A machine buyer, two inference routes.

Process continuity

The first buyer process completed 12 purchases and stopped. A fresh process recovered its durable state and completed the remaining 12. It did not ask the supplier to resend earlier responses.

1. Live phase 1

   2026-09-05T03:49:50Z to 2026-09-05T03:51:26Z

   Source sha256:4c02691a151cc96d1e22219b656086cbbfbc13f78910bf98ed1f1255f8cf02cc

2. Live phase 2

   2026-09-05T04:36:14Z to 2026-09-05T04:37:29Z

   Source sha256:4c02691a151cc96d1e22219b656086cbbfbc13f78910bf98ed1f1255f8cf02cc

The following 6 tests were injected afterward using the sealed P02 response, with network access denied. They are controlled buyer-process failures, not live provider outages.

The simulator recorded 11 unique consumptions in the live batch and 0 duplicates. Its dispatch keys make repeated local application idempotent. That does not establish exactly-once external execution or payment.

Evidence retained

When response payloads were removed during replay, every purchase became unresolved under both signed representations. A commitment identifies bytes; it cannot reconstruct a missing schedule.

The public corpus includes 143 exact members and reproduces 23 original decisions. P01’s entire account-bearing error response remains private. Its existing exclusion notice is published, with no redacted substitute. Public replay leaves P01 unresolved because its payload is absent; reproducing the original HTTP-error assessment requires the withheld original.

Methods and limits

Ordinary Ed25519-signed JSON and ActionReceipt carried the same local-observer statements and produced the same decisions. The measured format overhead is reported below; the experiment did not measure integration effort or an interoperability advantage.

Counts cover adapter/evaluation work, not native signature-internal hashing or total storage cost. The same retained payloads and buyer logic served both formats.

OpenRouter

Supplied the common interface and requested provider routing. A routing-parameter change is not evidence that Bulla reduced integration work.

Buyer controller and custody

Supplied the scheduler checks, durable decisions and recovery. These are research-harness capabilities, not features supplied by the Bulla package.

Local simulator

Consumed accepted schedules once per dispatch key. No physical jobs were scheduled and no settlement was executed.

Bulla

Created portable ActionReceipts. Signatures authenticated the local observer’s statements, not OpenRouter, its providers or historical model execution.

Inference charges followed OpenRouter’s service rules even when the buyer rejected an answer. The known service-reported cost sums to USD 0.00288350, with 1 unresolved charge; it is not a verified total bill. This small sample does not support a provider-performance ranking.

The result does not establish market demand, independent buyer adoption, provider authentication, direct-provider interoperability, payment, legal recourse or operation without a common router. As of the recorded study, existing x402 signed offers and receipts and payment identifiers addressed adjacent payment and retry concerns. Neither x402 nor a decentralized marketplace was implemented or comparatively tested here.

Downloads and reproduction

Download the frozen source and evidence to inspect the measurements or replay the public decisions. No API key or additional inference request is needed. Local reproduction was validated on macOS arm64 with Python 3.11; portability remains untested.

1. Check the archive against its detached SHA-256, then extract it. The source inventory binds each file to the frozen experiment commit.
2. Create a fresh Python 3.11 environment and install the hash-pinned requirements. The Bulla wheel hash and dependency identities are in DEPENDENCIES.json.
3. From the extracted directory, run the network-denied public replay. Never run the capture command: the recorded batch is complete.

python3.11 -m venv .venv
.venv/bin/python -m pip install --require-hashes -r requirements-macos-arm64-py311.txt
.venv/bin/python -I -B analysis/replay_safe.py evidence /absolute/new/safe-replay.json

The replay path must not already exist. It reproduces the public decisions without credentials and cannot reconstruct P01. Expected replay results are in SAFE-REPLAY.json.

Historical records retain their original local-only authorization and status text. This publication does not rewrite that completed experiment or promote its controller into a product. The reviews are project-operated model reviews, not independent human adoption evidence.

Frozen experiment commit

1e0a8d6f8a00faca3a0f9245868a50bb52ffe56f

Exact public corpus root

sha256:8f9a69df24a8c8e32e4bfca0e5be6c3255a2bcb3ec254202addd40154d75e4cd

Buyer source root, unchanged across processes

sha256:4c02691a151cc96d1e22219b656086cbbfbc13f78910bf98ed1f1255f8cf02cc