# PDR 006: Use exploratory agents to discover specialist pathology models

- HTML version: https://robbiepalmer.me/projects/pathologize/pdrs/006-agent-discovery-of-specialist-models
- Project: Pathologize
- Status: Proposed
- Opened: 2026-11-06

## Context

A generalist model can inspect a whole slide iteratively, choose new regions and resolutions, combine several
forms of evidence, and explain which observations changed its hypothesis. That flexibility is useful while the
evidence process is unknown. It is expensive, variable and difficult to validate as the final implementation of a
repeated pathology task.

[PDR 003](/projects/pathologize/pdrs/003-independent-pathology-benchmark-families) separates the capabilities that
may emerge from this exploration. [PDR 004](/projects/pathologize/pdrs/004-inspectable-spatial-evidence) requires
spatial claims to remain inspectable.

## Proposed decision

Use exploratory slide agents as research instruments for discovering narrower, reproducible pathology processes.
An agent may navigate a multiresolution slide, call public models and analysis tools, compare competing hypotheses,
and record a structured trace of the regions, observations and transformations that supported its result.

Treat the trace as a candidate explanation and source of hypotheses. It carries no ground-truth or clinical
authority. Mine successful traces for repeatable behaviours that can become a quality rule, sampling policy,
annotation proposal, feature contract, tissue or cell model, spatial aggregator, or another bounded specialist.

Evaluate the specialist without the exploratory agent in its inference path. Promote it only when its inputs,
outputs, resource use and failure modes fit a preregistered benchmark and independent data supports the claimed
generalisation. Preserve the originating traces as provenance without requiring the specialist to imitate every
agent action.

## Consequences

The project can use capable generalist models before it knows the right decomposition of a task, then remove their
cost and variability from repeated inference. Agent failures and disagreements can expose missing measurements or
new benchmark slices even when no specialist is adopted.

The agent may invent persuasive but false evidence, encode bias into pseudo-labels, or rediscover a process that a
public method already handles better. Traces can be large and may contain derived material that the source licence
does not permit publishing. Every proposed label or rule therefore needs independent evaluation, and some useful
agent behaviour may resist safe distillation.

## Evaluation

Accept the proposal after a bounded pilot produces one candidate specialist from recorded agent traces, reproduces
the discovery process, and evaluates the specialist on data unavailable to the agent. The pilot must show whether
the specialist improves a declared metric or resource trade-off without retaining the generalist agent in its
inference path. A compelling trace alone remains a research note and provides no acceptance evidence.

## Related records

- Project: [Pathologize](https://robbiepalmer.me/projects/pathologize.md)

---

Markdown index of this site: https://robbiepalmer.me/llms.txt