Sign in Start testing
Softsurance/AI investigationPreview

What it saw, and what it only suspects.

When a test fails, Softsurance gathers the evidence it actually captured and produces a structured report that keeps observed facts and hypotheses apart.

Start testing Read the docs

At a glance

InputsCaptured evidence only
OutputValidated schema
ProviderConfigurable
StatusPreview

Investigation needs a configured AI provider. Without one, execution and evidence work as normal, and the product says plainly that investigation is unavailable.

Observed

Facts from the execution

Assertion
Checkout button not visible after 5000 ms.
Network
GET /api/cart/summary returned HTTP 500.
Console
No errors recorded.
History
Same test passed in the previous run.
Suspected

Hypothesis, with alternatives

Likely cause
The failing cart summary request may prevent the checkout button from rendering.
Alternative
The button may have been renamed or moved in a recent change.
Confidence
Moderate
Next step
Check server logs for /api/cart/summary around the run time.

01 / Capabilities

What it checks9 items

What it checks.

Everything a run checks, and the evidence it keeps when something fails.

  • Failed assertion detailsPreview
  • Browser console outputPreview
  • Failed requests and status codesPreview
  • Screenshot and DOM contextPreview
  • Execution timelinePreview
  • Prior runs of the same testPreview
  • Cached reports, invalidated on new evidencePreview
  • Application logsRoadmap
  • Commit and deployment metadataRoadmap

02 / How it works

Step by step4 steps

  1. Collect

    The failed item’s assertions, console, network, screenshot reference, DOM excerpt and recent history are gathered — with secrets redacted.

  2. Separate

    Page content is passed as untrusted data, isolated from our instructions, so a page can’t tell the model what to conclude.

  3. Analyse

    The model returns a report in a fixed schema. Anything that doesn’t validate is rejected, not patched.

  4. Store

    Reports are saved with the provider, model and evidence they used. Rerun any time; new evidence invalidates the cache.

03 / Guardrails

Enforced by the backendNot optional

The limits come first.

These rules are enforced by the execution backend. The interface can’t switch them off, and neither can an AI.

G-01

Facts and hypotheses apart

A cause is never stated as fact unless the evidence shows it.

G-02

Nothing invented

No made-up stack traces, logs, commits or code references. If it isn’t captured, it isn’t cited.

G-03

Injection-aware

Instructions inside inspected pages are treated as data and ignored.

G-04

Inconclusive is allowed

When the evidence doesn’t support a conclusion, the report says so.

Report

The structure8 parts

Every report has the same eight parts.

A fixed structure, validated before it’s saved, so reports are comparable and nothing is quietly invented.

01

Summary

One paragraph on what failed.

02

Observed facts

Only what the evidence shows.

03

Suspected cause

Labelled as a hypothesis.

04

Supporting evidence

Linked to the captured artifacts.

05

Alternatives

Other explanations that fit.

06

Confidence

Low, moderate or high, with reasons.

07

Next steps

What to check to confirm.

08

Related failures

Only when evidence links them.

04 / Questions

AI investigation3 entries
Q.01Which AI providers work?

Any provider behind our interchangeable interface, configured by environment variable. Model choice is configurable.

Q.02What if I don’t configure one?

Everything else works. Investigation shows as unavailable with instructions to enable it.

Q.03Does my data train a model?

That depends on the provider you configure and its terms. We send only the redacted evidence for the failed item.

Start

Your first runStaging URL required

Add a staging URL. Read your first real result today.

Start testing Read the docs