01
Observe
Read the visible prompt, response, platform and displayed model label.
Measurement methodology
Every number has a defined meaning.
Veritiana separates directly observed facts, calculated estimates and interpretation. The dashboard never presents an estimate as provider billing.
Observed browser facts
Explicit calculations
Confidence shown separately
One measured browser turn
Visible browser evidence → estimates → classification → comparison value
Versioned method
Observed
Visible prompt + response
Content actually rendered in the supported browser surface.
Veritiana records what was visible, not whether the response itself was factually correct.
Estimated
Visible tokens + reconstructed context
Estimated from visible turns and conversation history.
It does not represent the provider's hidden internal context or billing record.
Classified
Task + complexity
Produced locally and assigned only when the relevant confidence threshold is met.
Task threshold
≥ 0.72
Complexity threshold
≥ 0.66
Calculated
Estimated API-equivalent cost
Estimated input and output token volumes are multiplied by the selected versioned model rates.
The result is a comparison estimate, never provider billing.
Method version
v1.0
Classifier version
Recorded per event
Tokenizer basis
Provider-aware estimate
Price basis
Versioned model rates
Each event preserves the method, classifier and price basis used at the time of measurement, so historical records remain interpretable even when the methodology changes.
Reproducible historical interpretation
Calculation flow
From visible browser content to a defensible estimate.
Each stage produces a distinct output. The dashboard can therefore show what was observed, what was estimated and what was inferred.
02
Estimate
Estimate visible input, output and accumulated context tokens.
03
Classify
Assign task, complexity and independent confidence values locally.
04
Value
Map visible usage to a versioned API-equivalent price model.
05
Interpret
Compare measured value, subscription price, coverage and confidence.
Token estimation
Visible tokens are estimates, not provider counters.
Veritiana estimates the text visible in the browser. It cannot see hidden system prompts, hidden reasoning, cache behavior or provider-side routing.
Visible input estimate
Estimated from the user message visible in the measured browser turn.
Visible output estimate
Estimated from the rendered AI response available to the extension.
Visible context estimate
Derived from the visible conversation trajectory available at the time of measurement.
Explicit coverage limit
The estimate covers browser-visible activity only and says so in the dashboard.
Local classification
Task and complexity are separate predictions.
The classifier does not compress every interpretation into one label. Task, complexity and confidence remain separate fields.
Task
What kind of work?
Examples include general chat, writing, translation, summarization, research, coding, mathematics, document analysis and high-stakes work.
Complexity
How demanding?
Complexity is estimated independently as low, medium or high. It is not inferred directly from task type.
Confidence
How certain?
Low-confidence results remain visibly uncertain. They are not converted into definitive claims.
Classifier training
How the local classifier was trained.
Veritiana uses a compact local ONNX classifier to interpret visible AI interactions by task and complexity. Classification runs inside the browser. Raw conversation content is not uploaded for classification.
Training data
The current release was trained from normalized prompt data derived from OASST1, CoEdIT, MBPP+ and IFEval, then filtered, balanced and deterministically augmented for minority classes.
Two outputs
Task and complexity are predicted separately. Complexity is not inferred directly from task type.
Confidence gates
Task is assigned at confidence ≥ 0.72. Complexity is assigned at confidence ≥ 0.66.
API-equivalent value
A comparison model, not the provider’s internal cost.
The estimate answers one practical question: what could similar visible usage cost through a public API pricing model?
Inputs
Visible token estimates, displayed model label and the pricing version active at measurement time.
Calculation
Estimated input and output quantities are mapped to versioned public price assumptions.
Meaning
The result is an API-equivalent comparison value, not billing, profit, compute cost or provider margin.
Subscription verdict
The verdict is conditional, not absolute.
A subscription comparison must include measured value, browser-only coverage and confidence. Without those limits, the verdict would be misleading.
Subscription paid
€20.00
Measured value
€8.40–€11.70
Coverage
Browser only
Verdict
Probably overpaying
Confidence: Medium
The wording remains probabilistic because unmeasured mobile, desktop-app or unsupported usage may materially change the comparison.
Method limits
What the methodology deliberately does not claim.
No hidden reasoning
The method does not infer private chain-of-thought or hidden model reasoning.
No provider internals
Caching, routing, batching and internal compute allocation remain unknown.
No billing replacement
API-equivalent value does not replace official billing or subscription invoices.
No native-app coverage
Mobile and desktop applications are outside browser-extension coverage.
No false precision
Ranges and confidence labels are used where a single exact number would mislead.
Versioned methodology
Pricing, classifiers and calculation rules are versioned so historical events remain interpretable.