Studio

Edge · build · one surface for Sharks, Snips, modules and rosters

Your first model predicts one thing — say, which side covers the spread. Pick that, and Train fills the other picks with their best option and tests the model on every NFL game since 2019, right here on your device.

8 picks to make · Train can make them all for you · nothing is saved until you choose

Forked from the Book Disagreement lens The Studio has no template for this lens yet. The board starts blank.

Sharks (models) · Snips (lenses)

8 slots · 0 decided · 54 of 56 options live · 2 dark · re-solved just now

Context · start here 0/8 decided · 54 live · 2 dark Browse the block catalog →

Templates one click replaces the board · skipped slots get their best option · one undo

Suggestions computed from your board · each one is a button · every action is one undo

Decide Target next — cuts 88% of what is left
Estimates assume 300 records — declare your projected records so time and n use your number
Start from ATS classic — 4 slots pre-set, the rest filled with their best option

Open frontiers zero-run combos worth a look · cost uniform until the registry carries estimates

S1 ats_sideS2 exclude_backups_startingS3 crowd_eloS4 exponential_decayS5 calibrated_ensembleS6 beta 6 picks · no prior record on any of them · lift unknown until a run lands
S1 ats_sideS2 exclude_backups_startingS3 crowd_eloS4 exponential_decayS5 calibrated_ensembleS6 isotonic 6 picks · no prior record on any of them · lift unknown until a run lands

3 more sampled · each Try is one undoable step

Remix start from an existing build, then change any pick Fork from artifact — paste a typed artifact's JSON

Honesty laws in force

? · what this block is

honesty laws

What

HARD_LAW rules the lattice carries right now -- what can never be negotiated away, however you build.

Why it is here

A wall the builder hits from a HARD_LAW is not a preference call; naming it here means no undo gets spent guessing which rule it was.

Example (illustrative rows, not measured)

id	why
law.graded-at-forbidden	graded_at is settlement data and never a usable feature or axis

Honesty

Every row is a real HARD_LAW rule from lawsInForce() over LatticeState.rules -- id and why come from the engine verbatim, never rewritten here. The example row below is illustrative, not measured.

Source

CollapseStudio.svelte lawsInForce() over LatticeState.rules

Checked live

law_ids_non_emptylaw_ids_uniquelaw_why_non_empty

law.graded-at-forbidden graded_at is settlement data and never a usable feature or axis
law.market-prices-pregame-provenance closing line is future data for a pregame target

Shark · model lattice S1–S8 core · one pick per slot, Features (S3) and Transforms (S4) many-pick · click a slot to focus it

? · what this block is

the board

What

The board: the slots you collapse, how many are decided, how many options are still live, and whether the build can train.

Why it is here

One place that answers what is decided, what is still possible, and what unlocks Train.

Example (illustrative rows, not measured)

slot	decided	live options	open
S1 Target	ats_side	1 of 8	no
S3 Features	—	14 of 14	yes

Honesty

Every count is read off the lattice state that the cards render; the Train hero is enabled exactly when no populated slot is open. Nothing here is a measurement of model quality.

Source

LatticeState (wfc-engine) via CollapseStudio.svelte openSlots/decidedSlotCount

Checked live

decided_within_slotslive_within_totalopen_within_slotstrain_ready_means_no_open_slot

Target S1 · pick one What answer do you want? Work backwards from it. 8 live · 0 dark

Focused slot

S1
? · what this block is

the focused slot

What

The focused slot: one card per registered module for that slot, live or dark, with its plain-language copy, the state the rules gave it, the knobs it carries and its how-it-works script.

Why it is here

A pick is only honest when you can see every option it was chosen against and why the dark ones went dark.

Example (illustrative rows, not measured)

module_id slot state why_dark
ats_side S1 live —
pitcher_mix S3 dark HARD_DATA · no pitcher rows in an NFL build

Honesty

Ranking uses the shrunk prior mean and its interval; popularity brightens a card but never ranks it; a module with no measured prior says so on the card. Example rows below are illustrative.

Source

`module_registry rows (module-registry-seed) + wfc-engine propagate() prune traces`

Checked live

`slot_has_registry_rowslive_ids_are_rows_of_the_slotranked_entries_are_liveranked_scores_descending`

module_id	slot	state	why_dark
ats_side	S1	live	—
pitcher_mix	S3	dark	HARD_DATA · no pitcher rows in an NFL build

Choose any module.

Read-out · what your picks do 8 open · not trainable yet · records ≈ 300 assumed

Estimates · time, cost, records

Trains in — pick a model family (S5) to estimate training time

Cost $0 free lanes only — this device or Kaggle CPU, never a paid GPU

Records ≈ 300 assumed until you declare projected records (Quant · Projected n)

Lands around · ATS Prior evidence as of Sep 1, 11:47 AM ET collapse a slot to project

Consequences

Decided slots
? · what this block is

decided slots

What

Every slot you have decided, in order, with the module it holds and what that pick did to the rest of the board.

Why it is here

The board re-solves after each pick; this rail is the receipt of what you asked for and what it cost.

Example (illustrative rows, not measured)

slot module_id options_pruned_by_this_pick
S1 ats_side 6
S3 epa_splits 2

Honesty

Counts come from the lattice observations returned by observe(), never from a cached page state; 0 decided means exactly that. Example rows below are illustrative.

Source

`LatticeState.observations + prune_trace (wfc-engine)`

Checked live

`one_decision_per_single_pick_slotdecided_modules_are_registry_rows`

0 decided

No slots decided yet.

Rules doing work right now

? · what this block is

rule ledger

What

Which rules are pruning options right now, and how many, highest first.

Why it is here

A build that looks stuck is usually one rule doing all the work -- this card names it instead of leaving the builder to guess from the board alone.

Example (illustrative rows, not measured)

rule_id	pruned_count
law.graded-at-forbidden	3

Honesty

Every row is a live count from rulesDoingWork() over LatticeState.domains[*].prune_trace -- nothing here is estimated or cached across builds. The example row below is illustrative, not measured.

Source

CollapseStudio.svelte rulesDoingWork() over LatticeState.domains[*].prune_trace

Checked live

rule_ids_uniquepruned_counts_positivesorted_pruned_count_desc_then_rule_id_asc

physics.sport-compatibility2

Experiments frontier

accuracy vs. cost, compatible runs only

0 of 0 compatible

No experiments recorded for this lattice yet.

Typed artifact

{
  "kind": "model",
  "config_hash": "a89aca28",
  "as_of": "2026-09-23T05:52:16.708Z",
  "quant": null,
  "projected_sample_size_records": null,
  "module_parameters": {},
  "decided": [],
  "open_slots": [
    {
      "slot": "S1",
      "live": 8,
      "registry": 8
    },
    {
      "slot": "S2",
      "live": 12,
      "registry": 12
    },
    {
      "slot": "S3",
      "live": 12,
      "registry": 14
    },
    {
      "slot": "S4",
      "live": 7,
      "registry": 7
    },
    {
      "slot": "S5",
      "live": 7,
      "registry": 7
    },
    {
      "slot": "S6",
      "live": 3,
      "registry": 3
    },
    {
      "slot": "S7",
      "live": 2,
      "registry": 2
    },
    {
      "slot": "S8",
      "live": 3,
      "registry": 3
    }
  ]
}

Name this model

Publish

Publish gates

All core slots decided 8 of 8 core slot(s) undecided: S1, S2, S3, S4, S5, S6, S7, S8
No slot domain went empty
Target sport recorded on S1 S1 (the target) is undecided, so no sport has been recorded
Forward ATS record pending — measured after publish: needs 20 graded forward picks (ruling k10)
CLV correlation pending — measured after publish: needs graded wagers priced against a closing line
Training run complete no training run recorded yet

Model card

Sport: —

No slots decided yet.

1,000 picks

10 games in the next 10 days · first kick Sep 25 · 102 models · priced once at pick time · lines moved 0.00 pt on average since (237 picks with a current line)

as of Sep 23

? · what this block is

this run

What

Watch your picks train: the pipeline your own picks make, walked live (a sketch until the training rail is wired), then the results screen.

Why it is here

The board is a program; seeing it run is how a collapse becomes a model card.

Example (illustrative rows, not measured)

stage	state	ticker
data	done	data · ATS side pick
features	done	features · 2 feature feeds
model	live	model · L2 logistic regression

Honesty

A sketch run walks the stages and folds without scoring anything — every number stays 0 or "not recorded"; the results screen prints only what public.experiments recorded, and shows "not emitted by this run" for feature contributions, the calibration curve, the forward ATS record and the CLV correlation until a trainer actually produces them.

Source

collapse-training stagesFromDecisions (sketch) / run-adapter reduceTrainingEvents (real run) / feed.ts feedFromExperiment (finished)

Checked live

active_stage_in_rangefolds_scored_within_totalloss_epochs_consistentfinished_metrics_finite_or_nullfinished_publish_readiness_finite_or_null

no run yet — decide the board, then Train model runs it on this device

? · what this block is

this code

What

The code you are building: a python or tf.js sketch generated from your decided picks, one authored body per module with the line saying what it assumes.

Why it is here

The lattice IS a program; reading it as code is the fastest audit of what a collapse would train.

Example (illustrative rows, not measured)

line	kind	text
0	comment	# collapse studio sketch — sketch, not a runnable pipeline
5	code	build_model("gbm") # Gradient boosted trees
6	body	model = GradientBoostingClassifier(max_depth=3, …).fit(X, y)

Honesty

Display only — never executed, never round-tripped through the engine. Bodies are authored sketches over invented, non-importable helpers; a family with no honest tf.js port says so instead of faking one. A hand edit decouples the caption from the config hash until the next collapse regenerates the sketch; edited text is still never run.

Source

collapse-codegen generate.ts over LatticeState.observations + MODULE_BODIES

Checked live

honesty_line_firstno_empty_sketchbody_follows_codebody_run_ends_in_honestynot_decided_is_placeholderdraft_is_a_real_edit

Codegen

The code you are building

display only — never executed, never round-tripped through the engine

Nothing decided yet — the sketch fills in as you collapse slots.

0/8
54 live

Unlock more Studio tools

Keep exploring for free, or unlock the complete Studio toolkit with Pro.

$1 first week, then $5.99/week.

Unlock for $1