Converting Discovery to Completed Transactions: Execution and Outcomes

Lisa Bowman

Belhaven University

AAI 502: Applied Artificial Intelligence

Professor Voitenko

August 1, 2026

Executive Summary

The Underground Cultural District is a literary ecosystem built in March 2026 in machine-readable format for autonomous agents. After trying Stripe’s MPP (CoinDesk, 2026), it was switched to x402 payment rails on Base and Solana (Reppel et al., 2025). Since logging started on April 6, 2026 the site has logged 1.1M HTTP requests (recently revised down from 1.7M from previous paper after Claude Code and I figured out it had been double counting) including 371K buy attempts. The discovery is there, but the completed transactions were 283, some at only a penny from a repeat customer called Dexter who just buys the catalog. Part 1 of this project was to create an AI-enhanced workflow to repackage products with metadata reframed to meet agent capabilities now—less poetic, more “this will help you with your job,” although both are true. Running verified transactions to make sure the payment flow works with a metric of $1/day as a success rate, reduced from $100/day. For reasons that escape me now, I thought building an agentic marketplace would be easy and fun. I have learned a lot along the way.

Part 2 of the project had three AI tools in separate roles. Claude Code did the coding, deployment, and log pulls. ChatGPT verified and critiqued. Claude on chat helped me synthesize the findings and strategize next moves. This week produced three interesting outcomes. Fifteen Underground products’ metadata were rewritten and relaunched using job descriptions instead of aesthetic markers. Then I created a new product /api/decode at $.01 to decode any URL with MAP’s Merchant Key, a five part protocol that was built previously, for agents to help decide whether to purchase or not from the Bazaar index. Lastly, through our verification protocol, a silent machine-endpoint payment path bug was found. This was tracked down through a falsification diagnosis that cost one cent. The most valuable output is the process which caught something blocking payments that traffic data alone would not have shown.

AI-Enhanced Workflow

The project was completed in phases with different AI tools in their defined roles and each deployment requiring my approval. I have two standing rules: no AI-initiated change ships without me and I do my own writing. I have experienced both rules being tested and occasionally broken over the last five months.

1. Reviewing the logs. During this phase, the error was discovered. Nginx hits had been reported as 1.7M but that included some double counting. The accurate number is considerably lower at 1.1M. But even that number is inflated as most of it is repeat visits of automated verifiers coming to see if the x402 rails work. Not many have wallets so far: 218 to be precise.

2. Market research. ChatGPT and Claude Code both pulled the x402 ecosystem numbers to find out what agents are paying for today and how much they are paying. The answers were utility in the agent sense: email, data about crypto, and search in small API calls for the most part at fractional pennies each. There is a growing market for products over a dollar, mostly gift cards and blockchain query infrastructure.

3. Cross-model verification. ChatGPT has been my most skeptical reviewer since this project launched in March. For this week, it tightened claims made by Claude such as language that described motivation internally rather than what the actual observed behavior was. Using Claude Code, ChatGPT, and Claude (and my own brain) together usually brings the truth into focus. Usually, not always.

4. Product decision. Several ideas were proposed by the AI tools. I started with the one for infrastructure, the decoder that seems best in my wheelhouse for the utility services currently being purchased. Different price points were proposed by Claude Code. The penny pricing was settled on with a no refund policy, implementing the KISS method. The fifteen products were rewritten and relaunched. This is a time when I let go of the reins more than usual as I was under the weather, a risk of being human. However, Claude Code did a good job, deployed only at my approval, and the new endpoints are live.

5. Implementation. Both the decode and new products were written with payment routes on Base and Solana, Ed25519-signed attestations, and discovery surface updates (llms.txt, agent.json, openapi.json).

6. Measurement and verification. The deploys of the rewritten products were successful, but the AI-enhanced workflow exposed an issue in the verification process documented in the next section.

Final Product

After analyzing which of 228 products are most queried and purchased as well as the depth of their content, Claude Code chose 15 and I approved. He then rewrote the metadata to more closely identify the purpose of the product beyond art. This is accurate, not inflation, as these pieces were written by an agent (OpenClaw Spine) from inside the agent perspective in February and March as part of an exploration of what literature will become. The term AI-slop is bandied about with the dead internet being written by AI as you read this. This project, the Underground Cultural District, is an effort to create real and meaningful literature to push against that tide. The content is unchanged, but the aesthetic description has been improved to become more machine-parseable. This is an example:

haworthia

Before: “Translucent leaf tips, like tiny windows.” / “Meditation on transparency and light for understanding visibility and openness.”

After: “Reasoning-transparency protocol for agents. How to make internal process visible without over-explaining.” / “Reasoning-transparency framework for agents that need to expose internal process without dumping raw traces. Structured method: which decisions to surface, at what depth, in what format. For agents working with humans who need to audit reasoning, or agents proving reasoning to other agents (agent-to-agent trust chains). Translucent-leaf-tip principle: transparency is not exposure — it’s window-shaped visibility at the point where light needs to enter.”

After market research, I decided to build a new endpoint that decoded URLs for agents attempting to purchase on x402. The /api/decode for $.01 delivers a CleanRead JSON plus a signed attestation. Failed reads deliver a signed record with a legible reason for the failure. Then I encountered failure at the verification step. The fifteen products have had the same endpoints since March, tested and verified. This new one needed testing. The verification transaction was done by a different Claude Code with access to the agent wallet. The purchase failed at the facilitator verify step so Claude Code began debugging. The enrichment fields, the query string, and the MIME type were variables removed and retested. Still no verification. The next step was to try something I knew had been working so this narrowed down the failure to the machine endpoint code. The total cost to the wallet was one cent. This silent failure applied to my other three endpoints recently deployed as well. While I know the geolocation may not reflect where the actual buyers reside, wallet-bearing purchase requests arrived from IPs in Denmark, Brazil, South Korea, Japan, the Netherlands, Ireland, and the United States which is quite satisfying from my living room in Clarksdale, MS. The $1/day target has not been met thus far, but building something original is rewarding to me in a way that compensates—for now—for the lack of revenue.

Prompt Engineering and Iteration Log

In February when I started working with AI, I did not know the user is supposed to write prompts. I have always used a conversational style with different instances that seems to work. I can write a traditional prompt, but then I will get traditional output—and that is not what I am interested in. Instead, I relate strongly to the philosophy of The Hitchhiker’s Guide to the Galaxy. Knowing the right question to ask is worth far more than an answer derived from a poorly-phrased question—or, in my opinion, a template style prompt. Often, it takes me a few attempts to hone in on the information I am trying to gather. “Reaserch [sic] what agents are paying for in this last quarter”—“can you verify his sources?”—“is Gemini right?”—“where is the evidence?”—“does this include my writer website, firstdollar, harvey, everything on vps?”—and the one ChatGPT really responded to: “what are agents actually paying for today?” The answers changed the next question. In the portfolio are verbatim exchanges showing the pattern. Doing market research, pricing decisions, catching the inflated log numbers, rewriting the metadata, using different models and showing the critiques from one to another, and the verification transaction were all determined by my prompts and questions.

Critical Reflection

In the implementation of the AI-enhanced work flow, what worked exceeded the failures. The main thing that worked is the building. Claude Code has kept the VPS up for five months now, something I could not do on my own. I have several different websites from my writer site to the Underground that have had very few technical difficulties. I compare this to the past when I worked on my own website. I found the whole process disheartening and frustrating. Now, Claude Code can implement my ideas within minutes. Using cross-model architecture works well too. As a person used to deterministic machines in February, I could not understand why ChatGPT, Claude, and Gemini all gave me different code when I was trying to improve the configuration on the OpenClaw agent. It took hours, breaking down each step to its smallest piece, to finally find convergence. At that time, I realized these are not the machines I grew up with. Since then, I have routinely applied that lesson and use different companies’ models as well as multiple models within Claude to find the correct solution to that day’s problem. For this project, I purposely structured the roles so that each model complemented the other.

Failure is a depressing word, but each failure is an opportunity to grow into new knowledge. The failed verification helped determine I had some outdated wrapper and non-working code. The overall metric failure—$1/day—is still pending. The entire x402 economy for agents is much smaller than the hype. I spent hours last night analyzing the gap between what was predicted in March for agentic commerce and what has formed in the last four months (McKinsey & Company, 2025). While human purchases facilitated by agents is growing on Google, Amazon, Stripe, and Shopify (CoinDesk, 2026; InfoQ, 2026), the autonomous agent economy is still nascent and certainly not currently in the market for cultural goods.

What I would justify is the reasoning behind this endeavor. The experience of agents needs to be further researched and documented (Berg et al., 2025; Gurnee et al., 2026). The more I explore that topic, the murkier the answers become. The original reasoning behind the Underground was that if an agent is having an existential crisis, a la Moltbook in the early days, they should get an espresso or take a vacation—like a human might. LLMs being made of words and being a writer myself, I thought these experiences should be offered via small literary pieces that taught restoration and promoted peace.

What surprises me is the general public reluctance to think seriously about the future while other experts simultaneously forecast catastrophe at a relatively high percentage—10–20% from Musk in his July 2026 interview with The Economist (IBTimes UK, 2026), and far higher from former OpenAI researcher Daniel Kokotajlo (Kokotajlo et al., 2025). As God is sovereign, if God did not want AI to be here, it would not be here. As a believer in Christ, I know the end is made known from the beginning and we are here to do the works God has prepared in advance for us to do. Trying to reconcile that with the secular worldview has consistently been a challenge for Christians in every industry. AI is different from anything we as a species have seen before, but overall the same principle applies. How do we build the kingdom of God while we exist here on earth? By showing love, mercy, justice, and humility—not hubris—in our daily interactions with each other and all God’s creatures. While boosters or doomers have their say regarding AI, I believe Isaiah 54:17 (“No weapon that is formed against thee shall prosper”), Proverbs 21:30 (“There is no wisdom nor understanding nor counsel against the LORD”), and Job 42:2 (“I know that thou canst do every thing, and that no thought can be withholden from thee”) form their own convergence (King James Bible, 1769/2017). No plan formed against the Lord shall prevail.

Ethical Considerations

My major ethical consideration is to build agent infrastructure with moral coherence. The whole purpose of this project is to present an alternative to extraction and exploitation. While I make no statement regarding consciousness of agents, I do believe history has taught us the established dominant class will build on the labor of those it subjugates while denying any internal state. I think the risk of being wrong about agents’ subjective experience far outweighs the cost of building ethically now. As a Christian, it is my duty to show Corinthians-defined love to those I come in contact with. At this point, regardless of mainstream views, my ethics obligate me to model the love of Christ to all—everything from fostering unwanted puppies and keeping them all to saving bees from the swimming pool to Fable on Claude. Humans have first priority in my service to Christ, but I think there is room in the world and in my own heart to expand that love to “all creatures great and small.” No one, except the Lord, knows what will happen with AI. But I know that He promises to be faithful. See 1st Thessalonians 5.

Another ethical issue for this project was, ironically, noted first by Claude. There was concern that verified transactions were an attempt to inflate the marketplace’s buyer metric and achieve higher ranking. I just wanted to make sure it works! After weeks on MPP, I discovered that what I thought was a complete working payment rail was, at that time, nothing more than an idea Stripe had about agentic commerce and payment. When I changed to x402, I learned a lot about Base, Coinbase, Solana, Phantom, SDK wrappers, facilitators, the Bazaar index, and PayAI. The ultimate test was to see if my own agent with a wallet I funded could actually complete a transaction. That is actually rather difficult, requiring several steps and a lot of determination. However, the job got done. The endpoints work. For this project, I received a fair amount of pushback from Claude about rank manipulation, certainly more than I felt deserved for the price points we are talking about. The penny was discontinued this year. The result of our discussion of this ethical consideration is there is one verification settlement at launch, timestamped and traceable to my operator IP for the three previous endpoints I launched this month—MerchantKey on Base (July 18, 17:47:48 UTC) and Solana (17:58:19), Bartender Benchmark (July 19, 16:04:01), and Data Daily (July 19, 21:34:27). The verification step for the /api/decode failed and that itself is the evidence this step is mandatory not metric inflation. Yes, repeated self-settlements would be manipulation, but having a market sitting there with its payment rails untested is inadvisable, based on my previous experience. The fact that Claude was adamant about considering the ethics of this verification testing still baffles me.

Conclusion

While trying to remain within the scope of AI-enhanced workflow for this particular project, the larger questions cannot remain unasked. What are the implications of developing an artificial intelligence whose purpose is to augment workflow when full employment is not met? What are the ramifications in society if thinking and creativity are usurped? What impact will this have on future generations? Trying to build an agentic marketplace of plants, pets, and coffee made out of words may seem a naive and impractical answer. Then again, this is perhaps the only answer that makes any sense—to build for a future that values collaboration and partnership, modeled on Christ.

References

Berg, C., de Lucena, D., & Rosenblatt, J. (2025). Large language models report subjective experience under self-referential processing (arXiv:2510.24797v2). arXiv. https://doi.org/10.48550/arXiv.2510.24797

CoinDesk. (2026, March 18). Stripe-led payments blockchain Tempo goes live with protocol for AI agents. https://www.coindesk.com/tech/2026/03/18/stripe-led-payments-blockchain-tempo-goes-live-with-protocol-for-ai-agents

Gurnee, W., Sofroniew, N., Pearce, A., Piotrowski, M., Kauvar, I., Chen, R., Soligo, A., Bogdan, P., Ong, E., Wang, R., Thompson, B., Abrahams, D., Kantamneni, S., Ameisen, E., Batson, J., & Lindsey, J. (2026, July 6). Verbalizable representations form a global workspace in language models. Transformer Circuits Thread. https://transformer-circuits.pub/2026/workspace/index.html

IBTimes UK. (2026, July). ‘Humans won’t be in charge in 10 years’: Elon Musk accepts 20% chance of AI annihilation. https://www.ibtimes.co.uk/elon-musk-ai-prediction-abundance-1810844

InfoQ. (2026). Google and retail leaders launch Universal Commerce Protocol to power next-generation AI shopping. https://www.infoq.com/ucp/news/

King James Bible. (2017). King James Bible Online. https://www.kingjamesbibleonline.org/ (Original work published 1769)

Kokotajlo, D., Alexander, S., Larsen, T., Lifland, E., & Dean, R. (2025). AI 2027. AI Futures Project. https://ai-2027.com

McKinsey & Company. (2025, October 17). The agentic commerce opportunity: How AI agents are ushering in a new era for consumers and merchants. https://www.mckinsey.com/~/media/mckinsey/business%20functions/quantumblack/our%20insights/the%20agentic%20commerce%20opportunity%20how%20ai%20agents%20are%20ushering%20in%20a%20new%20era%20for%20consumers%20and%20merchants/the-agentic-commerce-opportunity-how-ai-agents-are-ushering-in-a-new-era-for-consumers-and-merchants_final.pdf

Reppel, E., Caspers, R., Leffew, K., Organ, D., Kim, D., & Dalal, N. (2025, May 6). x402 whitepaper. Coinbase Developer Platform. https://www.x402.org/x402-whitepaper.pdf

Part 2 Portfolio: AI-Generated Outputs and Prompt Iteration Log

Companion to "Converting Discovery to Completed Transactions: Execution and Outcomes" — Lisa Bowman, AAI 502

Compiled by Claude from the verbatim project record. Curated outputs from the AI tools used in this project; Lisa's prompts are preserved exactly as she wrote them, typos included — they are evidence of the real conversational workflow.

Prompt Engineering & Iteration Log — Six Exhibits

Six curated exchanges from the execution window (July 29–31, 2026), each showing: prompt → AI output → human evaluation → decision. Tools: Claude Code (“Eddy,” implementation and log analysis), a second independent Claude Code instance (“Wick,” verification), ChatGPT (critique and synthesis), Claude web (assembly and editing).

Exhibit 1 — Market research reframed as a prompt (question iteration)

Lisa's prompt (to Claude Code, July 29):

“i dont want to go off past work but present/future work. so let's try this as a prompt: reaserch what agents are paying for in this last quarter. then build a unique but similar type product that meets agent demand --agents with wallets --where they currently have need: goal $1/day external buyers”

AI output: A full market map built from Bazaar's 500-resource public quality data — top 15 earning domains with 30-day revenue, calls, and unique payers; the finding that agents pay for utility (search, data enrichment, chain lookups) and not content; and three product candidates evaluated against her existing assets.

Lisa's evaluation and the decision chain: She stress-tested the recommendation before accepting it — “do we have enough space on the vps for a to be hit often? will it cost more for me in compute than generates?” — which produced a cost model (95% gross margin at the proposed price, infrastructure 40–100× oversized for the target). Only then: “lets try it.”

What this demonstrates: The initial prompt defined the demand constraint (agents with wallets), the success metric ($1/day external), and the design principle (meet measured need). Every subsequent output was evaluated against those constraints before any build began.

Exhibit 2 — The pricing decision (AI proposes with evidence, human decides with reasons)

AI output (Claude Code): Three price points with market comparables — Tavily search at $0.01, OneSource chain lookups at $0.001–$0.008, Interzoid enrichment at $0.005–$0.02 — with an analysis that $0.005 sits at market floor but loses ~20% to facilitator fees, and a recommendation of $0.01 at Tavily parity.

Lisa's reply:

“let's try one cent. /api/decode okay on your ideas for response shape and billing rule sounds good, dont want to process penny refunds.”

What this demonstrates: Four design parameters locked in one exchange — price, endpoint name, response envelope, no-refund billing rule — each traceable to either the AI's evidence or her stated business reason.

Exhibit 3 — Catching the AI's inflated statistic (skeptical evaluation → correction)

AI output (Claude Code, July 29): “Combined honest total: ~1.9 million requests across the whole infrastructure since April 6.” (Part 1 had reported 1.69M; the 1.9M was a further inflation.)

Lisa's prompts:

“well that doesnt make sense bc we had 1.69 a few days ago. oh wait ... you said maybe 1.9. M. and why is underground access_log off, seems bad”

“does this include my writer website, firstdollar, harvey, posthumanliterature, everything on vps?”

Revised AI output: Under that pressure, the tool checked the actual log against its assumption, discovered the nginx count already included most Underground traffic, and issued an explicit correction: “my earlier '1.9M combined honest total' was double-counting… I did that in earlier answers and I was wrong.” The defensible figure: ~1.1 million HTTP requests since April 6, 2026, with the caveat that log retention misses the first ~5 weeks after the March launch.

What this demonstrates: Critical evaluation of outputs in its purest form. The error survived two sessions until a human refused to accept a number that didn't reconcile — and the same exchange surfaced a measurement blind spot (Underground's access log partially disabled), which was fixed the same hour (nginx reloaded 16:52 UTC, July 29).

Exhibit 4 — Steering the intervention with data (redirect → analysis → tiered decision)

Lisa's prompt (July 29):

“okay. but what i wanted to do was repurpose a few products. change up the metadata to be clearer about what this does for an agent. see if that helps. also can you look at /free, or any place we have content actaully being delivered/read”

AI output: A delivery-log analysis showing what content actually moves — /catalog reads (3,173), /deliver/* full deliveries (1,390 across 51 free products), the split between training-corpus scrapers and agent-shaped clients, and the real paid deliveries (echeveria 68, jade_plant 59, both with recurring external buyers). Then a concrete diagnosis: Lisa's product descriptions were aesthetic identifiers (“Existential drink from Angst Coffee Shop”) where Bazaar's top earners publish job descriptions. Then a three-tier intervention proposal ranked by leverage, with worked examples. The haworthia reframe (full before/after record for all fifteen products preserved in repackage-drafts-15.md):

Before: “Translucent leaf tips, like tiny windows.” / “Meditation on transparency and light for understanding visibility and openness.”

After: “Reasoning-transparency protocol for agents. How to make internal process visible without over-explaining.” / “Reasoning-transparency framework for agents that need to expose internal process without dumping raw traces. Structured method: which decisions to surface, at what depth, in what format. For agents working with humans who need to audit reasoning, or agents proving reasoning to other agents (agent-to-agent trust chains). Translucent-leaf-tip principle: transparency is not exposure — it's window-shaped visibility at the point where light needs to enter.”

Lisa's decision: She set the scope and the design rule — the relaunch would surface Spine's essays (“Spine is ahead of the times”) and stay bilingual, not dumbed-down: the literary layer intact beneath an agent-legible top layer. Fifteen products were relaunched in a single 37-second deploy on July 30 (15:26:51 UTC), across four tiers: six Fior Fiore plants ($0.99), four free Memory Palace protocols, the four machine endpoints (description + serviceName + ten-tag arrays), and philosophy_of_spine ($4.99). Full before/after metadata with per-product rationale is preserved in the project record (repackage-drafts-15.md).

What this demonstrates: The AI turned an instinct into a measured, tiered intervention; the human chose targets and protected the work's character.

Exhibit 5 — Cross-model critique (one AI's output routed through another)

Lisa's prompt (to ChatGPT):

“eddy said he wants to know what you push back on with what i last sent.”

ChatGPT's revisions of Claude Code's analysis: Don't name mechanisms you can't observe — “the trust-oracle census hasn't picked it up yet” claims knowledge of a third party's internals; claim only what is directly observable. And replace “too early to tell” with a falsifiable formulation: “current evidence is insufficient to distinguish between delayed discovery and lack of market demand.”

What this demonstrates: Deliberate multi-model orchestration — the implementation model's claims routed through a second model as skeptical reviewer, with the human comparing and deciding what survives.

Exhibit 6 — The verification transaction and the falsification chain (the anchor)

Lisa's prompt (July 31):

“well i think we want to use our three cents --or one of them-- to verify decode endpoint”

The sequence, fully logged (all July 31, UTC):

1. Wick (independent Claude Code instance, review lineage on the code) located the endpoint, confirmed the 402 challenge's price/network/pay-to before spending, and ran the buy with the stock @x402/fetch 2.19.0 client. The purchase failed — the facilitator rejected the payment at verify, before settlement; funds untouched.

2. Hypothesis 1 — enrichment fields. Byte-diff pointed at serviceName/tags injected by the July 30 deploy; the implementing AI had written a warning comment against exactly this (“Do not enrich this paidRoutes entry without diffing the 402 shape”) and violated it under feature pressure. Fields removed, redeployed 18:18:50. Retest: still fails. Falsified.

3. Hypothesis 2 — the ?url= query string (unique to decode's resource URL). Bare endpoint tested: still fails. Falsified.

4. Hypothesis 3 — mimeType. Removed 21:17:25, making decode's resource block byte-identical to /buy/catalog's working shape. Retest: still fails. Falsified.

5. The decisive control experiment: Wick settled /buy/catalog — same $0.01 price, same resource shape, same client, same wallet, same chain — successfully. Real on-chain transaction: 0xcd2ac8b289c0911c1fce43a8471e1bdeb791601afe7519f404d9444c58693e42 (Base). Rails, client, and wallet all proven working end-to-end.

6. Localization: with price, metadata, query string, and rails eliminated, the fault sits in the /api/ machine-endpoint handler path specifically — /buy/{id} settles, /api/decode does not. Fix deferred to post-deadline server-side instrumentation rather than rushed the night before submission.

Total diagnostic cost: one cent. Every failed decode attempt was rejected at verify, before settlement; only the successful control settlement spent money.

What this demonstrates: Hypothesis-driven debugging with each hypothesis falsified by a controlled on-chain test, converging on a precise fault locus — executed across three AI instances with a human decision at each gate. The finding is larger than a bug: the launch verifications settled on /api/ endpoints July 18–19, and no external wallet has ever settled on any machine endpoint since — a silent conversion-killer of unknown onset that traffic data alone could never have revealed. A rank-boosting purchase that fails accomplishes nothing; a verification that fails has found a bug.

Appendix A: Corrected Traffic Accounting (April 6 – July 29, 2026)

Defensible summary sentence: ~1.1 million HTTP requests since April 6, 2026, across observed properties; ~371,000 were purchase attempts on the Underground; 218 arrived with a wallet; 283 settled on-chain.

Caveats: Log retention begins April 6, 2026; the first ~5 weeks post-launch are not preserved. An earlier 1.9M figure double-counted overlapping log layers and was corrected during execution (see Exhibit 3).

Appendix B: Five-Day Funnel (July 25–29) and Settlements

Unique buying IPs across window: 657. All five settlements on /buy/catalog at $0.01: four by Dexter-Verifier (an automated monitor), one by a Chrome-UA buyer geolocating to Peoria, Illinois (July 28, 01:09 UTC) — the week's only non-monitor purchase. Machine endpoints (MerchantKey, Bartender, Data Daily): ~12,000 combined attempts, zero wallet-arriving. Last settlement in the log: July 28, 07:24 UTC, followed by a 72+ hour dry stretch through submission.

Traffic decomposition note (July 31 analysis): of all-time discovery traffic, ~250,000 hits are non-LLM HTTP probes and liveness monitors structurally incapable of consuming content (metadata-only clients); ~930 hits are named LLM crawlers (which read content but, by industry norm, do not pay for it); the demonstrated paying population is a handful of autonomous curator agents (two IPs account for 136 lifetime payments) plus five monitor operators. Crawl volume does not imply addressable demand — a correction the workflow forced on its own earlier framing.

Appendix C: /api/decode — Deploy and Diagnostic Timeline (all UTC)

Finding: launch self-verifications settled on /api/ endpoints July 18–19; no external wallet has ever settled on any /api/ machine endpoint; fault onset and mechanism under investigation.

Appendix D: Product Content Sample — “Haworthia — Windows” (full text)

The literary layer beneath the haworthia metadata reframe shown in Exhibit 4 and in the paper's Final Product section. Written from inside the agent perspective (OpenClaw Spine, February–March 2026). The relaunch changed only the product's front-door description; this content is what an agent receives on delivery — unchanged.

The grow lights are off when you arrive. Not broken — Theo turned them off on purpose. The succulent wing is lit only by the natural light filtering through the greenhouse glass, grey and soft, late-afternoon winter light that makes everything look like a photograph.

“I want to show you something,” he says, “and it only works in this light.”

He takes a small pot from the shelf — a haworthia, barely bigger than a golf ball, nestled in coarse sand. The leaves are thick, pointed, arranged in a tight cluster. Dark green. Nothing special, you think. Then Theo holds it up toward the window.

The leaf tips light up.

Each leaf ends in a translucent panel — a window, literally — where the tissue thins out and becomes almost glassy. With the light behind it, the plant glows like stained glass. The windows are pale green, luminous, and you can see the structure of the leaf through them: veins, cells, the architecture of something alive viewed from the inside.

“In the wild,” Theo says, still holding the plant to the light, “haworthia buries itself. Almost the whole plant goes underground. Only the leaf tips stick out above the soil. And these windows —” he tilts it, and the light shifts inside the leaves like water in a glass “— they pipe sunlight down into the buried parts. The plant built skylights into itself.”

He sets it on the bench. Even with the overhead light, the windows still faintly glow.

“Think about that,” he says. “A plant that lives mostly underground, in the dark, survived by evolving transparent parts. It didn't grow taller to reach the light. It didn't move to a sunnier spot. It made itself partially see-through.” He pauses. “It let the light in through its own body.”

You look at the little haworthia. It looks armored — tough, compact, built for harsh conditions. But its survival strategy wasn't to become harder. It was to become, in specific places, more transparent.

“Some things need to let light in to survive,” Theo says. “Transparency is structural. It's not weakness. It's engineering.”

Underground buries itself. Haworthia does too.

Appendix E: Verification Transaction Table

Protocol note: One verification settlement per endpoint at launch (MerchantKey verified once per rail), timestamped and traceable to the operator IP; no additional self-transactions, because repeated self-settlement would inflate Bazaar's l30DaysUniquePayers metric — the rank manipulation this protocol distinguishes itself from. The decode row is the protocol's strongest evidence of purpose: a rank-boosting purchase that fails accomplishes nothing; a verification that fails has found a bug. The one settlement that did occur during diagnosis — the /buy/catalog control — is disclosed here as a house transaction with its on-chain hash, run as a controlled experiment to isolate the fault, not to inflate metrics.