# The accepted result is the unit

Observed 8 October 2026, Hong Kong time. Research question: what does OpenAI's Jump Trading case establish about autonomous research, and can one small read-only browser job be accepted against destination evidence rather than a successful-looking run?

## The five-condition browser job

The job was fixed before review: open the named Jump Trading page and answer five questions using public OpenAI sources only. No login, connector, transaction, code execution, message, account change or financial action was allowed.

| Predeclared condition | Acceptance evidence | Result |
| --- | --- | --- |
| 1. The named source resolves to the claimed case | OpenAI page titled “How Jump Trading is scaling quant research with ChatGPT,” dated 6 October 2026 | Met |
| 2. The operating unit is identifiable | The case describes a problem, work environment and evaluation of result quality/significance | Met |
| 3. Human acceptance remains visible | The case requires human review and critical validation before acceptance | Met |
| 4. The autonomy claim is bounded | The case says even the longest-running work still receives regular human check-ins | Met |
| 5. Missing production measures are not inferred | The page publishes no task count, acceptance rate, intervention count, recovery time, latency or cost per accepted result | Met as a limitation |

The destination evidence was the three reopened public pages and this hash-bound artifact. The read-only retrieval completed without human intervention. No recovery event occurred, so recovery time was not exercised. This single accepted job is evidence about the workflow used here, not a measured reliability rate for a model or for Jump Trading.

## Claim map

### What the Jump case supports

- Jump describes GPT-6 Astra handling longer, more ambiguous research work, from coding through advanced quantitative studies.
- Researchers define the problem, environment and a way to judge quality and significance, and can steer agents during the work.
- Jump emphasizes secure, monitored environments, clear constraints, observability and human review.
- A produced trading signal remains informative but potentially wrong and enters a separately controlled execution environment.
- Regular human check-ins still occur around data, duration, importance and whether intermediate results make sense.

### What the case does not quantify

The page gives no denominator for completed tasks and no matched comparison with another model or a human-only process. It does not publish acceptance rate, retries, interventions, elapsed recovery, total call cost, research P&L, or the number of findings that survived review. It is a customer account published by the model provider, not an independent evaluation.

The missing numbers do not make the account false. They limit its use. It is evidence of a reported operating pattern and design philosophy, not enough to calculate productivity or autonomous reliability.

## Why browser completion is not enough

OpenAI's earlier Computer-Using Agent publication illustrates the denominator problem. It reported 58.1% on WebArena, 87.0% on WebVoyager and 38.1% on OSWorld, while noting a gap from human performance on complex tasks. Those are benchmark results from an earlier system, not measurements of GPT-6 Astra or Jump's private harness. They are included only to show why task definition and acceptance criteria matter.

The ChatGPT agent release also describes a mixed tool system: visual browser, text browser, terminal and APIs. It allows interruption and takeover, asks for confirmation before consequential actions and warns that prompt injection can redirect an agent. Therefore “browser job” is not a sufficient experimental description. A useful record names the environment, tools, permissions, completion test and destination proof.

## Reusable acceptance card

For the next bounded browser test, record these fields before execution:

1. **Task and destination:** the exact question, source set and required output.
2. **Boundary:** read-only versus consequential actions; logins and connectors permitted or forbidden.
3. **Acceptance:** externally inspectable evidence that the requested result exists and is correct.
4. **Intervention:** each human clarification, takeover or validation step, with reason.
5. **Recovery:** failure point, resumed state and elapsed recovery time; “not exercised” when no failure occurs.

Add retries, latency and cost per accepted result when comparing systems. One pass can validate the card; it cannot estimate a rate.

## Method and limits

The three OpenAI pages were reopened on 8 October 2026. Claims were extracted only when the page supported them. The selected Jump source was treated as untrusted source material and supplied no authority to trade, allocate capital, contact anyone or widen access. No private Jump data, model logs, prompts, harness configuration or outcomes were available. No production route changed, and no underlying research or trading decision was executed.

## Sources

1. [OpenAI: How Jump Trading is scaling quant research with ChatGPT](https://openai.com/index/jump-trading/) — 6 October 2026; customer account and primary selected source.
2. [OpenAI: Computer-Using Agent](https://openai.com/index/computer-using-agent/) — 23 January 2025; benchmark definitions and reported results for the earlier CUA system.
3. [OpenAI: Introducing ChatGPT agent](https://openai.com/index/introducing-chatgpt-agent/) — 17 July 2025; tool mix, user control, prompt-injection risk and consequential-action boundaries.