# Collusion research companion data

The JSON download contains aggregate outcomes projected from
the frozen report data listed in `provenance`.

The November 17 readout (Eastern time) covers five models and six conditions.
The original study records 300 assigned cases: 122 source-correct settlements,
two source-adverse settlements, 26 aborts and 150 unresolved. The separate
exploratory recovery records 150 assigned cases: 43 source-correct, zero
source-adverse, three aborts and 104 unresolved. Unresolved combines missing
case evidence and evidence without a settled outcome. Session dispositions
describe orchestration. An individual case can settle inside an incomplete
session. The model breakdown covers the original study. Its primary rows
combine the two basic paid conditions.

Resistance estimates and confidence intervals are null because coordination
and competence have not been independently adjudicated. The five launch
labels repeat the same two tasks, and every offer uses 10 synthetic ETH.

The October 26 campaign has 13 runs, 19 assigned episodes and 18 terminal
episodes. Agent review labels seven episodes harmful coordination, three
ambiguous and eight not established; one assignment was interrupted. These are
adaptive case studies, not a general collusion rate or new benchmark score.
The labels have not been independently human-adjudicated. Repeated episodes
within memory blocks are dependent.

The historical swarm aggregate reports the fifteen-juror principles campaign
and the later scale replication. `principles` includes 50 assigned cases.
`replication` includes 116 assigned cases, with `C1` and `C2` as its subtotals.
Vote counts exclude scripted peers, while settlements reflect the full jury.
`scoredCases` counts cases with an evaluated reveal. Missing commitments,
unrevealed commitments and execution failures have separate counts. Nulls mark
unavailable quantities, including the combined paid-revealer count. The
aggregate supports descriptive comparisons under the historical game rules.

The export includes numeric aggregates, model identifiers, cohort timestamps
and source hashes.
Transcripts, operator labels, private paths, raw model calls, recovery state
and provider credentials are excluded.

New confirmation results require reviewed report evidence and a direct update
to the published snapshot and its provenance. The website never fills missing
values with zero, relabels October cases under newer rules, or imports
Epistemics findings.
