Accountability
Governance and red lines
Eight red lines
What Relacryx will never do
It never profiles a person by who they are.
Nationality, ethnicity, accent, appearance, location, an IP address or a VPN are never used to judge anyone. Only what is said, asked and paid counts.
It never suggests who does this.
Offenders and victims are of every gender, age and background. Wording, examples and images are chosen so that no gender looks like the typical offender or the typical victim, and no statistic here is used to suggest otherwise.
It never accuses.
A flag is a question to ask, never a verdict on a person. Genuine relationships and family support can look like scams, and every result says so.
It never acts by itself.
The system recommends and explains. A trained person decides, and records why. It does not block, report or contact anyone automatically.
It never blames.
No wording suggests a victim made a poor choice or is at fault. Blackmail is always the fault of the person making the threat.
It never keeps your answers.
The public check stores nothing, uses no cookies or trackers, and the site's security policy stops the browser sending anything out.
It never presents a model as evidence.
Modelled examples show how the system reasons. They are labelled as such everywhere, and never used to claim accuracy.
It never hides its sources.
Every figure, case and rule points to where it came from, and states what it cannot show.
Kept honest by tests
How these promises are checked
Automated checks on every release
A test suite confirms that no single signal raises the level alone, that adding signals never lowers it, that no wording blames a victim, that no signal profiles a person, that the source code contains no browser storage and no network calls, and that the decision trace detects any change to it.
Browser-enforced privacy
The site sends a content security policy that allows connections only to itself. A mistaken or malicious attempt to send data elsewhere would be blocked by the browser, not just discouraged by the code.
Self-run test results
What our own testing found, and what it does not prove
The public check was tested on hand-written profiles: 35 based on published descriptions of scams, and 33 of ordinary genuine relationships with no request for money. The profiles were written by the author of the rules, from the same sources, so the results show how the rules behave, not how they would perform on real people.
| Set | Scams reaching pause or stop | Genuine relationships wrongly flagged |
|---|---|---|
| First set (used while building) | 15 of 15 | 0 of 15 |
| Second set (its failures exposed a real bug, since fixed) | 10 of 10 | 0 of 8 |
| Blind set (written after the fix, run once) | 10 of 10 | 0 of 10 |
- The second set first showed the check calling intense but genuine relationships “a strong pattern” (5 of 8 wrongly flagged). We fixed the rule so that closeness alone never goes above “worth watching”, then wrote the blind set to check the fix honestly.
- Twenty-plus hand-written profiles are a small sample. Real relationships and real scams are far more varied.
- Two profiles that sit on the line (an investment tip with no request yet, and a one-off flight cost) are reported separately and not scored.
Limits
What this system cannot do
- It cannot confirm that anyone is, or is not, a scammer. It shows patterns that match published descriptions.
- Its rules come from official warnings, a six-firm regulatory review and academic work. They have not yet been validated on real data.
- It sees only what a person ticks. Someone who does not recognise a signal cannot be helped by it, which is why talking to another person always comes first.
- It does not replace the police, your bank, a lawyer or a counsellor.
Versions
What is published right now
- Engine
- 0.1.0
- Knowledge
- 2026.10.1
- Knowledge last read
- 2026-10-05
Every analysis and decision trace records these versions, so any result can be tied to the exact rules that produced it. Changes to the rules are logged and dated here.
Changelog
- 2026-10-05 · Version 0.1.0. First release: lifecycle model, public check, case files, console on modelled examples.