WhyLab Studio — Connected Investigation Workspace

STAGE 01/INGEST — Evidence & Log Acquisition

EVALUATION CSV · LOGS · PRESETS

01 / Investigation workspace

CLIENT-SIDE EVALUATION · ZERO DATA RETENTION

START WITH THE EVIDENCE

What went wrong?

Bring your logs, metrics, or dataset. Let’s connect the dots.

Analysis type Choose your investigation lens

Client-side processing · Your evidence never leaves your browser

· CASE FILE / 001SAMPLE INVESTIGATION

Great in validation.
Lost in production.

An image classifier with a real-world reality check.

Computer visionResNet-1830 epochs

Validation accuracy94.2%Looking good in the lab

Production accuracy61.8%A different story outside

ACCURACY ACROSS ENVIRONMENTS30 epochs

● Validation● ProductionIllustrative data

↘

32.4 percentage points. One important question.
What changed between validation and the real world?

3 hypotheses. A path to understanding.

04 / EXTENDED EVIDENCE IMPORT

Bring the whole experiment.

Select up to five related files (50 MB total). Map unfamiliar CSV headers to metric names. Files are analyzed independently; the first available epoch history supplies the chart. Parsed results replace the current investigation and invalidate its verification history.

Evidence files

STAGE 02/PROFILE — Data Ingestion & Drift Check

COLUMN TYPES · DISTRIBUTIONS · CLASS FREQUENCY

03 / DATASET INVESTIGATION

Look inside your data.

Profile a training CSV, then add validation or production data to compare. All analysis stays in your browser. Supports Kaggle, OpenML & platform datasets up to 100 MB, 500,000 rows, and 300 columns.

⚡Quick-Load Verified Multi-Split Experiment:

Training dataset

Empty

Baseline reference for column types, class distribution, and data profiling.

Choose Training CSVDrag & drop or browse

Validation dataset

Empty

Hold-out dataset to evaluate generalizability and detect distribution shift.

Choose Validation CSVDrag & drop or browse

Production dataset

Empty

Live inference stream to monitor real-world drift and novel anomalies.

Choose Production CSVDrag & drop or browse

No training dataset uploaded

Add a training CSV to begin profiling. Validation and production files are optional for comparative drift checks.

COMBINED DIAGNOSIS

Build the case.

Combines the completed investigation with dataset evidence. Priority reflects review order, not probability or a confirmed cause. User answers are context, not independently verified evidence.

0 hypotheses supported by current evidence.

No supported hypothesis yet. Complete a log investigation or add a training dataset. Absence of a triggered rule does not establish model health.

STAGE 03/INVESTIGATE — Autonomous AI & Case Diagnostics

ASTRA AGENT · DETERMINISTIC LOCAL PROTOCOL

AUTONOMOUS INVESTIGATION

Investigate with Astra

CHECKING CONNECTION

Give Astra evaluation predictions. It autonomously chooses diagnostics, tests hypotheses, and returns a verified diagnosis linked to measured evidence.

✨ Instant Demo · No Token RequiredExplore a fully worked Astra investigation

See Astra’s multi-step autonomous diagnostic reasoning, hypothesis tests, and verified evidence graph without setting up an API token or incurring OpenAI costs.

Evaluation Evidence & Baseline

Evaluation CSV files (up to three)

Required columns: y_true, y_pred, y_probability. Probability must refer to the positive label. Up to 2 MB per file and 20,000 rows total.

Or paste evaluation CSV

Training / log artifact (optional)

Or paste training / log artifact

Bounded TXT/LOG/JSON/CSV artifacts up to 16 KB and 200 lines. Astra receives source-linked raw observations plus parsed epoch history.

Positive labelNegative labelAccuracy-paradox gap (pp)

Investigation objective

Astra Specialist Role (§8)Delegates specialized diagnostic tests while maintaining a single auditable investigator loop

Or select specialist lens:

Mid-Investigation Steering & Priorities (§7, §116–117)Steer Astra's diagnostic priorities and domain-specific cost objectives

Quick Directives:

Keep a local draft in this browser

Deployment access token

Required for paid Astra requests. Kept only in this page session and sent to the same-origin server.

Restore 24-hour investigation ID

I agree to send diagnostics and logs for Astra analysis.

What data is sent?

I agree to send these CSVs to the WhyLab server and their metadata, diagnostic results, and training-log text to OpenAI for diagnosis.

⚡Interactive Mid-Turn SteeringREADY TO STEER

Redirect Astra mid-flight without losing completed evidence or corrupting diagnostic state. Click a quick directive below or type a free-text instruction to adjust investigation priorities.

Observable Activity Stream

Multi-Agent Specialist Swarm

Only executed actions appear here.

STAGE 04/VERIFY — Causal Evidence DAG & Falsification Tests

EVIDENCE GRAPH · COUNTERFACTUAL TESTS · LESSONS

No active investigation findings to verify yet

Run an autonomous Astra investigation in Stage 03, or load an evaluation dataset in Stage 01 to view the causal evidence graph, reliability profile, and counterfactual tests.

STAGE 05/REPAIR — Decision Threshold Sweeps & Policy Optimization

COST MODELS · CONFUSION MATRIX BEFORE/AFTER

MEASURE → REVIEW → APPLY

Repair Lab

State your objective and error costs, inspect the operating-point trade-offs, then apply a recommendation to an evaluation copy.

Repair Evidence & Cost Matrix

Repair evaluation CSV

Binary labels 1/0; probability refers to label 1. Paste up to 2 MB. Baseline threshold must reproduce the supplied predictions.

Repair objective

False-negative costWeight for missing true positive cases (e.g. 50x)

False-positive costWeight for false alarms (e.g. 1x)

Maximum total costTarget acceptance ceiling

Baseline thresholdCurrent decision boundary (default 0.5)

Numeric costs are authoritative only after you enter or confirm them; Astra proposals never apply automatically. Total cost = FN × false-negative cost + FP × false-positive cost. Sweep 0–1 in steps of 0.01, retaining the baseline; choose minimum eligible cost, highest threshold on ties.

Deployment access token (optional for demo)

Required for live OpenAI API requests. If left blank, verified recorded demonstration values are used.

I agree to send repair inputs for Astra policy translation.

What data is sent?

I agree to send the CSV to the server. OpenAI receives the objective and bounded dataset summaries for policy translation, or the confirmed cost policy, measured summaries and linked diagnosis context for repair.

STAGE 06/REPORT — Enterprise Governance & Audit Synthesis

COMPLIANCE GATE · INCIDENT EXPORT · PROVENANCE

Enterprise Model Governance & Executive Synthesis

Audit Ready

Produce certified verification artifacts, evaluate production deployment safety gates, and export reproducible JSON/Markdown audit evidence with full counterfactual test provenance.

01 / Investigation workspace

What went wrong?

Great in validation.Lost in production.

Bring the whole experiment.

Look inside your data.

Training dataset

Validation dataset

Production dataset

Build the case.

Investigate with Astra

Observable Activity Stream

No active investigation findings to verify yet

Repair Lab

Enterprise Model Governance & Executive Synthesis

Great in validation.
Lost in production.