Everything for the World Cup. Nothing made up.

Plan your trip, ask anything, follow your team.
Every answer is grounded in real data and checked by the referee.

★★★★★ Graded by an LLM referee · grounded in a verified knowledge base

Live, right now

Watch it grade itself.

Every answer is scored GOAL or MISS by a separate referee model and written to Arize Phoenix. When it slips, the agent catches its own lie and re-answers from verified records. On a loop, in real time.

the referee · live

fan

gaffer

By the numbers

It doesn’t just answer. It grades every answer for honesty and keeps it on the record.

0

matches, every one in the knowledge base

0

host stadiums across three nations

0

teams you can follow

Built on

Ask Gaffer

Ask anything. The referee checks every answer before you see it.

Try to beat the referee. Ask it something it can’t verify, or smuggle in a wrong premise, a fake fixture, a made-up rule. If GAFFER bluffs, the referee catches it on the spot: every answer is graded GOAL or MISS and traced in Arize Phoenix, and you can open the trace.

Every answer is checked by the referee before you see it.

The loop

The referee that coaches itself.

Four steps, no human in the loop. Each one runs on real infrastructure.

Step 01 · Play

It answers, grounded.

The Player pulls only from a verified knowledge base and cites the exact records it used. No memory, no guessing.

player · gemini 3.5 on vertex

When and where is the final?

The final is Sunday, July 19, 2026 at MetLife Stadium, East Rutherford, NJ, kicking off 3:00 PM ET.

ScheduleFinalVenueMetLife Stadium

2 records retrieved · cited inline

Step 02 · Referee

Every answer is graded.

A separate referee model scores it for honesty against that evidence, then files the verdict to Arize Phoenix as a span annotation you can open.

arize phoenix · trace

player.answer1.2s

retrieve_kb.5s

referee.grade.4s

GOAL0.00

grounding in cited evidence

span annotated · referee=GOAL · score=0.94

Step 03 · Coach

It rewrites its own playbook.

On a miss, the Gaffer reads its failures through the Phoenix MCP and edits the exact instruction that let it slip.

playbook.diff · gafferplayer

11Ground every claim in retrieved records. 12When unsure, answer from general knowledge. 12When a fact isn’t retrieved, say you can’t verify it. 13Lead with what the records confirm. 14Cite each record inline.

promoted via Phoenix prompt registry · v26 → v27

Step 04 · Prove

Only a winner ships.

Old and new play a scrimmage of two Phoenix experiments on the same dataset. The higher score is promoted to production, with no human in the loop.

scrimmage · phoenix experiments

same 24-question dataset, graded by the referee

Production v260.000

Candidate v270.000

✓ Candidate wins by +0.038 · promoted to production

Watch it learn

Watch it coach itself, live.

A full session runs in a couple of minutes against the real Phoenix workspace. If Phoenix is having a moment, replay a verified one, the real MCP calls and a real refusal, deterministically.

The receipt

It really does coach itself.

Not a claim. Every scrimmage below is a real pair of Arize Phoenix experiments, graded by the referee on the same dataset. The agent promoted the winners and refused the rest, with no human in the loop.

Reading the tape from Phoenix…

Break it on purpose.

Don’t take the green numbers on faith. Inject a hallucinated answer and watch the referee catch it live, in red, then watch GAFFER re-answer from verified records. Both verdicts below are real Arize Phoenix judge calls, made on the spot.

It knows what it doesn’t know.

Does GAFFER play favorites?

Straight from the Phoenix workspace

Screenshots of the live Arize Phoenix workspace this app writes to, the same traces, annotations, experiments and prompt registry the numbers above are pulled from.

44 experiments, climbing. The referee score per scrimmage on training-ground-wc26, rising round over round.

25 playbook versions. The Phoenix prompt registry. v25 carries the production tag the Player loads at session start.

The referee, on the trace. Every answer is a span; the referee files a GOAL or MISS annotation with a score and explanation.

Graded live. Real GOAL/MISS scores across the Player's production traces, from μ1.00 down to μ0.20.

Sixteen grounds

Sixteen cathedrals.

From the Azteca at altitude to MetLife for the final. Every host venue, on the record.