Testing & evidence

Does it find what is really there?

We generated 100 synthetic companies with problems planted at random strengths, ran SignalPath on each, and counted how often it found them. Run on 8 Oct 2026 (engine p2-2026-10-06).

• New hires leaving early in one small team (about 10 hires a year)Below the target
  Found in 43% of companies · target 75%

  With this few people, even the best possible single test is convinced only 50% of the time. SignalPath shows it as a weaker hint (“Watch”) in 78% of companies, and staff check-ins add the “why”. Targets were set before the run and are never lowered.

• New hires leaving within 60 days in Customer SupportBelow the target
  Found in 79% of companies · target 80%

  Best possible with this few people: 88%.

• Sick leave clustering after rest daysMeets the target
  Found in 90% of companies · target 70%
• The week overtime jumped in the night shift (within 2 weeks)Meets the target
  Found in 100% of companies · target 85%
• The guess “young hires leave” found to be about the team, not ageMeets the target
  Found in 98% of companies · target 85%
• Festive-season sick leave labelled as seasonal, not as a problemMeets the target
  Found in 100% of companies · target 90%
• A broken clock-in machine flagged as missing dataMeets the target
  Found in 100% of companies · target 100%

Does it stay quiet when nothing is there?

100 more synthetic companies with nothing planted. Any Strong or Moderate finding there is a false alarm.

• Companies with at least one false “Strong” findingMeets the target
  3% of companies · target at most 5%
• Companies with at least one false “Moderate” or “Strong” findingMeets the target
  10% of companies · target at most 15%

The sample company gives the planted answers

• Web & App new hires leaving early: Moderate (5 of 8 left, against 5 of 21 across the rest of the company combined)
• Customer Support new hires leaving early: Moderate (3 of 7, against 1 of 25)
• Worker-months above the 104-hour overtime limit: 8 (8 planted)
• The clock-in outage in Fulfilment Day: flagged

Are the AI's answers checked?

Every AI answer passes automatic checks before anyone sees it (numbers only from the engine, no personal details, no blaming words). If a check fails, SignalPath uses rules or a template instead. Recorded runs:

• Staff comments in Malay, English and Chinese: the right theme found in 138 of 138.gemini-3.5-flash-lite · 7 Oct 2026
• Praise never counted as a problem: 21 of 21.gemini-3.5-flash-lite · 7 Oct 2026
• Without AI (keyword rules only): the right theme found in 122 of 138.gemini-3.5-flash-lite · 7 Oct 2026
• A comment trying to give the AI orders was flagged and not followed.gemini-3.5-flash-lite · 7 Oct 2026
• Investigation explanations passing every check on the first try: 10 of 10.gemini-3.8-flash · 13 Oct 2026
• After one retry: 10 of 10 (none fell back to the template).gemini-3.8-flash · 13 Oct 2026
• “Test a guess”: the boss's own words mapped to the right test in 4 of 4 live checks, including one that tried to list everyone's MC (refused).gemini-3.5-flash-lite · 8 Oct 2026
• Ask SignalPath: questions in English, Malay, Chinese and Manglish read into the right query in 10 of 10 live checks (after one fix), and all three questions about people or protected traits refused.gemini-3.5-flash-lite · 8 Oct 2026
• Ask SignalPath answers that passed every check (numbers only from the result): 7 of 7.gemini-3.8-flash · 8 Oct 2026
• Exit chat: the follow-up questions and summaries for 3 leavers, in three languages, all passed the checks.gemini-3.5-flash-lite · 8 Oct 2026

The browser tests below run with AI switched off, so every feature is shown to work without it.

Automated tests

408 unit tests passed on 25 Oct 2026, and 194 browser tests on 25 Oct 2026, each in a laptop-sized and a phone-sized window.

• Privacy: groups under 5 people (and the rest of a group, when that would be under 5) are never shown, and there is no page about one person.
• Every statistic against reference values.
• The four messy sample files (Malay headers, title rows, a fake IC column) map themselves and rebuild the original data exactly.
• The 90-second sample journey from finding to verified result, in English and Malay.
• Every main page in Chinese as well as English and Malay, on a laptop and a phone, with no text left untranslated.
• Check-ins: one answer per phone per month, links that work once, a follow-up only when the person asks.
• No page scrolls sideways on a phone.

Does it improve the process?

Inside the product, every decision is tested: SignalPath checks the numbers 1, 3 and 6 months after an action, against the teams that didn't change, and says honestly when it is too early or too few to tell. In the sample, the months after 1 Oct 2026 are simulated and labelled as such, so the improvement shown there is an assumption, not a result.

Not tested yet

• With a real company's records and real staff check-ins.
• With bosses and HR using it week after week.
• The Malay and Chinese texts were drafted with AI help and still need a review by native speakers.
• The synthetic companies show the engine works as designed, not that real companies behave like them.