FOLDRA

Flat information in. Folded structure out.
Every answer ships with its crease pattern.

Everything your organization knows arrives flat: contracts, tickets, ledgers, transcripts — millions of true sentences with no shape. Flat data can be searched. It cannot be trusted with a decision.

FOLDRA is an inference engine built on the mathematics of folding. It scores creases in the sheet, folds along them into the structure you asked for, and records every move — so any answer unfolds back to the exact lines that made it.

Structure is a sequence of folds.

Three operations, composed in order. Nothing else ever happens to your data.


Crease

Score the sheet

The engine reads the flat stream and marks latent structure as creases: entities, claims, relations, contradictions. Nothing bends yet — a crease is a hypothesis about where the sheet wants to fold.

Fold

Collapse ambiguity

Along chosen creases, flat text collapses into the target shape: a schema, a triage queue, a close package. Every fold is one of three typed moves — valley, mountain, reverse — applied in a recorded order. There is no fourth move, and no leap.

Unfold

Audit the shape

Any output unfolds. Run the crease pattern backward and the shape flattens into the exact source lines that produced each field. If a fold is wrong, you correct the crease and refold — you never hand-edit an answer.

One answer, unfolded

bench run 0117 · target: obligation_map · synthetic corpus

One bench answer from the Contract Valley shape, replayed backward to flat. The folded field says sixty days; the base contract says thirty. The trace shows why the engine is right — and where you would look if it were wrong.

shape
notice_days: 60 · auto_renew: true — the folded field a team would act on.
move 3 · reverse
A later amendment overrides the base clause. The reverse fold keeps the newer date on top and the older one creased beneath it, still reachable. at amendment-02.pdf p.3 ¶1 · "Section 11.2 is amended to read sixty (60) days"
move 2 · mountain
Two renewal sentences collapse into one obligation. at msa-2026.pdf p.212 ¶4 · "shall renew automatically for successive one-year terms"
move 1 · valley
Crease scored on the termination window — at this point a hypothesis; nothing has bent yet. at msa-2026.pdf p.211 ¶2 · "written notice of non-renewal no fewer than thirty (30) days"
flat sheet
Three source lines, zero unplaced fragments. Nothing in the answer exists outside the sheet.
audit
Replay reached flat in 0.4 s. Verdict: flat-foldable. The sixty is not a guess — it is the amendment beating the base clause, and both lines sit one unfold away.

Shapes teams fold every day.

Five reference shapes ship with the engine. Each is a target geometry plus the crease logic that reaches it — start from one, or define your own.


Ledger Crane

Finance operations

Folds a quarter of raw transactions, invoices, and bank feeds into a close package where every figure carries its crease — the journal line, file, and timestamp it folded from. Auditors replay the pattern instead of sampling it.

target: close_package · unfolds to journal line

Intake Fan

Support & operations

Folds an unbounded ticket stream into a pleated queue: duplicates collapse into one pleat, urgency orders the folds, and each pleat opens back to every customer voice inside it.

target: triage_fan · duplicates collapse losslessly

Contract Valley

Legal & procurement

Folds four hundred pages of agreement into an obligation map. Each duty, date, and dollar sits on its own crease, one unfold away from the clause that created it — renewal cliffs included.

target: obligation_map · clause-level provenance

Signal Pleat

Research & intelligence

Folds a literature stream into a claim graph. Edges exist only where two sources actually touch, and every edge unfolds to the paragraphs on both sides — disagreement stays visible as an unclosed pleat.

target: claim_graph · edges cite both sides

Atlas Base

Data platforms

The bird base of the library: folds heterogeneous records into one governed schema that every other shape folds from. Change the base and every downstream shape refolds from the same creases — no drift, no forks.

target: atlas_schema · downstream shapes refold automatically

Built to the theorems.

Origami has laws — real ones, proved. We adopted them as engineering constraints, because a fold that breaks the mathematics is a fold you cannot trust.


Maekawa’s theorem

At every flat vertex, mountain and valley creases differ by exactly two.

In the engine: every decision point must balance its evidence the same way. A fold whose support does not balance is rejected before it bends the sheet — imbalance is detected as geometry, not reviewed as prose.

Kawasaki’s theorem

Around every flat vertex, alternating angles sum to a straight line.

In the engine: before folding, the readings around a decision must close the circle — exhaustive, exclusive, complete. If the alternatives do not sum, the crease stays a hypothesis and the sheet stays flat.

Flat-foldability

A crease pattern is only valid if the sheet can return to flat.

In the engine: no output ships unless it fully unfolds to its sources. An answer that cannot be flattened back to evidence is treated as a defect, not a result — the server refuses it before you see it.

One sheet, no cuts

Classical origami adds nothing and removes nothing: one square, folded.

In the engine: outputs contain only folded input. If a fact is not somewhere in your sheet, it cannot appear in your shape — the engine has no scissors and no glue. What it cannot fold, it returns flat and says so.

What we can honestly claim. FOLDRA is a concept study — a fictional company built to argue a real standard. The theorems above are genuine mathematics from the flat-folding literature; the argument is that machine reasoning should be held to the same test: if it will not unfold, it does not ship.

A fold is one call.

The API has a single verb. Everything else is geometry you declare, not behavior you hope for.


Give it the sheet

Point the engine at flat sources — files, streams, tables. It concatenates them into one sheet and alters nothing. The sheet is immutable for the life of the fold.

Name the target shape

Pick a reference shape or define a geometry of your own. The engine proposes creases and shows you the full pattern before anything bends.

Ship flat-foldable

Fold. The response carries the shape and its complete crease pattern. With flat_foldable: true, any response that cannot unfold is refused server-side — you never receive it.

POST /v1/fold

{
  "sheet":  ["s3://ops/tickets/2026-08/*.json"],
  "target": "triage_fan",
  "moves":  ["valley", "mountain", "reverse"],
  "flat_foldable": true
}

200 OK
{
  "shape": { "pleats": 41, "open": 7 },
  "crease_pattern": [
    { "op": "valley",  "at": "tickets/8812.json:14" },
    { "op": "valley",  "at": "tickets/8813.json:02" },
    { "op": "reverse", "at": "tickets/8812.json:31" }
  ]
}

// replay forward to rebuild the answer.
// replay backward to reach flat.

Fold something with us.

Access opens in small batches while the engine is in study. Write one line about your flattest data — the pile nobody can make decisions from — and we will read it.

hello@foldra.ai

FOLDRA is a design fiction — a concept studio piece, not a live service. The mathematics is real. How this page was built, crease by crease, is in the field guide.