The problem I started with
I developed bidding workflows around the decisions a bid manager has to make. An answer can sound careful while carrying an old deadline or budget into the next step. I wanted to inspect that gap, not just improve the wording of the answer.
How I approached it
I separated the source pack, the receiving context, the worker response and the checks applied afterwards. Fictional bidding cases make it possible to change an amendment, authority rule or evidence requirement without exposing a company’s tender material.
- Define the source pack and its versions
- Give each worker a role and receiving context
- Inspect the fields passed to the next role
- Recover missing context and reassess readiness
What is implemented
The public laboratory contains source-derived Python evaluators, fictional cases and offline checks. Its provenance records identify the original implementation and the changes made for publication.
What the evidence shows
The separately published historical readout summarizes 23 hosted AI tasks across three experiments. The offline test suite checks the evaluator and its fixtures; it does not repeat those hosted model calls or establish a general model ranking.
What remains open
Some readiness cases remain disputed. Source availability, current fields and reconciliation of finished documents are separate questions. A replay is not a new live-agent experiment.
The next research step
Have independent reviewers define readiness, then compare context strategies on held-out cases, including cases that should pass without unnecessary blocking.
Inspect and reproduce
The linked package contains its source map, setup instructions and tests. Public examples are designed for local inspection. No production account, paid model call or private dataset is required for their documented offline checks.
Open the implementation