Get started
01 / 06The method4 checks · 3 outcomes
Method · 4 checks · 3 outcomes

The order of the questions is the method.

A watched move becomes a reviewable decision by passing the same checks in the same order, every time. This page is that order, printed from the function that applies it rather than described beside it.

Every figure below was returned by the engine when this page rendered. Nothing is typed in.

What the engine readsMarketEvaluationProductionConstraintsOne openingShould the strategy change?And what may leaveStayTestSwitchA person rules, whichever it is

Aperture · the decision engineMarket, evaluation, production and constraints meet in one place, and what leaves is not a score but a position: should the strategy change, and what evidence would change the answer. Every verdict is reviewable down to the threshold that decided it.

02 / 06Four inputs10 of 10 component classes wired
Four inputs

What the engine reads, and who holds it.

Two of the four are things Vitna does not hold. The record supplies the market; the evaluation, the production readings and the constraints are yours, and the engine supplies none of them.

  1. Held by the record
    Market

    713 skus from 34 providers on the register, read from 62 registered sources on their own cadence.

  2. Held by you
    Evaluation

    Your representative cases, scored by whatever you already run. Vitna holds none of them and generates none.

  3. Held by you
    Production

    What your traffic actually did: success rates, the cases that failed, the p90 the run measured.

  4. Held by you
    Constraints

    What you wrote down: a quality floor, an economic threshold, the cases that must not fail, and 4 kinds of term a provider must have published.

10 of 10 component classes wired

A class earns a live market only when its pricing can be read from a document a provider published. The rest are labelled not yet wired rather than dressed in illustrative prices, and the engine is not run over them.

Read the tape
03 / 06The loop and check order4 checks in decide()
Three entries

A loop, an order, and what the method refuses.

Detection is cheap and constant. Everything downstream of it is deliberate, and it happens in one fixed order.

Entry 01 / 03The loop

Detection is cheap and constant. Everything downstream of it is deliberate.

  • A threshold somebody armed, crossed once
  • Candidates run against your own cases
  • Every brief names the trigger that reopens it
01 Detect · a threshold somebody armedThe market02Gate03Test04DecideThe ruling05 Re-armEvery brief names the trigger that reopens it
  1. 01
    Detect

    A watched market moves: a price, a release, a deprecation, an availability change. Most moves are noise, so detection is cheap and constant and nothing downstream of it is.

  2. 02
    Gate

    The move is weighed against the threshold somebody armed. A move that does not cross it spends nothing, and a watch fires once rather than every hour until it is muted.

  3. 03
    Test

    Candidates run against your representative cases: real work, edge cases, regression sets. The measurements come back as evidence rows the brief can cite.

  4. 04
    Decide

    The checks below, in the order below. What comes out is a brief with the thresholds, the measurements and the failures on its face, and a person rules on it.

  5. 05
    Re-arm

    Every brief names the trigger that reopens the question. Decisions expire, and the trigger is written down before anyone has forgotten why the decision was taken.

Entry 02 / 03The checks

The order is not a style. It is the judgement.

  • 4 checks, measured on every run
  • The earliest failure is the one that decides
  • Printed from the function that applies them
The run, left to right01020304A run that fails one stops, and the brief says whichRuled

Every check below is measured on every run, and the earliest one that fails is the one that decides. They are printed in the order platform/engine/decide.mjs applies them, and there are as many of them as that file applies: 4, today.

  1. 01

    eligibility

    nothing required

    May this candidate serve this work at all?

    Asked before anything is weighed, because a candidate that may not serve this work is not more expensive, it is unavailable. A requirement the record cannot check is not satisfied: where no provider term answers it, the run stops and the brief names the term that is missing rather than assuming either answer.

  2. 02

    quality floor

    floor 96

    Does it clear the lowest measured success rate this workload accepts?

    First of the measured checks, because a candidate that cannot do the work is not cheaper at anything. The floor is the customer's, the rate is their own evidence, and Vitna supplies neither.

  3. 03

    critical failures

    0 allowed

    Does it still fail a case this workload cannot afford to fail?

    Second, because an aggregate score hides the cases that would page somebody at 2am. Where a workload names the cases that must not fail, those carry a veto over the average; where it names none, every reported failure is critical, which is the stricter reading.

  4. 04

    economic threshold

    requires 25%

    Is the saving large enough to pay for the migration?

    Last, because a saving that does not pay for the work is not a reason to do the work. What counts as material is the workload's own, and it is larger when the workload is run for something other than economics.

04 / 06Three runs at renderSupport · Extraction · Research
The run · at render

The order never changes. The answer does.

Three illustrative workloads, each with the constraints it declares, put through the engine when this page was rendered. Read across a row to compare three runs at one check; read down a column to watch one run stop.

Decisions / Three illustrative runsRun at render · the committed fixture tape
The workloadIncumbent, candidate, and what it is run for
01eligibilitynothing required
02quality floorfloor 96
03critical failures0 allowed
04economic thresholdrequires 25%
Recommendationstay, test, switch
Carried, not readreported on the brief, decisive nowhere
Customer supportGPT-5.6 Luna → Claude Haiku 4.5Run for quality · 3,800 resolved conversations
eligibilitymeets itnothing requiredpassed
quality floor95.1floor 96faileddecides this run
critical failures00 allowedpassed
economic threshold15requires 25%failed
RecommendationStay

The candidate falls below the 96.0 quality floor. Keep the incumbent and continue watching the market.

Carried, not read
Annual opportunity
no volume declared
p90 latency
not measured here
Invoice extractionClaude Haiku 4.5 → Gemini 3.5 Flash-LiteRun for economics · 1,200 representative documents
eligibilitymeets itnothing requiredpassed
quality floor93.4floor 92passed
critical failures20 allowedfaileddecides this run
economic threshold31requires 15%passed
RecommendationTest

The economics are material, but 2 critical edge cases still fail. Run a focused validation before moving traffic.

Carried, not read
Annual opportunity
no volume declared
p90 latency
not measured here
Research synthesisClaude Haiku 4.5 → GPT-5.6 LunaRun for economics · 640 source-grounded reports
eligibilitymeets itnothing requiredpassed
quality floor96.4floor 92passed
critical failures00 allowedpassed
economic threshold36requires 15%passed
RecommendationSwitch

The candidate clears the quality floor with no critical failures and improves cost per successful task by 36%.

Carried, not read
Annual opportunity
no volume declared
p90 latency
not measured here

Illustrative workloads on the committed fixture tape, so this board is the same on every render and its arithmetic is checked by tests/engine-decide.test.mjs. The same function against the record's live prices, with its read date printed on the band, is the Decision Lab on the homepage.

05 / 06Boundaries and refusalsThree boundaries · three refusals
Entry 03 / 03The boundaries

What a method refuses is part of the method.

Three things Vitna is not, and three refusals the engine performs rather than promises.

  1. Never a dependency

    Our software can sit in your request path. Our service never has to. The optional runtime reads a signed policy off your own disk and calls us at no point, so we can be entirely unavailable and your traffic keeps flowing.

  2. Not an eval platform

    Evals are an input here, not the product. Vitna consumes evidence from the tools you already run and turns it into a defensible operating decision.

  3. Not an autopilot

    Vitna recommends and documents. Humans own the switch. Nothing reroutes production traffic silently, and no lane serves a model nobody accepted.

Three refusals the engine performsEach one a throw or an early return
  1. A requirement the record cannot check is not satisfied.

    A workload that requires zero retention, weighed against a provider whose retention the record has never read, could be passed, failed, or quietly ignored. The first two invent a fact. The third looks harmless and is worst, because the brief would then be silent about a constraint the customer wrote down.

    decide · eligibility check
  2. A decision is not taken over a price the record does not hold.

    A unit profile that spends a dimension the tape does not carry is a refusal rather than a zero. Interpolating the missing rate would produce a confident brief resting on a number nobody read off a document.

    decide · cost per unit
  3. A volume nobody measured buys no annual figure.

    Where a workload has not declared how many units it runs, the saving per unit is reported and the annual figure is null. All three runs on the board are in that state, and it says so in words instead of printing a zero.

    decide · derived figures
Carried, not read

Measured, recorded, read by nothing

Latency is measured on every run and carried into the brief, and no check reads it. No workload has yet shown where that threshold should sit, and inventing one here would be a judgement with nothing behind it. This paragraph is written from the checks themselves, so the day a latency check is added it stops being printed.

The same restraint holds a level up. 10 of 10 component classes are wired to a live market; are not, and they are labelled not yet wired rather than dressed in illustrative prices.

Measured every runp90 latencyNo check reads it01020304The checks, and the ruling they reach
06 / 06Sign-offRegister · Operate · Record

Every verdict resolves to the threshold that decided it.

Register what you run, arm a watch against the market, and rule on the brief that comes back.