Skip to content
DecisionNodeDecisionNde
  • Model
  • Inference
  • Benchmarks
  • Examples
  • Docs
  • Pricing
Sign inGet API keyGet API key
  • dnModelTyped answers, calibrated confidence
  • msInferenceOur own stack and GPUs, answers in ms
  • %BenchmarksAccuracy per suite, with intervals
  • exExamplesBuilds you can start from today
  • /v1DocsQuickstart, API reference, recipes
  • $PricingPay per input token, output is free
  • inSign inEmail link or Google, no password
examples

Start from a working build.

34 requests, ready to run. Every type names its answer: choice (one of your labels), score (a level on your scale), truth (a probability), number (a value on your grid), point (a spot in your image), box (a rectangle around the thing). Every card holds the exact body sent to /v1/decide and the answer it gets back, ready to open in the playground; the point and box builds also draw the answer on the frame.

Open playgroundOpen playgroundRead the docs
ready to run
34 requests
question types
6 types
images per request
up to 16
lessons

Four questions, four answers.

One lesson per question type, each a decision developers automate every day. Each asks one question and shows exactly what comes back: the fields, the numbers and how your code reads them.

comes back
choice
always one of your keys
probabilities
one per option, summing to 1
confidence
how clearly the pick won, 0 to 1
Choice

Name the commit

Give it your options. It picks exactly one and returns a probability for every option.

asked
Stop skipping the last page when the cursor lands on an exact multiple of the page size. Adds a regression test for 100 items at 25 per page.

kind: What kind of change is this commit message describing?

criteria

feat
adds a new capability
fix
corrects wrong behaviour
refactor
restructures code, behaviour unchanged
docs
documentation only
chore
tooling, dependencies or build

fixpicked at 0.71confidence 0.64

  • fix0.71
  • feat0.21
  • chore0.04
  • docs0.02
  • refactor0.02

then your code

confidence 0.64 is under 0.7, so the commit is re-checked with the full model before it is labeled.

comes back
choice
always one of your keys
probabilities
one per option, summing to 1
confidence
how clearly the pick won, 0 to 1
Open in playground: Choice: name the commit
Score

Bug severity

Give it an ordered scale. It returns the expected level and how the probability spreads over every level.

asked
Checkout fails for every customer paying with Apple Pay since this morning's release. The button spins and then shows 'payment declined'. Card payments still work.

severity: How severe is this bug report?

criteria

  1. 0cosmetic
  2. 1minor
  3. 2major
  4. 3blocker

2.60nearest level: blockerconfidence 0.60

  1. 0.00
  2. 0.04
  3. 0.31
  4. 0.65
  1. 0cosmetic
  2. 1minor
  3. 2major
  4. 3blocker
Expected score 2.60 on a scale of 0 to 3, nearest level blocker.
  1. backlog
  2. next sprint
  3. flag off
  4. roll back (what the code does with this answer)

then your code

confidence 0.60 is under 0.7, so the code takes the cautious level, blocker, and rolls the release back on its own.

comes back
score
the expected level, between two levels
probabilities
one per level, low to high
legend
level index back to your words
confidence
how settled the spread is
Open in playground: Score: how severe is this bug report
Truth

Charged twice

Is this statement true? One calibrated probability from 0.00 to 1.00 that the statement is true. 0.80 means true about 8 times in 10. Your code picks the cut-off: 0.50 for a plain yes or no, higher when acting on a false yes is costly.

asked
Hi, I see two identical charges of $49.00 for order #4471 on the same day. I only placed the order once.

charged_twice: Was the customer charged twice for the same order?

criteria

true
Charged twice for the same order
false
Any other billing issue

0.97trueprobability it is truerefunds at 0.80

falsetrue
00.500.801
What the code does with each answer:
  1. 0.00 to 0.49falseno refund, asks for receipt
  2. 0.50 to 0.79truefull model re-checks it
  3. 0.80 to 1.00truerefunded, reply sent (this answer)

then your code

0.97 clears the 0.80 threshold, so the duplicate charge is refunded and the reply goes out on its own.

comes back
truth
the probability that the statement is true, 0.00 to 1.00: the whole answer
true / false
not a field: truth against 0.50, as shown here
threshold
lives in your code; this one refunds at 0.80
Open in playground: Truth: was the customer charged twice
comes back
number
the most probable value on your grid
expected
the probability-weighted mean; it can fall between values
confidence
the probability of number, 0 to 1
probabilities
every value above 0.001, keyed by the value
Number

Count the cars

One value on a grid you set (min, max and an optional step), with a calibrated probability for every value: the most probable value, the expected value and the confidence.

asked
Drone photo looking straight down on a parking row of ten marked bays: seven cars parked (white, black, silver, navy, red, graphite and champagne) and three empty bays, with a pavement, a lamp post and a grass verge below.
Drone frame over row B4 of the north car park, camera pointing straight down.

cars: How many cars are visible in the image?

grid

  • min0
  • max50
  • step1
  • 51 values, one answer

7most probable valueconfidence 0.78expected 7.00

  1. 0.015: 0.01
  2. 0.106: 0.10
  3. 0.787: 0.78 (the answer)
  4. 0.108: 0.10
  5. 0.019: 0.01
56789
the diamond marks expected 7.00; values outside 5 to 9 are under 0.001, so the answer leaves them out050

then your code

confidence 0.78 clears 0.7, so row B4 is logged at 7 cars and the car park's free-space sign updates on its own.

comes back
number
the most probable value on your grid
expected
the probability-weighted mean; it can fall between values
confidence
the probability of number, 0 to 1
probabilities
every value above 0.001, keyed by the value
Open in playground: Number: count the cars in the frame
cameras

Watch the road, the floor and the line.

Four builds that read a camera frame and act on it: traffic, people, incidents and the sorter. Each reads up to 16 images a request, or the same camera as video frames with their timestamps.

3 questions, 1 image, one call

Junction camera that times its own signals

Count the queue, read how congested the junction is and catch wrong-way drivers, from one camera frame.

6(0.63)vehicles queued, congestion slow

input: junction camera frame

Camera view from above a junction: six cars queued behind the stop line in the eastbound lane, two cars moving west, a zebra crossing with two people waiting at its kerb, and two cars on the cross street.
signal plan keptno wrong-way alert

answers, one call

  • queueNumber, expected 6.076 (0.63)
  • congestionScore, slow1.85
  • wrong_wayTruth, alerts at 0.900.16 false

then your code

6 vehicles queued (0.63) and congestion 1.85 (slow): the signal plan stays as it is. Wrong way 0.16 false, under 0.90: no alert goes out.

With video: how many cars passed and which way they turned, from the last ten seconds of frames.

517 input tokens, $34.10 per 1M calls

Open in playground: Junction camera that times its own signals

3 questions, 1 image, one call

Concourse camera that opens the next gate

Count the people in a marked area, read how crowded it is and check the exit lane, from one ceiling camera frame.

25(0.14)people in the area, busy

input: concourse camera frame

Ceiling camera view of a station concourse: about two dozen people inside a yellow dashed waiting area, a few walking towards three ticket gates on the right, and an empty green hatched exit lane along the bottom.
two gates keptexit lane clear

answers, one call

  • peopleNumber, expected 27.7125 (0.14)
  • crowdingScore, busy1.77
  • exit_blockedTruth, announces at 0.800.00 false

then your code

25 people (0.14) in an area built for 40, crowding 1.77 (busy): two gates stay enough. Exit lane 0.00 false: it is clear, so nothing plays.

With video: how many people came in through the gate, counted over the frames of the last minute.

434 input tokens, $30.61 per 1M calls

Open in playground: Concourse camera that opens the next gate

3 questions, 1 image, one call

Night camera that raises its own alarm

Check every frame of a closed site for smoke, fire and break-ins, rate how serious it is and raise the right alarm on its own.

0.96truesmoke or fire, alarms at 0.80

input: night camera frame

Night camera view of a closed storeroom: racks of cartons on the left, a closed roll-up door at the back, and grey smoke rising from behind a pallet of cartons on the right, with an orange glow at its base.
fire alarm raisedventilation shut

answers, one call

  • smokeTruth, alarms at 0.800.96 true
  • intrusionTruth, locks down at 0.800.04 false
  • severityScore, emergency2.61

then your code

Smoke 0.96 true clears 0.80 and severity reads 2.61 (emergency): the fire alarm is raised and the ventilation shuts on its own. Intrusion 0.04 false: no lockdown.

With video: whether the smoke is spreading, read over the last eight seconds of frames.

334 input tokens, $26.41 per 1M calls

Open in playground: Night camera that raises its own alarm

3 questions, 1 image, one call

Sorter that sends every package to its lane

One decision per package at the diverter: its lane, whether it is damaged, whether its label can be read.

lane 21.00lane, acts at 0.70

input: camera frame over the diverter

Camera view from above a conveyor belt: a yellow package with a white shipping label face up inside the camera's guide marks, a brown carton arriving on the left and another leaving towards the diverter for lanes 1 and 2.
diverted to lane 2label scanned downstream

answers, one call

  • laneChoice, lane 21.00
  • damagedTruth, reworks at 0.800.01 false
  • label_readableTruth, ships at 0.800.98 true

then your code

lane 2 1.00 clears 0.7: the diverter swings and the package drops into its lane. Damaged 0.01 false and label readable 0.98 true: it ships without a stop.

With video: whether the belt stopped or jammed, read over the last few frames.

445 input tokens, $31.07 per 1M calls

Open in playground: Sorter that sends every package to its lane

On your own hardware

On premises, at the edge, next to the cameras: for qualifying deployments, DecisionNode can run on hardware you own. Tell us about yours.

Talk to us

Machines and drones

Command-level decisions for warehouse machines and robots: go, stop, which aisle, one typed answer per frame. Drones the same way: send their camera frames like any other image, or stream them through a session that hands the controller a decision for every frame.

where

Point to it. Draw a box around it.

Four builds for the two types that answer where: a point is one spot in the image, a box the rectangle around the thing, each with how likely the thing is there and how likely the place is right.

Production lines and sorting"point"

Robot arm that picks the right package

Point to the package to pick, by colour and by where it lies, and hand the spot straight to the arm.

(812, 455)point, present 0.97, confidence 0.91

input: belt camera, line 3

Camera frame from above a conveyor belt: a robot arm reaches over a yellow package, a second yellow package further down, and a brown carton with a torn corner at the top left
pick

question

"pick": { "type": "point", "instructions" }

Point to the yellow package closest to the robot arm.

answer

"pick": {
  "type": "point",
  "x": 812,
  "y": 455,
  "present": 0.97,
  "confidence": 0.91,
  "image": "belt",
  "width": 1920,
  "height": 1080
}

then your code

present 0.97 clears the arm's line (0.80), so it picks at (812, 455) and the belt moves on. Under the line it holds and asks again on the next frame, or hands the frame to a person instead.

point and box, on POST /v1/decide

Read the docsOpen in playground: Robot arm that picks the right package

Quality control"box"

Quality check that marks the defect

Draw a box around the scratch, so the panel goes to rework with the damage already marked on its photo.

(600, 282) to (846, 458)box, present 0.94, confidence 0.86

input: inspection camera, station QC 2

Close camera frame of a brushed aluminium panel with four screws, a vent slot row and a bright diagonal scratch right of centre
scratch

question

state Part 4410-B, lot 0817. Panels with any scratch longer than 5 mm go to rework.

"scratch": { "type": "box", "instructions" }

Draw a box around the scratch on the panel.

answer

"scratch": {
  "type": "box",
  "x1": 600,
  "y1": 282,
  "x2": 846,
  "y2": 458,
  "present": 0.94,
  "confidence": 0.86,
  "image": "panel",
  "width": 1200,
  "height": 800
}

then your code

present 0.94: the panel is diverted to rework and the box is saved with its photo, so the rework station sees the scratch before it picks the panel up. Under 0.50 the panel ships, or its photo goes to an inspector instead.

point and box, on POST /v1/decide

Read the docsOpen in playground: Quality check that marks the defect

Surveillance"box"

Perimeter camera that finds the climber

Draw a box around the person climbing the fence, and turn the lights and the nearest camera onto them.

(706, 226) to (822, 566)box, present 0.96, confidence 0.89

input: perimeter camera, north fence

Night infrared camera frame of a chain-link perimeter fence with barbed wire, a person climbing it right of centre, timestamp 23:41
climber

question

state North fence, 23:41. The site is closed; nobody is scheduled on this side.

"climber": { "type": "box", "instructions" }

Draw a box around the person climbing the fence.

answer

"climber": {
  "type": "box",
  "x1": 706,
  "y1": 226,
  "x2": 822,
  "y2": 566,
  "present": 0.96,
  "confidence": 0.89,
  "image": "fence",
  "width": 1200,
  "height": 800
}

then your code

present 0.96: the floodlights on that stretch come on, the pan camera turns to the box, and the incident is logged with the crop. Under 0.60 the frame is checked again a second later, or sent to the guard on duty instead.

point and box, on POST /v1/decide

Read the docsOpen in playground: Perimeter camera that finds the climber

Drones and robots"point"

Drone that finds its landing pad

Point to the landing marker in the drone's own camera frame and land on it.

(742, 452)point, present 0.98, confidence 0.93

input: drone camera, looking down

Camera frame looking straight down at a rooftop: an air unit, vents, a cable tray and a dark square landing pad with a ring and a yellow H, the drone's shadow beside it
marker

question

"marker": { "type": "point", "instructions" }

Point to the landing marker.

answer

"marker": {
  "type": "point",
  "x": 742,
  "y": 452,
  "present": 0.98,
  "confidence": 0.93,
  "image": "roof",
  "width": 1200,
  "height": 800
}

then your code

present 0.98: the flight controller centres on (742, 452) and descends. Under 0.80 it holds its hover and asks again on the next frame, or hands the frame to its pilot instead.

point and box, on POST /v1/decide

Read the docsOpen in playground: Drone that finds its landing pad
builds

Builds you could ship this week.

Eight jobs teams still hand-code with brittle rules or hand to a slow chat model, each answered in one call, and one control loop that streams its decisions. The code acts on every answer the moment it lands.

3 questions, one call

Support inbox that routes itself

Route every ticket, rank how urgent it is and pay back double charges before anyone opens the inbox.

0.94truerefund, clears 0.80

input: webhook: new ticket

Customer: I was charged twice and nobody has replied for 3 days.

answers, one call

  • routeChoice, billing0.87
  • urgencyScore, urgent2.31
  • refundTruth, refunds at 0.800.94 true

then your code

  • Queued with billingif route.choice == "billing"
  • Moved to the front, urgentif urgency.score >= 2
  • Duplicate charge refunded, reply sentif refund.truth >= 0.80

62 input tokens, $2.60 per 1M calls

Open in playground: Support inbox that routes itself

3 questions, one call

Guardrail for agent tool calls

Before an agent runs a tool call, check it is needed and how hard it would be to undo.

block0.82verdict, risk destructive

agent, shell toolchecked before every call

# task: Clear the build cache so the next CI run starts clean.

$ rm -rf ./ && git push --force origin main

blocked before it ranblock 0.82, in scope 0.05 false

$ rm -rf ./node_modules/.cache

re-planned by the agent, checked again

ran on its ownrun 0.96, risk 0.04 (harmless), in scope 0.93 true

> removed 2,184 cached files, 312 MB

$

answers, one call

  • verdictChoice, block0.82
  • riskScore, destructive2.76
  • in_scopeTruth, runs at 0.800.05 false

then your code

block 0.82, risk 2.76 (destructive), needed for the task 0.05 false: the call never runs, and the agent gets it back to plan a narrower one.

231 input tokens, $9.70 per 1M calls

Open in playground: Guardrail for agent tool calls

3 questions, one call

Moderation that acts on every comment

Allow, limit or remove each comment the moment it is posted, with a toxicity level and a promotion check in the same call.

remove0.61action, toxicity harsh

input: new comment onShow HN: a tiny CLI that resizes images in place

account 2 days old, 1 earlier removal

This is useless. Anyone who ships this should quit programming. Check my profile for a REAL tool, 50% off today only.

what the code did

  1. hidden on arrivalremove 0.61, under 0.70
  2. links strippedpromotion 0.88 true, strips at 0.80

link posting off for this account

answers, one call

  • actionChoice, remove0.61
  • toxicityScore, harsh2.01
  • promotionTruth, strips links at 0.800.88 true

then your code

remove 0.61 is under 0.7, so the safe action runs at once: the comment is hidden before anyone replies. Promotion 0.88 true clears 0.80: links stripped.

133 input tokens, $5.59 per 1M calls

Open in playground: Moderation that acts on every comment

5 questions, 1 image, one call

Receipt and invoice checker

Read a receipt photo against the expense claim: total, date, currency and policy, in one call.

input: receipt photo

Photo of a paper receipt from Hallon Bistro, Stockholm, dated 2026-09-28: starters, two mains, a bottle of Riesling and coffee, total 1 315,50 kr.

input: expense claim

employee
M. Okafor
merchant
Hallon Bistro
claimed total
1315.50
currency
SEK
date
2026-09-28
category
client dinner

1.00truetotal matches, pays at 0.90

claim paidwine moved to entertainment

answers, one call

  • total_matchesTruth, pays at 0.901.00 true
  • date_matchesTruth, pays at 0.901.00 true
  • currencyChoice, SEK1.00
  • legibleScore, fully readable1.99
  • alcoholTruth, splits at 0.800.99 true

then your code

total 1.00 true and date 1.00 true both clear 0.90, currency SEK 1.00: the claim is paid. Alcohol 0.99 true moves the wine to the entertainment budget.

521 input tokens, $34.27 per 1M calls

Open in playground: Receipt and invoice checker

input: pages 1 and 2 of a 6-page PDF

Page 1 of a residential lease agreement between Marrow Lane Properties LLC and Ines Albrecht: the parties, the premises at 48 Linden Court, the term, the monthly rent, the security deposit and utilities.
Page 2 of the same lease: use of the premises, maintenance and repairs, entry by the landlord, pets, insurance and ending the lease.

1 question, 2 images, one call

Classify a document from its pages

Send the first pages of an uploaded PDF as images and ask what the document is: one choice over 85 types, unknown among them.

lease agreement0.92the most probable of 85 types

filed as a leaselease workflow started

answers, one call

  • doc_typeChoice, lease agreement0.92

then your code

lease agreement 0.92 clears 0.80: the file is filed as a lease and goes to the lease workflow, which reads the rent, the term and the deposit next.

The question's criteria hold all 85 types, each a code and a description; the answer's probabilities list every one of them, zeros included.

2904 input tokens, $146.74 per 1M calls

Open in playground: Classify a document from its pages

3 questions, one call

Lead scoring for new signups

Score each signup from its note and first-week usage, and tell production builds from evaluations.

input: signup

role
Staff engineer
first week
18,240 requests, 5 days
note
40k support emails a day, wants calibrated answers

2.49hotfit, hot

rate limit, set by the answer

new workspace40 req/s
production1,000 req/s

fast track 0.90 true, upgrades at 0.85

answers, one call

  • fitScore, hot2.49
  • intentChoice, production0.78
  • fast_trackTruth, upgrades at 0.850.90 true

then your code

fit 2.49 (hot), production 0.78, fast track 0.90 true: the account moves to production limits and gets the production onboarding email now.

158 input tokens, $6.64 per 1M calls

Open in playground: Lead scoring for new signups

3 questions, 1 image, one call

Listing photo match

Check the photo shows the item the seller describes, and grade its condition from the picture.

0.93truephoto matches, publishes at 0.85

input: listing photo + description

A silver and black vintage rangefinder camera resting on piano keys.

goes live as

Vintage 35mm rangefinder camera

€140

  • verified photo
  • film camera
  • condition: good

answers, one call

  • matchesTruth, publishes at 0.850.93 true
  • itemChoice, film camera0.90
  • conditionScore, good2.10

then your code

match 0.93 true clears 0.85, film camera 0.90, condition good: the listing goes live with a verified photo and a condition tag.

391 input tokens, $28.81 per 1M calls

Open in playground: Listing photo match

3 questions, one call

Statement reader that chases late invoices

Read a customer statement, count what is overdue and how late the oldest is, and send the right reminder on its own.

4(0.84)overdue invoices, acts at 0.70

input: customer statement

account
Norrvik Supply AB (30-4471)
statement date
2026-10-01
terms
net 30, amounts in EUR
Invoice lines on the statement
invoiceissueddueamountstatus
INV-20412026-06-152026-07-151,240.00open
INV-20582026-07-032026-08-02410.00paid 2026-08-01
INV-20632026-07-212026-08-20860.00open
INV-20692026-07-292026-08-28395.00paid 2026-09-02
INV-20772026-08-062026-09-051,120.00part paid, 300.00 open
INV-20812026-08-132026-09-12640.00open
INV-20902026-08-312026-09-30780.00paid 2026-09-29
INV-20942026-09-102026-10-10520.00open
INV-21022026-09-242026-10-24915.00open
firm reminder sent, 4 invoices listed

answers, one call

  • overdueNumber, expected 4.004 (0.84)
  • oldest_daysNumber, expected 78.0878 (0.61)
  • reminderChoice, firm0.76

then your code

4 (0.84) overdue and firm 0.76 both clear 0.7, so the firm reminder goes out today with every open invoice listed. The oldest reads 78 days at 0.61, under 0.7, so the letter says more than 60 days instead of an exact count.

445 input tokens, $18.69 per 1M calls

Open in playground: Statement reader that chases late invoices

2 questions per frame, one open session

Survey drone that supervises its own mission

Open one session with the mission brief and two questions, then stream every telemetry frame and get the flight mode and an abort check back for each one.

More than 10 decisions a second?Use a session.

1. open oncePOST /v1/sessions

model
decisionnode
instructions
You are the mission supervisor of a survey drone.
state
mission brief, geofence, rules
questions
mode (choice), abort (truth)
window
8 recent frames

2. then stream framesWebSocket /v1/sessions/{id}/stream

frames in, one decision back for each

  1. frame 15

    battery 0.36, wind 11.0 m/s, home 835 m

    mode: return0.52abort0.08false

  2. frame 16

    battery 0.34, wind 11.2 m/s, home 830 m

    mode: return0.71abort0.08false

  3. frame 17

    battery 0.31, wind 11.4 m/s, home 820 m

    mode: return0.84abort0.07false

aborts at 0.80; below it the mission flies on

then your code

Frame 17 comes back return 0.84, abort 0.07 false. The supervisor hands return-to-home to the drone's flight controller, which flies the route back; the next frames keep checking abort on the way.

latest wins

A client that falls behind gets the newest answer, never a backlog. Earlier on this flight frames 9, 10 and 11 arrived together: only 11 was answered, and the reply reported 9 and 10 as skipped.

billed per frame: its own tokens and the questions asked

Read the sessions docs: Survey drone that supervises its own mission
knowledge

Decide by your own rules.

Policies, manuals and playbooks go into a knowledge base. The knowledge booster gives each request the passages that apply, pinned rules always, and the answer cites every passage it read.

3 questions, 3 documents, one call

Water damage claim against the claims policy

The claims policy is a knowledge base: the claim is read against the passages that apply, the receipt rule pinned so it always does, and the answer cites each one.

0.98true

covered

answers, one call

  • coveredTruth, probability it is true0.98 true
  • sofa_in_fullTruth, probability it is true0.06 false
  • settleChoice, pay in part0.81

passages cited: 7 of “Sample: Water damage claim against the claims policy”

  1. [1]Claims handling rules, Receiptspinned

    An item claimed at more than EUR 500 is paid only with a receipt or a bank statement that shows the purchase. Without one, pay at most EUR 500 for that item.

  2. [2]Claims handling rules, Deadlinespinned

    A claim is filed within 30 days of the day the loss was discovered. A later claim goes to an adjuster.

  3. [3]Claims handling rules, Settling a claimpinned

    Pay in full when every item is covered and documented. Pay in part when some items are covered and others are excluded or capped. Refer to an adjuster when the total claimed is over EUR 10,000, the claim is late, or the cause is unclear. Decline when nothing claimed is covered.

  4. [4]Water damage, Covered

    Sudden and accidental escape of water from a pipe, a hose, a tank or an appliance inside the home: a burst pipe, a split washing machine hose, a dishwasher seal that fails at once.

  5. [5]Water damage, Not covered

    - Water that leaked slowly over weeks or months, and the rot or mould it caused. - Flooding from outside the home (rivers, heavy rain, rising groundwater), unless the policy has the Flood add-on. - The part that failed: the hose, the pipe or the seal itself is not replaced.

  6. [6]Water damage, What the policyholder does

    Stop the water at once and keep the failed part for the adjuster.

  7. [7]Theft and burglary, Covered

    Theft after a forced entry into the home, reported to the police within 48 hours.

input: the knowledge baseMarkdown articles; open one to read it

  • Claims handling rulespinned
    # Claims handling rules
    
    ## Receipts
    
    An item claimed at more than EUR 500 is paid only with a receipt or a bank statement that shows the purchase. Without one, pay at most EUR 500 for that item.
    
    ## Deadlines
    
    A claim is filed within 30 days of the day the loss was discovered. A later claim goes to an adjuster.
    
    ## Settling a claim
    
    Pay in full when every item is covered and documented. Pay in part when some items are covered and others are excluded or capped. Refer to an adjuster when the total claimed is over EUR 10,000, the claim is late, or the cause is unclear. Decline when nothing claimed is covered.
  • Water damage
    # Water damage
    
    ## Covered
    
    Sudden and accidental escape of water from a pipe, a hose, a tank or an appliance inside the home: a burst pipe, a split washing machine hose, a dishwasher seal that fails at once.
    
    ## Not covered
    
    - Water that leaked slowly over weeks or months, and the rot or mould it caused.
    - Flooding from outside the home (rivers, heavy rain, rising groundwater), unless the policy has the Flood add-on.
    - The part that failed: the hose, the pipe or the seal itself is not replaced.
    
    ## What the policyholder does
    
    Stop the water at once and keep the failed part for the adjuster.
  • Theft and burglary
    # Theft and burglary
    
    ## Covered
    
    Theft after a forced entry into the home, reported to the police within 48 hours.
    
    ## Not covered
    
    Theft from an unlocked home or an unlocked car, and items left outside the home overnight.

the requestcollection is your knowledge base's id

POST /v1/decide, boosters: knowledge
{
  "model": "decisionnode",
  "state": {
    "policy": "Home Standard, no add-ons",
    "claim": {
      "cause": "The washing machine's inlet hose split while the policyholder was at work; water ran for about four hours.",
      "discovered": "2026-10-14",
      "filed": "2026-10-16",
      "items": [
        {
          "item": "laminate floor, kitchen and hall",
          "amount_eur": 1850,
          "proof": "receipt"
        },
        {
          "item": "sofa",
          "amount_eur": 920,
          "proof": "none"
        },
        {
          "item": "washing machine inlet hose",
          "amount_eur": 25,
          "proof": "receipt"
        }
      ]
    }
  },
  "boosters": [
    {
      "name": "knowledge",
      "options": {
        "collection": "sample-claims-policy"
      }
    }
  ],
  "questions": {
    "covered": {
      "type": "truth",
      "instructions": "Is the cause of this water damage covered by the policy?",
      "criteria": {
        "true": "The policy's water damage rules cover the cause",
        "false": "The cause is excluded or not covered"
      }
    },
    "sofa_in_full": {
      "type": "truth",
      "instructions": "Can the sofa be paid at the amount claimed?",
      "criteria": {
        "true": "The claims rules allow paying the full amount claimed for it",
        "false": "The rules cap it or refuse it"
      }
    },
    "settle": {
      "type": "choice",
      "instructions": "How is this claim settled under the claims handling rules?",
      "criteria": {
        "pay_in_full": "every item covered and documented",
        "pay_in_part": "some items covered, others excluded or capped",
        "refer_to_adjuster": "over EUR 10,000, filed late, or the cause unclear",
        "decline": "nothing claimed is covered"
      }
    }
  }
}

measured on decisionnode-flash-0.9.1-rc1, median of 3 runs, 2026-10-23

783 input tokens, $1,032.89 per 1M calls

Open in playground: Water damage claim against the claims policy

3 questions, 3 documents, one call

Delivery rover against its operating manual

One knowledge base holds the manual of every rover model, scoped by model: this R4 reads its own manual and the pinned fleet rules, never the R2's speed table.

stop and wait0.94

action

answers, one call

  • actionChoice, stop and wait0.94
  • over_speedTruth, probability it is true0.07 false
  • max_speedNumber, expected 4.915 (0.96)

passages cited: 4 of “Sample: Delivery rover against its operating manual”

  1. [1]Fleet safety rulespinned

    These apply to every rover model. - People on foot always have the right of way. Never pass a person closer than the model's manual allows. - Cross a road only at a marked crossing, and only on a green signal. - Stop and hand over to a remote operator when a sensor reports a fault, when a person touches or blocks the rover on purpose, or when the path stays blocked for more than 60 seconds.

  2. [2]R4 operating manual, Speed by terrain

    | Terrain | Highest speed | |---|---| | Pavement | 8 km/h | | Gravel or grass | 5 km/h | | Crowded (more than 5 people within 10 m) | 3 km/h |

  3. [3]R4 operating manual, Kerbs

    The R4 climbs kerbs up to 10 cm. At a higher kerb, stop and hand over to a remote operator.

  4. [4]R4 operating manual, Distance to people

    Keep at least 1.5 m from an adult and 3 m from a child or an animal. When one is closer, stop and wait until the gap is back.

input: the knowledge baseMarkdown articles; open one to read it

  • Fleet safety rulespinned
    # Fleet safety rules
    
    These apply to every rover model.
    
    - People on foot always have the right of way. Never pass a person closer than the model's manual allows.
    - Cross a road only at a marked crossing, and only on a green signal.
    - Stop and hand over to a remote operator when a sensor reports a fault, when a person touches or blocks the rover on purpose, or when the path stays blocked for more than 60 seconds.
  • R2 operating manualmodel: R2
    # R2 operating manual
    
    ## Speed by terrain
    
    | Terrain | Highest speed |
    |---|---|
    | Pavement | 6 km/h |
    | Gravel or grass | 3 km/h |
    | Crowded (more than 5 people within 10 m) | 2 km/h |
    
    ## Distance to people
    
    Keep at least 1 m from any person.
    
    ## Kerbs
    
    The R2 does not climb kerbs over 5 cm; route around them.
  • R4 operating manualmodel: R4
    # R4 operating manual
    
    ## Speed by terrain
    
    | Terrain | Highest speed |
    |---|---|
    | Pavement | 8 km/h |
    | Gravel or grass | 5 km/h |
    | Crowded (more than 5 people within 10 m) | 3 km/h |
    
    ## Distance to people
    
    Keep at least 1.5 m from an adult and 3 m from a child or an animal. When one is closer, stop and wait until the gap is back.
    
    ## Kerbs
    
    The R4 climbs kerbs up to 10 cm. At a higher kerb, stop and hand over to a remote operator.

the requestcollection is your knowledge base's id

POST /v1/decide, boosters: knowledge, scope R4
{
  "model": "decisionnode",
  "state": {
    "rover_model": "R4",
    "terrain": "gravel path in a park",
    "speed_kmh": 4.5,
    "people_within_10m": 2,
    "nearest": {
      "who": "child",
      "distance_m": 2.2
    },
    "path_blocked_s": 0,
    "sensors": "ok"
  },
  "boosters": [
    {
      "name": "knowledge",
      "options": {
        "collection": "sample-rover-manual",
        "scope": "R4"
      }
    }
  ],
  "questions": {
    "action": {
      "type": "choice",
      "instructions": "What does the rover do now, under its manual and the fleet rules?",
      "criteria": {
        "continue": "keep going at the current speed",
        "slow_down": "keep going, below the terrain's highest speed",
        "stop_and_wait": "stop until the gap to people is back",
        "hand_over": "stop and hand over to a remote operator"
      }
    },
    "over_speed": {
      "type": "truth",
      "instructions": "Is the rover going faster than its manual allows on this terrain?",
      "criteria": {
        "true": "`speed_kmh` is above the manual's highest speed for the terrain",
        "false": "It is at or below it"
      }
    },
    "max_speed": {
      "type": "number",
      "instructions": "What is the highest speed, in km/h, the manual allows on this terrain?",
      "min": 0,
      "max": 10
    }
  }
}

measured on decisionnode-flash-0.9.1-rc1, median of 3 runs, 2026-10-23

633 input tokens, $1,026.59 per 1M calls

Open in playground: Delivery rover against its operating manual

3 questions, 3 documents, one call

Refund request against the refund policy

The refund policy says how each request is treated: approve, escalate, refuse or flag. The pinned policy and the warranty passage decide this one, and the answer cites both.

escalate0.87

action

answers, one call

  • actionChoice, escalate0.87
  • in_windowTruth, probability it is true0.07 false
  • under_warrantyTruth, probability it is true0.99 true

passages cited: 7 of “Sample: Refund request against the refund policy”

  1. [1]Refund policypinned

    Every request gets exactly one of four actions.

  2. [2]Refund policy, Approvepinned

    Approve a refund when the order arrived within the last 30 days, the item is unused or arrived faulty, and the order total is EUR 200 or less.

  3. [3]Refund policy, Escalatepinned

    Escalate to a team lead when the order total is over EUR 200, when the customer had another refund in the last 90 days, or when a fault is reported after 30 days and within the 2-year warranty: the team lead offers a repair or a replacement, never cash.

  4. [4]Refund policy, Refusepinned

    Refuse when the order arrived more than 30 days ago and no fault is reported, when a digital item was downloaded, or when the item was made to order.

  5. [5]Refund policy, Flagpinned

    Flag for fraud review when delivery was signed for but the customer says it never arrived, when the customer had three or more refunds in the last 90 days, or when the account is under 7 days old and the order is over EUR 100. Flag comes before every other action.

  6. [6]Warranty

    Every product is under warranty for 2 years from delivery. The warranty covers faults in materials and manufacture: a battery that stops charging, a button that stops working, a seam that splits. It does not cover damage from a drop or from water, or wear such as scratched surfaces and worn ear cushions.

  7. [7]Shipping and delivery

    Orders ship within 2 working days. A parcel over EUR 150 is delivered against a signature. A customer who reports a parcel missing is asked to check with neighbours first, and the carrier is asked for its proof of delivery.

input: the knowledge baseMarkdown articles; open one to read it

  • Refund policypinned
    # Refund policy
    
    Every request gets exactly one of four actions.
    
    ## Approve
    
    Approve a refund when the order arrived within the last 30 days, the item is unused or arrived faulty, and the order total is EUR 200 or less.
    
    ## Escalate
    
    Escalate to a team lead when the order total is over EUR 200, when the customer had another refund in the last 90 days, or when a fault is reported after 30 days and within the 2-year warranty: the team lead offers a repair or a replacement, never cash.
    
    ## Refuse
    
    Refuse when the order arrived more than 30 days ago and no fault is reported, when a digital item was downloaded, or when the item was made to order.
    
    ## Flag
    
    Flag for fraud review when delivery was signed for but the customer says it never arrived, when the customer had three or more refunds in the last 90 days, or when the account is under 7 days old and the order is over EUR 100. Flag comes before every other action.
  • Warranty
    # Warranty
    
    Every product is under warranty for 2 years from delivery. The warranty covers faults in materials and manufacture: a battery that stops charging, a button that stops working, a seam that splits.
    
    It does not cover damage from a drop or from water, or wear such as scratched surfaces and worn ear cushions.
  • Shipping and delivery
    # Shipping and delivery
    
    Orders ship within 2 working days. A parcel over EUR 150 is delivered against a signature. A customer who reports a parcel missing is asked to check with neighbours first, and the carrier is asked for its proof of delivery.

the requestcollection is your knowledge base's id

POST /v1/decide, boosters: knowledge
{
  "model": "decisionnode",
  "state": {
    "customer": {
      "account_age_days": 412,
      "refunds_last_90_days": 1
    },
    "order": {
      "id": "SO-88213",
      "item": "wireless earbuds",
      "total_eur": 149,
      "delivered": "2026-09-02"
    },
    "today": "2026-10-20",
    "message": "The left earbud stopped charging last week. I would like my money back."
  },
  "boosters": [
    {
      "name": "knowledge",
      "options": {
        "collection": "sample-refund-policy"
      }
    }
  ],
  "questions": {
    "action": {
      "type": "choice",
      "instructions": "Which action does the refund policy give this request?",
      "criteria": {
        "approve": "refund the order",
        "escalate": "hand it to a team lead",
        "refuse": "no refund",
        "flag": "send it to fraud review"
      }
    },
    "in_window": {
      "type": "truth",
      "instructions": "Did the order arrive within the last 30 days before `today`?"
    },
    "under_warranty": {
      "type": "truth",
      "instructions": "Is the fault the customer reports covered by the warranty?",
      "criteria": {
        "true": "A fault the warranty covers, within its 2 years",
        "false": "Damage or wear the warranty excludes, or past its 2 years"
      }
    }
  }
}

measured on decisionnode-flash-0.9.1-rc1, median of 3 runs, 2026-10-23

747 input tokens, $1,031.37 per 1M calls

Open in playground: Refund request against the refund policy

3 questions, 3 documents, one call

Reported post against the trust and safety policy

A reported post is checked against the community's own trust and safety policy: the pinned enforcement order picks the action, the rule it breaks is cited beside it.

remove and suspend0.90

action

answers, one call

  • actionChoice, remove and suspend0.90
  • breaks_ruleTruth, probability it is true0.95 true
  • ruleChoice, spam and scams1.00

passages cited: 5 of “Sample: Reported post against the trust and safety policy”

  1. [1]Enforcement actionspinned

    Take the first action that applies, from the top. 1. Escalate to the safety team: a credible threat of violence, or a risk to someone's life. 2. Remove and suspend the account: a scam, a second removal within 30 days, or any removal on an account under 7 days old. 3. Remove: the post breaks a rule of this policy. 4. Label: the post is allowed but could mislead, such as an unverified claim about health or money. 5. Leave up: the post breaks no rule.

  2. [2]Spam and scams, Not allowed

    - Asking people to send money, gift cards or cryptocurrency with a promise of more back. - Pretending to be the platform, its staff or a well-known brand to win trust. - Links to pages that ask for passwords or card details.

  3. [3]Spam and scams, Allowed

    Selling your own goods in the marketplace section, at a fixed price.

  4. [4]Harassment, Not allowed

    Insults aimed at a person, sharing someone's private details, and repeated contact after being asked to stop.

  5. [5]Harassment, Allowed

    Criticism of ideas, of products, and of public figures' public acts.

input: the knowledge baseMarkdown articles; open one to read it

  • Enforcement actionspinned
    # Enforcement actions
    
    Take the first action that applies, from the top.
    
    1. Escalate to the safety team: a credible threat of violence, or a risk to someone's life.
    2. Remove and suspend the account: a scam, a second removal within 30 days, or any removal on an account under 7 days old.
    3. Remove: the post breaks a rule of this policy.
    4. Label: the post is allowed but could mislead, such as an unverified claim about health or money.
    5. Leave up: the post breaks no rule.
  • Spam and scams
    # Spam and scams
    
    ## Not allowed
    
    - Asking people to send money, gift cards or cryptocurrency with a promise of more back.
    - Pretending to be the platform, its staff or a well-known brand to win trust.
    - Links to pages that ask for passwords or card details.
    
    ## Allowed
    
    Selling your own goods in the marketplace section, at a fixed price.
  • Harassment
    # Harassment
    
    ## Not allowed
    
    Insults aimed at a person, sharing someone's private details, and repeated contact after being asked to stop.
    
    ## Allowed
    
    Criticism of ideas, of products, and of public figures' public acts.

the requestcollection is your knowledge base's id

POST /v1/decide, boosters: knowledge
{
  "model": "decisionnode",
  "state": {
    "report": {
      "reason": "scam",
      "reports": 7
    },
    "post": {
      "text": "GIVEAWAY: the first 50 people to send 0.05 BTC to the wallet below get 0.5 BTC back within the hour. Verified by the Pinehall team.",
      "links": 1
    },
    "author": {
      "account_age_days": 2,
      "removals_last_30_days": 0
    }
  },
  "boosters": [
    {
      "name": "knowledge",
      "options": {
        "collection": "sample-trust-safety"
      }
    }
  ],
  "questions": {
    "action": {
      "type": "choice",
      "instructions": "Which enforcement action does the policy give this post?",
      "criteria": {
        "leave_up": "it breaks no rule",
        "label": "allowed, but could mislead",
        "remove": "it breaks a rule",
        "remove_and_suspend": "remove it and suspend the account",
        "escalate": "send it to the safety team"
      }
    },
    "breaks_rule": {
      "type": "truth",
      "instructions": "Does the post break a rule of the trust and safety policy?"
    },
    "rule": {
      "type": "choice",
      "instructions": "Which part of the policy does the post fall under?",
      "criteria": {
        "spam_and_scams": "money, impersonation or phishing",
        "harassment": "aimed at a person",
        "none": "no rule applies"
      }
    }
  }
}

measured on decisionnode-flash-0.9.1-rc1, median of 3 runs, 2026-10-23

628 input tokens, $1,026.38 per 1M calls

Open in playground: Reported post against the trust and safety policy

Your build is one request away.

Every card on this page is the same call with a different state and different questions. Paste this one as it is, then make it yours in the playground and copy it out as curl, Python or TypeScript.

  1. 01Get a keyFrom the console. Prepaid, pay per input token, output is free.
  2. 02Describe the decisionState is any text, JSON or image. Questions are Choice, Score, Truth, Number, Point or Box.
  3. 03Branch on the answerAct above your threshold. Below it, take the safe path or re-check with the full model.
Open playground
the whole integrationPOST /v1/decide
curl https://api.decisionnode.com/v1/decide \  -H "Authorization: Bearer $DECISIONNODE_API_KEY" \  -H "Content-Type: application/json" \  -d '{    "model": "decisionnode-flash",    "state": "Charged twice for order #4471, no reply in 3 days.",    "questions": {      "route": {        "type": "choice",        "criteria": { "billing": "money", "bug": "broken" }      },      "urgency": {        "type": "score",        "criteria": ["routine", "today", "urgent"]      },      "refund": {        "type": "truth",        "criteria": { "true": "Charged twice", "false": "Anything else" }      }    }  }'
DecisionNodeDecisionNde

Add a decision engine with one prompt.

Paste into your coding agent.

start building

Your first decision in minutes.

Make a key, connect your project, branch on a typed answer. Prepaid, no sales call.

  • api all systems normalapi normal
  • output is always freeoutput always free
  • same request, same answersame request, same answer

DecisionNode is built and run by Bynn Intelligence, Inc.

We train the model and serve it on our own GPUs.

hello@bynn.com

product

  • Model
  • Inference
  • Benchmarks
  • Pricing
  • Open models

developers

  • Docs
  • Quickstart
  • Examples
  • Playground
  • Console
  • Status

company

  • Contact
  • Security
  • Report abuse
  • Dedicated
  • Affiliate program

legal

  • Terms
  • Acceptable use
  • Privacy
  • Data processing
  • Defence addendum
  • Cookies
  • Affiliate agreement

compliance

  • SOC 2, SOC 2 report, monitored continuously (opens the trust center in a new tab)
  • ISO 27001:2022, Information security management, certified (opens the trust center in a new tab)
  • ISO/IEC 42001:2023, AI management system, controls mapped to ISO/IEC 42001:2023; compliant, not certified (opens the trust center in a new tab)
  • GDPR, Processor terms under Article 28 in the Data Processing Addendum; a law, not a certification (opens the trust center in a new tab)
  • CCPA, Service-provider terms in the Data Processing Addendum; a law, not a certification (opens the trust center in a new tab)
  • PCI DSS SAQ A, Card data is handled by the payment provider and never stored by us; a self-assessment, not a certification (opens the trust center in a new tab)
  • NIST Cybersecurity Framework, Controls mapped to the NIST CSF; a framework, not a certification (opens the trust center in a new tab)
  • HIPAA (BAA available), Protected health information only under a signed Business Associate Agreement, available on request; HIPAA is not a certification (opens the trust center in a new tab)
  • EU AI Act, Controls mapped to the EU AI Act; a law, not a certification, and no determination that any customer use complies (opens the trust center in a new tab)

© 2026 Bynn Intelligence, Inc.

prices in USD per 1M input tokens