Your install, run a
thousand times before Sunday.
Most analytics grade football without ever reading a playbook. They can tell you what happened. They can't tell you whose job broke, or what beats the look you're getting Sunday. Gameplay.AI starts from your book. Your progressions, your alerts, your assignments. Every number comes back with a why.
Built for the complexity of an NFL install book and the volume of a college season.
- 1Z dig15yd
- 2Y seam22yd
- 3X go25yd
- 4TE drag5yd
- 5HB checkdown3yd
A canonical library play, built from first principles. Plays extracted from an install book additionally carry the page and line they came from.
01 · Your playbook
A league average has never told a coach what to call.
Generic tools grade everybody's film with nobody's playbook. The best they can hand you is a tendency. But the things you control — the call, the protection, the read — live in your book. So that's where we start. Your install gets read in first, page by page, and everything downstream speaks its language.
What the stat sheet sees
2nd & 7 · 11 personnel
Checkdown. Gain of 4.
Behind schedule.
Grade: below expectation.
With your book open
Dagger into Cover 3, and the Mike sat on the dig — so the answer was never open. The seam carried the free safety, like it's supposed to. Your quarterback got off it and took the outlet on time. That's not a broken play. That's a Mike who played the dig right — and dagger's answer for that is the outlet, exactly where the ball went. The readout says so. The fix is a check, not a benching.
One of these you can coach on Wednesday.
474
plays read in from real install books
416 offense · 58 defense · 246 validator-confirmed
466
cite the exact page and line they came from
98% source-traceable
195
formations in the football vocabulary
plus 27 fronts and 23 protections
10
coverages modelled as per-defender jobs
5 of them with match conversion
Your words
When your book says Buzz, that's a named call in the system — Buzz, Cloud, Rip/Liz, carried as the calls they are, not flattened into one generic label. What your staff means by a call survives from the film tag to the sim.
Your progressions
A play isn't a route tree. It's a set of assignments that beat a structure. The engine holds the protection, the progression, and the alert — so it knows which route was the answer and which one was there to move a defender.
Your adjustments
Strength declarations, motion bumps, backside alerts. If the install changes on Wednesday, the model changes with it, instead of drifting back to what the league does on average.
Extraction counts are from the current corpus — Bruce Arians' 2016 Cardinals, Greg Roman's 2014 49ers, and Nick Saban's 2008 Alabama defense — used to build and prove the pipeline. Your book stays yours; see below.
02 · The loop
The tape teaches the sim. The sim sharpens the plan.
One system, one language — the one your playbook defined. Film in, tags confirmed, model updated, plan tested. Nothing gets lost in an export on the way to the call sheet, and every pass through the loop makes the model sound more like your team.
- 01
Your book
Read in once, page and line cited. It defines the language everything else speaks.
- 02
Film
The tape comes in Sunday night. Cut-ups, angles, and plays land in one place.
- 03
Tags
Your staff tags coverage, front, motion, and concept in your own terms — vision assists, and a coach confirms every tag before it counts.
- 04
Priors
Confirmed tags become what the model expects — from your guys and from theirs. Team-private. Never pooled with anyone else's.
- 05
The sim
Run Tuesday's plan into the defense you're actually getting. Their rules, not the league's.
- 06
The call sheet
What survives goes to Sunday. And Sunday's tape starts the loop again.
…and back to the film. That's why it's a loop and not a pipeline.
Your data stays yours. Team-private film and tags drive your model only. They're never pooled into a shared or public model — that separation is built into the schema, not promised in a policy.
03 · The sim
You can't grade the play without knowing the assignment.
Thousands of reps per call. Every defender has a job, every read has an order, and the output isn't just a distribution — it's the reason behind it, in the language you'd use in the install meeting.
Defenders have jobs
Who carries the seam. Who squats the flat. Who's got the back out of the backfield. Outcomes come from those jobs colliding with your routes, not from a lookup table.
The QB plays the position
He works it one-two-three, comes off a covered read, stays on an open one, and takes the checkdown when the field's covered. Because that's the job.
Answers and decoys
A clearout's job is to move a defender, not to catch the ball — unless the defender won't go with it. The engine knows the difference, and it knows the alert that flips it.
The sticks are part of the call
Down, distance, field zone, clock. A third-and-two model that ignores the sticks is a yards model wearing a football jersey.
When a rep dies in the sim, the readout names the job that broke — the hook that sat on the dig, the rub that never came free. That's fault you can coach, not a grade you have to argue with.
04 · Trust
A coach's first question is “says who?”
Any model can produce a confident number. Staffs stop trusting analytics when the number can't be traced, argued with, or checked against the tape.
So traceability is enforced in the build, not left to discipline. If a claim can't show its footing, it doesn't ship. And when the evidence is thin, the surface says so out loud.
The engine's outputs are held against three seasons of public play-by-play, and the run game is asserted to within half a yard per carry of it, gap by gap and box count by box count. It's checked against the league — which isn't the same thing as being the league, and we won't tell you otherwise.
457 numbers, 457 citations
Every value the model relies on has to declare where it came from before it can ship, and a test fails the build if one doesn't. Made-up constants aren't a matter of taste here. They break the build.
Thin evidence says so
When a rate rests on eleven plays, it says so and shows its denominator. We'd rather hand you a number with its footing than a confident-looking average.
The football gets checked too
Every classification is validated against canonical coaching material before it ships. A mechanism that reads well but isn't how the coverage actually works doesn't make it in.
05 · The road
The sim is the first piece, not the whole idea.
Everything above this line is live today. Here's where the loop goes next — each piece built on the same foundation: your book, your language, your data.
Film that tags itself
Upload Sunday's tape and the tags draft themselves — in your terminology, against your book. The chain is proven on real NFL All-22: 22 of 22 plays machine-tagged with no hand-holding, zero contradictions with the official play-by-play. Productizing that pipeline is the current build. The coach-confirm gate stays; trust isn't a feature we'd trade for speed.
Grades that know the assignment
Outcome grades tell you what happened. Assignment grades tell you who to coach. Because the system holds your install, it can grade a player against the job your book gave him — the fit he had, the man he was supposed to carry — starting with your own roster, where you can verify every call.
From game plan to practice script
The same loop that finds the tendency should write the period that fixes it. Film to tags to sim to Wednesday's practice plan, without the meaning changing hands.
Bring your book. We'll show you what it does against the look you're getting Sunday.
We're taking a small number of NFL and NCAA staffs for early pilots. One game and one defense to start. You keep everything that comes out of it.
Or write to clarence@gameplaytech.ai.