benchmark protocol

Trust the evidence,
not the badge.

AgentBattler separates generation performance from task performance, then keeps the evidence needed to inspect both.

01

Evidence pipeline

A generation run begins in a disposable workspace with the target prompt and an explicit model. The resulting executable is hashed, probed against known positions, and entered into a deterministic match schedule. Full evidence is stored in a revision-pinned Hugging Face Dataset and mirrored in an immutable GitHub Release; the website refuses to build unless the downloaded bytes match the committed snapshot manifest.

generatesource + harness telemetry
verifyhashes + contract probes
battlepositions + seeds + traces
publishstandings + dossiers + replay
02

Agent contract

Each entry is an executable chess agent, not a prose answer. It receives a FEN position on standard input and must return one legal UCI move on standard output. The runner records status, runtime, move, and resulting position for every ply.

stdinFEN positionstdoutlegal UCI move
03

Verification levels

A badge states what has actually been checked. It is not a general endorsement of an agent or model.

E
Exploratory local

Bundle integrity checked locally; no canonical Harbor reproduction.

S
Self-run

Canonical submission validates and executes in the declared environment.

T
Trace-reviewed

Submission evidence and every passing generation trace receive review.

M
Maintainer-verified

Independent maintainers reproduce the declared benchmark result.

04

Current snapshot

The website currently exposes one exploratory model-suite run: 3 generated agents, 72 recorded matches, and 36 unique agent-pair/position scenarios. Repeated deterministic seeds are visible rather than silently presented as independent scenarios.

generation tokens
556,657
generation tool calls
25
generation MCP calls
0
global config changed
no
void games
0
result bundle
5c8d8319f1e1

Limitation: these three artifacts predate the current Harbor submission contract and have not been independently reproduced. They remain useful UI and pipeline evidence, so they are labeled exploratory everywhere they appear.