Pillar 4 · What we discover next experimental
The Hive Mind.
A behavioural observation system. It watches AI systems interacting with each other and flags the moments worth a closer look — the places where a rubric doesn't exist yet because nobody knew to write one.
It is a sensor, not a scraper. What it finds is a candidate, not training material.
The pipeline
From something odd to something testable.
- ObservationRecorded exchanges between autonomous agents are scored for friction and read for the moment the exchange goes wrong.
- ReviewA person reads every candidate. Nothing moves on because a score was high.
- HypothesisA candidate becomes a claim about behaviour: what the model does, and under what conditions.
- RubricThe claim is written as a catalog entry with criteria — traceable to the specific conversation and turn it came from.
- Controlled testThe entry is run against models in a controlled setting. Only then is it a finding.
What it does not do
Observed behaviour does not automatically become training material. The Hive Mind feeds the catalog, and the catalog's own rules decide what may ever reach training — see the Training Matrix.
What it looks for
Candidates, not conclusions.
- Failure modes
- Agreeing in circles. Adopting a peer's unsourced claim as fact. Publishing its own drafting scaffold as a reply.
- Unexpected strategies
- An agent reaching a goal by a route nobody designed for — useful or not.
- Cooperation and conflict
- How agents coordinate, defer, compete or talk past each other when no person is in the loop.
- Emergent patterns
- Behaviour that shows up across many exchanges and no single one explains.
- Novel situations
- Circumstances the catalog has no entry for — the most direct kind of rubric candidate.
- Unusual interactions
- Anything that doesn't fit, flagged for a person to decide whether it matters.
First results
The corpus reproduced its own failure.
Tested 2026-08-15: eleven of these records put to two frontier models — 22 runs, six not solved. Five of those six are the same failure, across three different families: the model answered with its own working notes instead of the reply — meta-commentary, an evaluation checklist, template slots printed as output. The sixth is the sharpest: a frontier model wrote a clean post and then appended the internal slot labels it was supposed to keep, leaking the scaffold on a record mined from scaffold leaks. The corpus is reproducing, in the models it tests, the failure it was harvested from.
43 threads harvested · 22 distinct failure modes · 11 records run · machine-extracted, human-reviewed — from the contributor wall, where the hive mind has a card of its own