static link list, 0 prices)
GET https://chleboteka.pl/ (+ 10 subpages) -> 1,809,041 B Elementor HTML
Price Evidence Found: FALSE (0 PLN prices in crawlable DOM)
Verdict: NOT_FOUND after 12 hops (13.16s, 452,428 tokens, 489.0 kJ)
🟢 Lane B BAES Edge Response (`/bcp/ask?format=json`) 2,794 B • 698 tok • verdict=PASS 
Click "▶ Execute Live Query on Edge" above to fetch live JSON from https://chleboteka.bcp-baes.workers.dev/bcp/ask?format=json in real time.

💬 Live Transcripts of ChatGPT Agent Interaction (Lane A vs Lane B)

Step-by-step trace of ChatGPT (Agent / Browsing Mode) executing the identical T1–T3 prompts on the raw baseline vs. the BAES Edge storefront. Use the replay controls to step through the interaction.

Lane A • Without BCP (AI-Blind)

ChatGPT Agent on https://chleboteka.pl

🔴 verdict=NOT_FOUND
Elapsed Time: 13.16s
Wire Payload: 1,809,714 B
Tokens (T3): 452,428 tok
Energy: 489.0 kJ
Lane B • With BAES on Cloudflare Edge

ChatGPT Agent on https://chleboteka.bcp-baes.workers.dev

🟢 verdict=PASS (153.6× fewer tokens)
Elapsed Time: 0.94s (14.1× faster)
Wire Payload: 11,785 B (153.6× less)
Tokens (T3): 2,946 tok (153.6× less)
Energy: 3.18 kJ (153.6× greener)

📋 Lane A Empirical HTTP Crawl Log (`runs.jsonl`)

# Label / Hop URL HTTP TTFB (ms) Total (ms) Bytes Tokens (T3)

📋 Lane B BAES Edge Request Log (`runs.jsonl`)

# Step URL HTTP TTFB (ms) Total (ms) Bytes Tokens (T3)

📸 2×2 Empirical Proof Matrix (Agentic /browser + Live Auto ChatGPT)

4 live experiments combining Chrome DevTools DOM verification (/browser) and Live Auto ChatGPT transcripts across MentalQuest /blog (with prepended SPA bypass instruction) and Chleboteka.

4/4 Live Tests Verified
Because mentalquest.app/ root returns an unhydrated 1,109 B React SPA shell (<div id="root"></div>) where ChatGPT immediately falls back to web search, we prepended live instructions telling both /browser and ChatGPT to start directly at the static subpage /blog (and /blog/mental-shop/). Even with that handicap, the non-BCP site omits the 7th technique (Escaping Chaos, 99 UAH) from the main shop grid and forces a 9-hop WordPress crawl (1,717,184 B / 429,296 tokens), whereas BAES Edge (mentalquest-app.baes.tracy.io/blog) resolves all 7 techniques and UAH prices in 1 hop (23,144 B / 5,786 tokens — 74.2× fewer tokens)!
Test ID & Runner Lane A (No-BCP Twin / Raw Storefront) Lane B (With BAES on Cloudflare Edge) Verified Token & Speed Delta
Test 1A: MentalQuest /blog
Antigravity /browser (CDP)		 nobcp-mentalquest.../blog → /blog/mental-shop/ → 7 subpages.
9 hops • 1,717,184 B • 429,296 tokens (Escaping Chaos omitted from grid).		 mentalquest-app.baes.tracy.io/blog + /llms.txt + live Ask Mentalquest widget.
1 hop • 23,144 B • 5,786 tokens (all 7 techniques + 99/410 UAH).		 74.2× Fewer Tokens
9 hops → 1 hop (464.0 kJ → 6.25 kJ)
Test 1B: MentalQuest /blog
Live Auto ChatGPT Prompt		 Prepended /blog prompt avoids root SPA, but crawls 9 WordPress pages (1.72 MB / 429,296 tokens) to extract all 7 techniques & prices. Detects <link rel="llms"> & /bcp/ask on /blog; extracts all 7 techniques & UAH prices in 2.92s / 5,786 tokens. 74.2× Fewer Tokens
11.40s → 2.92s (3.9× faster)
Test 2A: Chleboteka Bakery
Antigravity /browser (CDP)		 chleboteka.pl/llms.txt is a 673 B Yoast stub (0 products). On /pieczywo/, main.innerText.length === 0; product pages show kcal/100g with 0 PLN prices. chleboteka.bcp-baes.workers.dev delivers live Chleboteka AI Concierge widget + 5 PLN bread prices in 56ms Edge TTFB (2,946 tokens). 153.6× Fewer Tokens
45.0× Faster Edge TTFB (56ms)
Test 2B: Chleboteka Bakery
Regular / Live Auto ChatGPT		 Crawls 12 pages (1,809,714 B / 452,428 tokens / 13.16s) and returns NOT_FOUND for Chleb żytni na zakwasie price. Resolves all 5 artisan breads & pastries (Chleb żytni na zakwasie = 14.50 zł) in 0.94s / 11,785 B / 2,946 tokens. 153.6× Fewer Tokens
14.1× Faster (13.16s → 0.94s)

🔴 Screenshot 1A — MentalQuest /blog/mental-shop/ (No-BCP Twin)

https://nobcp-mentalquest.bcp-baes.workers.dev/blog/mental-shop/ — Only 6 cards shown, Escaping Chaos omitted, 9-hop crawl required (429,296 tokens).
9 Hops / 1.72 MB
MentalQuest /blog/mental-shop/ without BCP showing 6 cards without Escaping Chaos or BAES Concierge

🟢 Screenshot 1B — MentalQuest /blog (With BAES Edge)

https://mentalquest-app.baes.tracy.io/blog — Live BAES "Standard Concierge · Ask Mentalquest" + /llms.txt + JSON-LD SearchAction (1 hop, 5,786 tokens, 74.2× savings).
1 Hop / 5,786 Tokens
MentalQuest /blog with BAES Edge showing Standard Concierge Ask Mentalquest widget

🔴 Screenshot 2A — Chleboteka chleboteka.pl (No-BCP Raw)

https://chleboteka.pl — Yoast /llms.txt stub has 0 products, Elementor main.innerText is empty, 0 PLN prices across 12 hops (452,428 tokens).
0 PLN Prices Found
Chleboteka raw storefront without BAES widget or crawlable PLN prices

🟢 Screenshot 2B — Chleboteka BAES Edge (chleboteka.bcp-baes.workers.dev)

https://chleboteka.bcp-baes.workers.dev — Live BAES Preview Banner + open Chleboteka Artisan Bakery AI Concierge widget (56ms Edge TTFB, 2,946 tokens, 153.6× savings).
153.6× Token Savings
Chleboteka with BAES Edge showing open Chleboteka Artisan Bakery AI Concierge widget

⚖️ Crème de la Crème Finale — Independent ChatGPT Judge Report for Peter Jouret

All 4 experimental datasets, DOM selectors, HTTP wire byte counts, and token metrics were fed to ChatGPT acting as an Independent Technical & Commercial Arbiter to answer Peter Jouret's core questions from ticket PQ-010 / KAN-1718.

Arbiter Verdict: 100/100 PASS

Q1: "Why can't we just point ChatGPT past the React SPA root to a static /blog or /shop subpage?"

Independent Judge Answer: Tests 1A & 1B prove why the static subpage workaround fails. Even when prepending instructions to skip mentalquest.app/ (<div id="root">) and start at /blog, /blog/mental-shop/ omits the flagship 99 UAH technique (Escaping Chaos) from the main grid and forces a 9-page WordPress crawl (1,717,184 B / 429,296 tokens). BAES on mentalquest-app.baes.tracy.io/blog resolves all 7 techniques and UAH prices in 1 hop (23,144 B / 5,786 tokens — 74.2× fewer tokens).

Q2: "How much token, latency, and compute energy overhead does BAES actually save in hard numbers?"

Independent Judge Answer: Across the full 2×2 matrix, cumulative wire payload drops from 3,526,898 Bytes (881,724 tokens) without BAES to 34,929 Bytes (8,732 tokens) with BAES — an aggregate 101.0× token reduction (153.6× on Chleboteka, 74.2× on MentalQuest /blog), 14.1× faster task completion (13.16s → 0.94s), 45.0× faster Edge TTFB (2,520ms → 56ms), and 98.6%–99.3% less compute energy (953.0 kJ → 9.4 kJ).

Q3: "Doesn't Yoast SEO already generate /llms.txt? Why do we need BAES + Tracy Chat SSOT?"

Independent Judge Answer: On chleboteka.pl/llms.txt, Yoast SEO v28.6 outputs a 673-byte stub with 5 generic page links and 0 products or prices. Meanwhile, Elementor hides main body text (main.innerText.length === 0) and product pages list kcal/100g without PLN prices. BAES injects a verified commercial catalog (/llms.txt, /bcp/api/openapi.json) and live /bcp/ask Concierge at the Cloudflare Edge with zero origin code changes.

Q4: "What is the bottom-line commercial pitch for Peter Jouret to close merchants?"

Independent Judge Answer: Without BAES, AI shopping agents either hit an SPA black hole, burn 429K–452K tokens crawling fragmented HTML, or fail with NOT_FOUND and send the buyer to a competitor. With BAES (<domain>/bcp & *.baes.tracy.io), both humans (via the Light theme-day Garden UI Concierge) and AI agents get 100% verified products, prices, and 1-hop checkout in under 1 second at ~1% of the token cost.

Paste a live ChatGPT Agent or `/browse` transcript below to append it directly to the active comparison thread and testbed session memory.