πOne Step Ahead (Concepts)
Three differentiation-axis concepts (never trust the voice / graduating to the edge / police-report package) + deepfake-detection candidates. Everything β³ concept stage.
β³ Concept stage
The attackers (scams) have finished going AI (voice cloning, fraud-as-a-service). The industry competes on AI that answers calls; AI that doubts calls is an empty space β these three differentiation axes aim at that space. **Everything on this page is at the concept stage**: nothing is implemented or verified yet (per our conventions, we never write "it can do X").Even one step before this siteβs usual rule (distinguishing βimplementedβ from βenabled by defaultβ) β
everything on this page is at the concept stage. Nothing is implemented or device-verified yet, so we
write βthe idea is toβ¦β / βaims atβ¦β and never βit can do Xβ.
Source of truth: aegis-platform docs/specs/ROADMAP_voice_patterns_2026-07-12.md and docs/specs/DESIGN_verification_layer_2026-07-12.md.
One step ahead in differentiation (3 axes)
Axis 1: A design that never trusts the voice (behavioral verification)
aegis-platform docs/specs/DESIGN_verification_layer_2026-07-12.md The idea: protect with behavioral verification, not by judging whether a voice is genuine β designed on the premise that voice cloning has already crossed the indistinguishability line.
Axis 2: Graduating to the edge (distillation lands on the Pi)
aegis-platform docs/specs/ROADMAP_voice_patterns_2026-07-12.md The idea: the graduation target of distillation (stage 4) is on-Pi verdicts β something cloud SaaS structurally cannot copy (the device is the moat).
Axis 3: Proof of protection β a police-report package
aegis-platform docs/specs/ROADMAP_voice_patterns_2026-07-12.md The idea: store the record of every repelled or encountered scam call (caller number, timestamp, transcript, verdict rationale, tactic classification) with tamper-evidence, and export it in one tap in a format usable for a police report.
Inside Axis 1: the 3 principles and the trigger β³ Concept stage
- Principle 1 β never trust the voice: who the voice sounds like is never grounds for identity verification
- Principle 2 β never used to let calls through: verification only clears suspicion. Even on success, the scorer's verdict is never relaxed (no bonus points). Failure or non-completion always falls to the safe side (a human)
- Principle 3 β fail-safe: no failure in the verification layer may break existing call handling
- Trigger: only when an "urgency signal" AND a "money / personal-info signal" both fire (either alone over-triggers; voice similarity and the caller's claims are never trigger conditions)
Inside Axis 1: the 3 verification steps (V1 β V2 β V3, tried top-down) β³ Concept stage
V1: In-conversation verification (passphrase)
The AI weaves a pre-registered passphrase or a question only the real person can answer naturally into the conversation. Correct β a V record is kept and the normal flow continues (the scorer's verdict is still never relaxed). Wrong or evasive β escalate to V2.
V2: Out-of-band verification (call back the registered number)
Declare "we will call your registered number back to confirm," hang up, then dial the registered number and confirm the facts with the real person. A confirmation call to our own customer, so no brand contradiction. As the first outbound feature it presupposes reusing the outbound-restriction design plus legal review.
V3: Human fallback
If V1/V2 cannot complete (nothing registered, no answer), guide with "please confirm directly with your family" and end on the safe side. Rollout order: V3 (guidance only, minimal) β V1 (passphrase) β V2 (call-back).
Axes 2 & 3 in brief β³ Concept stage
- [Axis 2] Aims at zero marginal cost (no cloud inference fees)
- [Axis 2] Aims to keep working even when the line is down
- [Axis 2] Aims at a topology where audio never leaves the house (a structural privacy advantage)
- [Axis 3] Business value (a): counters "nothing ever happens = churn" by making the protection visible
- [Axis 3] Business value (b): positions us against victims' silent resignation (today, neither the recording nor the number survives)
- [Axis 3] Business value (c): extends into a family dashboard and monthly reports
- [Axis 3] Tamper-evidence (hash chain or similar; the method is chosen at implementation time). The verification log (Axis 1) doubles as the raw material β defense and evidence generation in one
- [Axis 3] Prerequisite: legal review of the recording notice. Acceptance is up to police-side procedure, so we say "a format usable for a report" (we never promise acceptance)
A legal review of the recording notice comes first. Acceptance is up to police-side procedure, so we say βa format usable for a reportβ and never promise acceptance.
Deepfake-detection candidates (external tools) β³ Concept stage
Figures and claims below are transcribed from vendor statements and independent test reports. Measurement against our own gold set is still ahead (eval first).
| Tool | Target | Published / reported strengths | Pricing |
|---|---|---|---|
| π’ Pindrop Pulse | Real-time, for call centers | Claims ~2-second verdicts, 1,300+ voice features, >90% detection even on unseen synthesis; offers a detection warranty program | Enterprise pricing |
| π§ͺ Reality Defender | Developer-facing RealAPI (RealCall available for phone) | Free tier of 50 scans/month β the entry point for offline checks of recorded data, measuring real phone-quality performance against our own gold set | Usage-based |
| π¬ Resemble Detect | Developer API | Reported ~94% detection on clean audio | Usage-based API. Accuracy degrades at phone quality |
| 𧩠OSS η³» | Research use | Independent tests report ~78% plus false positives β not keeping up with commercial synthesis | Free (accuracy concerns) |