# An agent finding you can check

This case asks a narrow question: did NASA GISS report a 2020 global surface temperature anomaly of at least **+1.00 °C** relative to the 1951–1980 mean? The retained [NASA Earth Observatory page extraction](source-web-retrieval.txt) includes its **+1.02 °C** statement. A Synatrail research agent follows a source lead, reads a selected excerpt, proposes a cited finding, and asks for a numerical test. Synatrail rejects a fabricated quote, compiles the plan, computes the margin, and derives the result **met**. The Agent finding remains pending for a person to review.

![Native Synatrail card rendering of the question, source, pending finding, test and review decision](canvas.png)

The source path has two deliberately distinct records. The official [NASA article](https://science.nasa.gov/earth/earth-observatory/2020-tied-for-warmest-year-on-record-147794/) was fetched through the web tool on 5 October 2026; the retained [web extraction](source-web-retrieval.txt) identifies the article, line 268 and its extracted text. This is an actual fetched page extraction, **not raw HTTP response bytes**. The local shell could not resolve NASA's host, so the platform `read_source` step used [source-response.txt](source-response.txt), a selected excerpt whose quote is checked against the independently fetched page. The case saves the replay's original bytes, SHA-256, source version, passage IDs, read coverage and date in [source-receipt.json](source-receipt.json). The selected excerpt is shorter than the full article. Re-running this case does not fetch the current NASA page.

The agent calls are a deterministic OpenAI Responses **protocol fixture**. They pass through Synatrail's actual `callResearchProvider` dispatch, which checks the allowed tool, emits a function-body authority, and parses each function call. There are six recorded calls: `search_public`, `read_source`, `propose_finding`, `assess_evidence_passages`, `compile_test_plan`, and `finish_research`. No paid model call is made, and the script does not claim the fixture independently discovered the source or chose the claim. Its purpose is to make the platform's research and verification path reproducible without a credential. Prompt hashes, selected tools, source read and results appear in [investigation.json](investigation.json).

The agent's pending finding cites `sourceVersionId:passageId`, so a reader can return to the exact retained text. The evidence verifier accepts one copied phrase from that passage and drops one invented sentence with `not-verbatim`. The accepted quote supports the reported number; it does not by itself settle scientific uncertainty or prove causation. A person has not accepted the finding onto My canvas.

The test compiler checks the metric, units, threshold, sample rule, falsifying observation, parameter bindings, and deterministic tool configuration. The compute engine runs a one-scenario sweep for **1.02 − 1.00 = +0.02**. Temperature anomalies are differences, so the computation uses kelvin intervals; 0.02 K is the same interval as 0.02 °C. Synatrail's result path extracts the computed metric and sample count, then `cleanResult` derives `observedVerdict: "met"` against the registered `gte 0` threshold. [test-receipt.json](test-receipt.json) contains the compiled plan, evidence check, computation hashes, metric, sample and platform-produced result. This verifies only the arithmetic threshold for the retained source excerpt.

The [canvas image](canvas.png) is rendered with Synatrail's actual `ResearchNode`, `Relationship` and React Flow components from the recorded case data. It is an authored presentation of the run, not a live account screenshot. The [source inspector image](source-inspector.png) is a separate capture of the current app: an isolated signed-in board ingests the same retained excerpt as a text file and opens its real source viewer. That reproduction has its own [capture receipt](inspector-capture.json); it is not a claim that the agent research receipt created that temporary board.

To reproduce from the repository root:

```sh
node scripts/cookbook-agent-evidence.mjs
```

The script writes only this case folder. It uses the repository's cached Chromium and an ephemeral `127.0.0.1` listener for the native canvas and inspector captures. The app's temporary database is removed after capture. Run from a workspace that permits local listening and has the existing Synatrail browser dependencies installed. It asserts the official web extraction contains the replayed quote; that the source reader makes exactly one bounded replay read; that the cited passage exists; that the fabricated quote is rejected; that the compiler accepts the executable test; that the compute and platform result paths produce the measured value, sample and `met` verdict; and that the native graph has no clipped text or label collisions. Browser capture checks for page errors and unexpected external requests.

The next human step is to inspect the original NASA page and [readable finding](findings.md), then accept or challenge the pending Agent finding. A fresh full data-table read and NASA's uncertainty notes remain open if the question grows beyond this one reported value.
