# Field Test 001: Erratum and Run Notes, October 13, 2026

**Protocol:** E.D.A.I. Field Test 001, protocol v1.1 ([PROTOCOL.md](PROTOCOL.md), SHA-256 3121c204d9cfa8535ba322e91b15386b7180a3e04fcdfa7c6b3a103f17f4c185)
**Status at the time of writing:** no request has been sent to the TypeSafe API under this protocol. No model output has been seen.

The protocol says any error found later is appended as a dated erratum and the original text stands. This is that note. PROTOCOL.md and the frozen files are unchanged. Everything below is fixed before the first request, and this note is fingerprinted on Hedera, topic 0.0.9376001, before the run begins.

## 1. Where the requests come from

Section 5.5 says requests originate from an Anthropic-hosted Linux sandbox. That sandbox cannot reach TypeSafe's API. The run will instead come from the author's own computer in Austin, Texas. This changes only the latency measurement, which the protocol already calls secondary. Latency will be reported as measured from Austin. The Claude baseline runs from the same computer, as Section 7 requires.

## 2. The baseline prompt, now fixed

Section 7 fixed the instruction's intent but not its exact words. They are fixed now:

- Model: `claude-haiku-4-5-20251001`, maximum 16 output tokens, default sampling settings.
- System text: "You will be given a passage and a proposition about it. Reply with the probability, between 0 and 1, that the proposition is true. Reply with the number only and nothing else."
- User text: "Passage:" followed by the item's state, a blank line, then "Proposition:" followed by the question.
- One request per item for the primary question, one for the paraphrase, and three repeats of the primary for jev-0000 through jev-0199: 4,600 requests.
- A reply that is not a single number between 0 and 1 is counted as missing. The number of missing replies is reported, and they are excluded from the statistics, never re-requested.

## 3. Harness details the protocol left open

The harness is published here as [run-ft001.cjs](run-ft001.cjs), SHA-256 42c9bf4658ddf25bbc0bc21e536f3a9d33b20dd3e6587f57b2b6a85395b30497. It will not change before or during the run. It fixes these details:

- **Order:** the 2,000 main requests in id order, then three repeat rounds over jev-0000 through jev-0199, each round in id order.
- **Request ids:** the three questions are sent under the ids `primary`, `paraphrase` and `choice`. TypeSafe's documentation says question ids are not sent to the model.
- **Party labels** for the EQUAL questions are read from the item's own state line (Petitioner/Respondent or Appellant/Appellee), never from the labeled file.
- **Pause:** a fixed 250 milliseconds between requests.
- **Retries:** network errors, timeouts (60 seconds), HTTP 429 and HTTP 5xx are retried up to 3 times, waiting 2, 4 and 8 seconds. Every attempt is logged. Any other error stops the run, and the run resumes from where it stopped. A successful request is never sent again.
- **Blindness:** the harness never opens `items_labeled.jsonl`. The answer key is not read until analysis.

## 4. The item files were rebuilt and match

The item files were not kept from September 30. They were rebuilt with the frozen `build_items.py` from the public transcript mirror (walkerdb/supreme_court_transcripts, commit 1171edccc208b759f3ee227aab75c15062adaaf7, the newest commit as of September 30). The rebuilt files match the frozen hashes exactly:

```
items_public.jsonl   b97e66b1b30852ad0af58724c844c578335e4d4b6768335a8ab767df51ffffd6
items_labeled.jsonl  5a12c05c3a5f2fdacadefa619e2b20803fb2d2ecbf827daa76e2bc69e835ef66
```

## 5. One connectivity check

Before the run, one request is sent to each service on a made-up line that is not a test item ("This is a connectivity check, not a test item."), to confirm the keys and the request format work. Its output is not part of any statistic.

## 6. Arm B may not run

The transcript mirror has not been updated since May 9, 2026. Arm B runs only if transcripts from the October 2026 session appear there. If they never do, Arm B does not run, and the write-up says so.

## 7. Publication, and sealing the results

TypeSafe has not given consent to publish (see the [status note of October 13, 2026](https://github.com/ethicsbuild/receipts/blob/main/field-test-001/status-2026-10-13.md)). So the run is private and the results are sealed, under the commitment made there. When each run finishes, the SHA-256 of its response file is posted to topic 0.0.9376001. That proves the run happened once and the responses were not changed afterward, without revealing them.

Gage Cass Woodle