How it works · v1
Inspect the reasoning.
Compare exposes differences between models. Guide exposes how interpretations are constructed. Chat lets you examine both as a conversation develops.
What ASSAY is — and isn’t
ASSAY is a workspace for inspecting model reasoning: what a response says, how it relates to a source, and how it differs from another response. It is not a benchmark leaderboard or a universal ranking of models. A model’s judgment is another interpretation to inspect; the final decision stays with you.
Chat: follow a conversation across models
Choose which model responds next and switch providers within a thread. Relay context shares the conversation across models; parallel context keeps each model’s response history separate. In relay context, you can ask multiple models to respond to the same turn, inspect their alternatives, and use inline judgment before choosing a response to carry forward.
The conversation and the judgments remain available to inspect. An inline judgment can help you see a tradeoff, but it does not certify a response as correct.
Guide: read the source beside its interpretation
Supply a source text and choose a guide model, or run several models against the same source. ASSAY deterministically divides the text into paragraphs or sentences with stable segment IDs. Models receive those segments and are asked to return claims citing the IDs they rely on.
The reading view keeps the original text beside those claims. Select a citation to inspect the passage behind it. You can compare guides on the same source to see where their readings converge or diverge. A guide offers suggested paths through a document; it is not a replacement for the document.
Compare: make differences inspectable
Send the same prompt to selected models and read their responses side by side. Choose a judgment mode and, where useful, provide criteria for the task. Structured judgments help you inspect shared ground, key differences, tradeoffs, and fit for those criteria while keeping the underlying responses visible.
A result describes this prompt, these responses, and these criteria. It is not evidence that one model is best for every task, and agreement between models is not proof of truth.
What the checks tell you
Guide’s citation contract is checked by code. If a claim cites an ID that does not exist in the source segmentation, ASSAY records a fabrication flag and marks the dangling citation as invalid. Quoted excerpts are also checked for a verbatim match in the cited segments. These checks establish that a reference or excerpt exists; they do not establish that the passage supports the claim or that an interpretation is sound.
- Noted Gaps
- The guide model’s own observations about what the source leaves out or does not substantiate. They are separate from cited claims and from computed coverage. Code checks the shape of these observations, not whether they are correct. In the current reading view, this band appears when the model reports gaps; an empty list does not prove there are no omissions.
- Not Reflected
- A code-computed list of source segments that no claim cites. This band stays visible even when empty, when it states that every segment was cited by at least one claim. Citation coverage does not establish completeness or accuracy.
Omissions and invalid citations are part of the reading, not something to hide behind a polished summary. The checks give you specific things to inspect against the source.
Bring your own API keys
Users bring their own provider API keys and configure them in the workspace. ASSAY doesn’t sell model access. Usage and any resulting charges belong to your accounts with those providers.
Session boundaries
There is no persistent cross-session memory in v1. Sessions are self-contained: a new session does not automatically inherit facts, preferences, or conclusions from earlier sessions. Saved history lets you revisit work; it is not memory carried into future model calls.