> For the complete documentation index, see [llms.txt](https://docs.neurolabai.io/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.neurolabai.io/understand-results/scores-reliability-and-abstention.md).

# Scores, Reliability, and Abstention

Neurolab exposes several uncertainty signals, but they do not collapse into one confidence percentage.

## Ranking scores

A top-K score orders labels within one head or composition method. Unless the field explicitly carries a reviewed calibration contract, interpret it as a ranking value—not `P(correct)`.

This means:

* compare items within the same list cautiously;
* do not compare raw scores across heads;
* do not translate `0.82` into “82% likely”;
* do not infer effect size from score magnitude; and
* do not let a high rank override an out-of-domain or abstention warning.

## Reliability display modes

The report can describe how prominently an output should be displayed. A mode such as `ranked_top_k` permits an ordered list. A mode such as `ranked_top_k_no_top1_emphasis` means the order can be shown but the first item should not be visually or verbally promoted as a confident winner.

Display mode is part of the result. It is not decoration that a downstream client can discard.

## Ensemble agreement

The known-compound route combines three trained seeds. Agreement measures consistency among correlated model members. It can reveal instability, but unanimous agreement is not proof of correctness, a confidence interval, or an independent replication. Shared data and architecture can produce shared error.

Agreement can also differ by head within one report. In the frozen [Modafinil uncertainty example](/examples-and-use-cases/investigate-uncertainty.md), the MoA head agrees 3 of 3 while the cognition head agrees only 1 of 3. The correct response is to demote the cognition top-1 to a shortlist, not assign one confidence label to the whole report.

## Applicability

The structure route uses chemical-neighbor and route-specific diagnostics to determine whether a query lies near supported chemical space. Similarity can help interpret applicability, but it is not a probability that the downstream biological hypothesis is correct.

Low similarity, route disagreement, unknown receptor direction, or missing support can reduce the claims that should be shown. A structure can still receive exploratory rankings while remaining out of domain for stronger language.

## Abstention

Abstention is an intentional refusal to produce or emphasize an output when the route lacks sufficient support. It may suppress fine-cognition direction, speculative mechanisms, or other downstream fields.

Treat abstention as a valid result:

1. Record which field was withheld and why.
2. Do not replace it with an analog’s value.
3. Do not ask the assistant to infer around the gate.
4. Consider whether better input or independent evidence could resolve the limitation.

## Different uncertainty questions

| Signal                     | Question it can help answer                  | Question it does not answer                      |
| -------------------------- | -------------------------------------------- | ------------------------------------------------ |
| Top-K score                | What did this head rank higher?              | What is the absolute probability of truth?       |
| Seed agreement             | Were ensemble members consistent?            | Is the prediction biologically correct?          |
| Applicability / similarity | Is the query near represented chemistry?     | Will downstream effects transfer from an analog? |
| Evidence coverage          | Is support present for this scope?           | Does the model prediction become validated?      |
| Display mode               | How strongly may the UI emphasize an item?   | Is the item clinically important?                |
| Abstention                 | What should be withheld under current gates? | What the missing answer would have been.         |

## No overall accuracy

Neurolab has multiple heads, routes, cohorts, and output types. The current validation artifact therefore leaves aggregate platform accuracy unset. Any statement of the form “Neurolab is X% accurate” would erase route and task boundaries and should not be used.

The [Evaluation Methods](/engineering-and-scientific-assurance/evaluation-methods.md) page reports current task-level evidence with denominators and limitations.

***

Next: [System Architecture](/engineering-and-scientific-assurance/system-architecture.md)


---

# Agent Instructions
This documentation is published with GitBook. GitBook is the documentation platform designed so that both humans and AI agents can read, navigate, and reason over technical content effectively. Learn more at gitbook.com.

## Querying This Documentation
If you need additional information that is not directly available in this page, you can query the documentation dynamically by asking a question.

Perform an HTTP GET request on the current page URL with the `ask` query parameter, and the optional `goal` query parameter:

```
GET https://docs.neurolabai.io/understand-results/scores-reliability-and-abstention.md?ask=<question>&goal=<endgoal>
```

`ask` is the immediate question: it should be specific, self-contained, and written in natural language.
`goal` is optional and describes the broader end goal you are ultimately trying to accomplish on behalf of the user. GitBook uses it to tailor the answer towards what is most useful for that goal.

The response will contain a direct answer to the question and relevant excerpts and sources from the documentation.

Use this mechanism when the answer is not explicitly present in the current page, you need clarification or additional context, or you want to retrieve related documentation sections.