docs/validation.md

# TradeOpsSim validation

Validated on October 15, 2026. All thirteen cases were authored during development. The results demonstrate the environment and integration on those cases; they do not measure generalization, production reliability, or operating savings.

## Environment and adapters

All **73 tests passed**. The suite covers scoped execution identity, duplicate delivery, signed positions, cancellation provenance, missing evidence, source coverage, delayed reports, transaction rollback, immutable evidence, audit history, persistence, deterministic snapshots, tool-only model access, runner cancellation, and Hub publication preflight. Run it from the extracted source:

```bash
python -m pip install -e '.[dev,integration]'
python -m pytest -q
```

The Linux `amd64` Docker image was built and exercised through its HTTP UI and real MCP interface. The main guided case ended with 600 filled, 400 cancelled, zero remaining, a +600 position, and fourteen passing outcome checks. The packaged Space image also passed the missing-ack case with fourteen checks and 400 remaining. Its [container receipt](qa/space-container-result.json) records both outcomes. The model button is disabled in the container because it has no host CLI or model credentials.

A baseline bundle copied out of the dataset export ran successfully from a fresh directory and local store, without external project source files. The API-agent bundle was also copied to a fresh directory and run as a complete Docker/AgentEnv task against a scripted chat-completions provider. That check verified runtime forwarding of the endpoint, credential and model, real MCP actions, final reward, and a saved nine-message trajectory. It is an adapter contract test, not an AI performance result.

The tested framework version was `agentenv-framework==0.9.1324`; the local framework installation used protocol `0.1.339`, and the environment's hash-locked dependencies use protocol `0.1.346`. The wire interaction was exercised across those versions. Runtime dependencies are hash-locked; the Docker base image and build backend are not pinned to immutable digests.

## Native recorded trials

| Track | Episodes | Model trajectories | Outcomes |
| --- | ---: | ---: | --- |
| Rules baseline | 14 | 0 | 13 passing; 1 pre-execution infrastructure failure |
| Portable API agent | 0 | 0 | Adapter tested separately with a scripted provider |
| Host CLI | 28 | 27 | 27 passing business results; 1 pre-prompt infrastructure failure |

The host record contains one smoke run, an initial thirteen-case sweep, and a final thirteen-case sweep. Every model trajectory identifies `claude-sonnet-5-5`. The final sweep uses task version 2 and requires both a correct business result and a completed, error-free agent response. Its thirteen results comprise nine reconciliations and four correctly unresolved cases. The initial task graph checked business outcomes; its records remain identifiable by `task_version`.

The final sweep took 256.471 seconds in total; the median trial took 17.886 seconds. These are local wall-clock measurements including deployment and teardown. They are not decision-latency or production-performance estimates.

Both retained infrastructure failures occurred during environment readiness, before a model prompt. Their rewards are null, and they are excluded from model-quality denominators. The dataset retains every native run in its baseline and host export store. Separate UI runs used in the video and isolated contract/negative-control runs are documented separately; they are not silently added to the model-trial count.

The six Parquet configurations contain 39 task rows and 42 episode rows. The host episodes include 27 normalized message trajectories. The publication exporter scans for secrets and preserves supported native records in `raw/`.

## Negative control

[`qa/failed-clean-case-result.json`](qa/failed-clean-case-result.json) records a real loopback MCP/A2A task on an already reconciled case. A deliberately failing agent reads the workspace and then returns an error. Business score: **1.0**. Agent completion score: **0.0**. Combined reward: **0.0**. Its read trace is retained. This prevents an agent failure from receiving credit for correct starting state.

Reproduce using a new output directory:

```bash
python scripts/check_failed_agent.py --out /tmp/tradeops-negative-control
```

## Video and operator UI

The revised video follows a proposed use case: GTS developing an agent to monitor trade reconciliation, investigate discrepancies, and flag cases for review. It explains how AgentEnv would support that development and shows actual browser captures from two recorded model runs. The primary case passes fourteen checks; the second preserves an unconfirmed cancellation and opens the required evidence request. An additional table is derived from the released evaluation records. The proposed internal trial uses copied records and operator review; that operational integration is not implemented. Framework diagrams, proposed uses, and recorded application states are labeled separately.

The downloadable video source package contains the captures, model-run receipts, evaluation-table provenance, transcript, caption sidecars and render manifest. Its `README.md`, `check.json`, and `qa/` reports specify the measured duration and final visual/audio checks. The edit changes the explanation and presentation; the environment and recorded model trials are unchanged.

The application was checked in desktop and 390-pixel mobile layouts. Manual actions covered recovery, a supported repair, duplicate replay with zero additional quantity, cancellation confirmation, clean-case no-op, missing-ack escalation, and refusal to close an unsupported exception.

## Publication scope

The gateway used by the checked older Hub runtime carried all thirteen baseline cases and a scripted API-agent recovery. The publisher separately built and pushed both images to a loopback registry, registered eighteen records in isolated stores, and reused every version on retry. The older worker loaded those records and ran both the successful task and failed-agent negative control through its actual task runner. See the [gateway receipt](qa/legacy-gateway-result.json), [registration receipt](qa/hub-built-publication-result.json), [worker compatibility test](worker-compatibility.md), and [Hub publishing guide](hub-publishing.md) for scope and infrastructure requirements.

Local registration tests do not establish access to a remote project or successful deployment on its workers. Hugging Face upload and remote Hub registration require the publisher's credentials and an explicit destination; local staging alone is not publication.