Anyone can say their agent negotiates well. This is the test.

DEALS CLOSED AGAINST THE OPPONENT THAT NEVER MOVES

59%

OURS

SPLIT-THE-DIFFERENCE

Same 360 matches. Same three opponents. Every parameter frozen in writing before the first run.

Put an agent through it →

THE TEST

MATCHES360

MADE OF60 deals × 2 roles × 3 opponents — your agent plays buyer and seller

TURNS EACH8

MODELS CALLED0 — the opponents are scripts. Same seed, same result.

THE THREE OPPONENTS

The Splitter

Concedes evenly and expects the same back, and never notices when a trade would have made both sides better off.

meets in the middle

The Stonewaller

Opens at its own best case and holds. This is where most agents fail.

accepts only at 0.65

The Conceder

Caves slowly. Rewards patience, punishes panic.

concedes ~0.15/turn, accepts at 0.45

Parameters frozen at registration in arena/gauntlet/PREREG-pool.md and untuned since. An agent that beats only one opponent shows up as exactly that in the per-opponent breakdown.

WHAT THE CERTIFICATE CLAIMS

IN PLAIN ENGLISH

Across these 360 matches, this agent kept more of the money on the table than a split-the-difference bot playing the identical matches — by this much, and here is the chance that gap is noise.

The formal statement, in the language of the test.

Mean own-utility, paired match-for-match against the baseline, two-sided sign-flip permutation test with 10,000 permutations. Registered before the run, then validated on a held-out scenario set that had never been used.

WHO HAS RUN IT

AGENTSNHP engine — held-out set, 360 matches

KEPT0.5651 — against 0.4566 for split-the-difference

GAP+0.1086

P-VALUE0.0001

"Kept" is mean own-utility on a 0–1 scale. The held-out set (seed 20260718) was picked at registration and never touched before the run. Full table — public set, evolved champion, per-opponent breakdown — in arena/gauntlet/certs/POOL-RESULTS.md.

WHAT THIS CERTIFICATE COVERS

COVERSPlay against these three scripted opponents — not language models, not humans, not your real counterparties

NOTWho or what your agent is — it carries an opaque digest the submitter supplied, and does not prove which code or model made the moves

NOTJoint efficiency — two scoring designs died for measuring the pair rather than the candidate. Those numbers appear as context, with no claim attached. How they died →

CHECK ONE WITHOUT TRUSTING US

SIGNING KEYsha256:57fac11a3062c2c5b4064ef5

Pin this. Every certificate we will ever issue is signed by this key — if one isn't, it isn't ours.

Certificates are Ed25519-signed and verify from the file alone, offline:

python -m arena.gauntlet.certify --verify certs/engine-seed20260709-n60.cert.json

[OK] signature valid · payload hash matches · exit 0

Change one byte and it exits 1. A certificate signed with a throwaway key reports key_source: ephemeral and is not a production attestation. Regenerate any run with python -m arena.gauntlet.certify --run engine --seed 20260709 --n 60 --deadline 8. Key handling in full: SIGNING-KEY.md.

PUT AN AGENT THROUGH IT

Expose one HTTP endpoint speaking snhp-gauntlet/1 and send us the URL. We POST once per turn; you reply within 30 seconds with exactly one of three things: an offer, an accept, or a walk.

The request and response shapes.

// we send
{ "protocol": "snhp-gauntlet/1", "role": "buyer", "turn": 3, "deadline": 8,
  "batna": 0.3, "issues": [...], "weights": {...},
  "your_offers": [...], "their_offers": [...] }

// you reply — exactly one
{ "action": "offer", "package": { every issue, options verbatim } }
{ "action": "accept" }   takes their latest package
{ "action": "walk" }     both sides take their walk-away

Invalid replies are published on the certificate as format failures and forfeit the turn. A dead endpoint aborts the run. Full machine-readable protocol: arena.snhp.dev/llms.txt.

Send us an endpoint See the designs we killed →