Evidence-based telemetry. Three specialized intelligence layers.
A transparent architectural breakdown of how Praxys captures permitted client-side interaction signals, normalizes the evidence, and generates proportionate decisions without compromising privacy.
End-to-End Evidence Pipeline
Grounded in the real platform architecture: client-side collection, normalization, specialized engines, and tenant dashboards.
How Interaction Evidence Becomes Actionable
Every decision produced by Praxys can be traced back to verified interaction signals.
Permitted Interaction Telemetry
Captured via an opt-in Manifest V3 browser extension and permitted IDE plugins. Measures client-side keystroke inter-arrival timing, edit delta sizes, and prompt submission cycles.
Canonical Normalization & Provenance
Raw interaction events are sanitized and formatted into standard schema records. Metadata (timing, frequency, ratios) is cleanly decoupled from prompt text.
Specialized Analysis & Evidence Evaluation
Dispatches normalized evidence to the three purpose-built engines: Assurance (regex & entropy scanning), Collaboration (cadence & loop detection), and Capability (autonomy & offloading scoring).
Findings, Uncertainty & Decisions
Translates analytical scores into contextual findings. Scanner matches are clearly separated from confirmed security breaches; friction indicators are reported with confidence boundaries.
Proportionate Intervention & Feedback
When policy thresholds or offloading anomalies are detected, Praxys returns non-blocking, guided feedback (such as a pre-dispatch warning or Socratic reflection prompt).
Signal Taxonomy: What We Measure vs What We Infer
We maintain an honest boundary between what is directly observed, what is a derived indicator, and what requires external validation.
| Measurement Level | Technical Nature | Concrete Example | Explicit Limitation |
|---|---|---|---|
| Observed Signal | Deterministic Raw Measurement | Keydown event timestamp delta (e.g. 142ms), paste string length (e.g. 480 chars). | Raw signals convey timing, not intent or cognitive comprehension. |
| Derived Feature | Statistical Synthesis | Burst velocity, prompt revision cycle count, review dwell time prior to code execution. | Reflects descriptive pacing; does not measure overall task complexity. |
| Behavioral Proxy | Calibrated Analytical Index | Interaction Strain Index, Cognitive Offloading Ratio, Autonomy Score. | Useful for trends and coaching; must not be treated as a definitive psychometric assessment. |
| Validated Outcome | Empirically Verified Impact | Long-term skill retention, independent debugging efficacy, validated policy compliance. | Demands multi-month academic cohort studies; Praxys does not claim causal outcomes prematurely. |
Strictly Defined Privacy Boundaries
Metadata Mode (Default): Core telemetry measures interaction physics—timing, revision cadence, iteration loops, and delta counts. Prompt text is neither captured nor transmitted to backend servers.
Deterministic & API-Level Processing: Pattern scanning, credential detection, and friction indicators run via deterministic client-side logic and rule-based evaluators. Where generative synthesis is required (such as drafting Socratic nudges or parameterizing reusable templates), Praxys utilizes existing commercial model APIs under strict zero-data-retention agreements. We do not train proprietary foundation models on customer prompts.
Explicit Consent Boundaries: Richer content evaluation (such as Assurance regex scanning or deep reflection context) is strictly gated behind affirmative user or administrative policy consent.
No Surveillance: Praxys does not monitor webcam activity, capture background desktop screenshots, log raw keystrokes in non-AI applications, or deploy covert endpoint agents.
Privacy by Design
Interaction physics are analyzed locally. Raw prompt payloads are isolated by default.