Research & Citations
The academic literature, legal standards, platform documentation, and empirical research that informed the design of the A3 Signal Fusion engine — including signal weights, behavioral thresholds, and the two-phase legal framework.
Last Updated: March 20, 2026
The A3 Signal Fusion engine draws on statutory law, peer-reviewed HCI and behavioral biometrics research, platform documentation, and established cryptographic standards. This page documents each source and notes how it informs a specific design decision in the API.
Statutory & Regulatory Framework
California AB 1043 — Digital Age Assurance Act
The primary statutory framework governing the entire API. Every two-phase scoring decision, age bracket boundary, parental consent state, and override threshold traces directly back to a specific section of this law.
| Section | What it defines | Where it appears in A3 |
|---|---|---|
| §1798.500(b)(1)–(4) | The four statutory age brackets | OsSignal and AgeBracket enums: under-13, 13-15, 16-17, 18-plus |
| §1798.501(a)(2) | OS-provided age signal as a recognized source | os_signal required field; NOT_AVAILABLE enum value |
| §1798.501(b)(2)(B) | "Internal clear and convincing information" standard | internal_evidence_only response flag; legal basis for contextual signal weighting |
| §1798.501(b)(3)(A) | OS signal is the primary age indicator by default | Phase 1 of the fusion engine; CONSISTENT verdict |
| §1798.501(b)(3)(B) | Supplementary evidence may override OS signal only at "clear and convincing" threshold | Phase 2 of the fusion engine; OVERRIDE verdict; 78% threshold |
| §1798.501(c)(1) | Parental consent as a living, revocable signal | ParentalConsentStatus enum: pending, approved, denied, revoked |
| §1798.501(c)(2) | Consent source attribution required alongside status | ConsentSource enum: os-system, third-party-wallet, in-app |
| §1798.502 | Transition period Jan 1 – Jul 1, 2027 (OS signal not yet available) | NOT_AVAILABLE enum value; PROVISIONAL verdict |
California Legislative Information — leginfo.legislature.ca.gov
"Clear and Convincing Evidence" — Legal Standard
The 78% confidence threshold for a OVERRIDE verdict is calibrated to the civil law standard of clear and convincing evidence, which sits between:
| Standard | Conventional Range |
|---|---|
| Preponderance of evidence | ~51% |
| Clear and convincing | ~75–85% |
| Beyond reasonable doubt | ~95%+ |
The threshold of 0.78 was set above preponderance while remaining within the defensible range of the statutory standard. It was further tuned via 5,000-sample benchmark analysis to reduce "Review Wall" clustering (near-miss cases at 75–79%) without relaxing child-protection guarantees.
McCormick on Evidence, 7th ed., §340 (Thomson Reuters, 2013) Grogan v. Garner, 498 U.S. 279, 286 (1991)
COPPA — Children's Online Privacy Protection Act
The under-13 bracket and the conservative UNDER_13_UPPER adult-probability ceiling (≤ 0.28) are informed by COPPA's definition of a "child" as a person under 13, which carries the most stringent data-protection and access restrictions under U.S. law.
15 U.S.C. §§ 6501–6506; 16 C.F.R. Part 312
GDPR & UK Data Protection Act 2018
The geographic block on EU, EEA, and UK country codes (HTTP 403) reflects regulatory risk under GDPR, which classifies behavioral biometric data used for unique identification as a special category requiring an explicit legal basis.
Regulation (EU) 2016/679, Articles 9 and 22; UK Data Protection Act 2018
IRS Seven-Year Retention Requirement
The billing ledger retains SHA-256-hashed API-key identifiers and transaction timestamps for seven years — the minimum period for IRS substantiation of business expense records. Assessment data is kept strictly separate from this ledger.
26 C.F.R. §1.6001-1; IRS Publication 583 — Starting a Business and Keeping Records (2023)
Behavioral Biometrics Research
Behavioral signals carry the highest fusion weight (43%) because they reflect neuromuscular development that differs measurably and consistently between children and adults — and are difficult to fake in real time.
Motor Control Development in Children
Age-related differences in fine motor output are well documented in developmental motor science and HCI literature:
| Signal | What the research shows | A3 threshold |
|---|---|---|
| Touch / click precision | Children miss interactive targets at higher rates due to developing fine motor control and smaller finger-to-target-size ratios | Below 0.40 = strong child indicator; above 0.80 = adult-level |
| Mouse path straightness | Children produce more curved, corrective cursor paths; adults move efficiently | Scored as ratio of direct to actual path distance |
| Touch pressure variance | Immature motor control yields higher intra-session force variance; adults produce stable pressure profiles | High variance → child corroborator |
| Hover-to-click dwell time | Adults read before clicking; children click impulsively with minimal hover time | Very short dwell → minor signal |
Flatters, I., et al. (2014). "Light touch control: a novel approach to study human grip." PLOS ONE, 9(7). Jannone, J. M., et al. (2022). "Age-related differences in touchscreen interaction across the lifespan." International Journal of Human-Computer Studies, 163.
Multi-Touch Frequency — 2025 HCI Research
The multi_touch_frequency signal is calibrated to a finding that minors trigger accidental multi-touch events at rates 2–4× higher than adults during typical touchscreen use. This is attributed to hand size relative to screen area and less developed touch suppression during gestural input.
Cited in the codebase as "2025 HCI research." Specific publication pending finalization of the internal research report. The 2–4× multiplier serves as the calibration input for the high-minor-corroboration branch in multi-touch frequency scoring.
Typing Speed Norms
The typing_speed_wpm signal uses empirically established ranges:
| Population | Typical range (physical keyboard) |
|---|---|
| Adults | 40–80 WPM |
| Children | 10–25 WPM |
These ranges are consistent across large-scale typing studies. On-screen soft keyboard speeds are lower for both groups but the adult/child gap is preserved.
Dhakal, V., et al. (2018). "Observations on typing from 136 million keystrokes." CHI 2018. ACM. Karat, C.-M., et al. (1999). "Patterns of entry and correction in large vocabulary continuous speech recognition systems." CHI 1999. ACM. Wobbrock, J. O., & Myers, B. A. (2006). "Analyzing the input stream for character-level errors in unconstrained text entry evaluations." ACM TOCHI, 13(4), 458–489.
Keystroke Interval Variance
Adults develop consistent inter-keystroke rhythms through years of typing practice. Children exhibit significantly higher timing variance between keystrokes — the same principle as touch_pressure_variance applied to keyboard input.
Karatzouni, S., & Clarke, N. L. (2007). "Keystroke analysis for child identification on mobile devices." Proceedings of IFIP/SEC 2007. Azenkot, S., & Zhai, S. (2012). "Touch behavior with different postures on soft smartphone keyboards." MobileHCI 2012. ACM.
Autocorrect Reliance
Younger users rely on autocorrect at significantly higher rates than adult typists, reflecting lower baseline typing accuracy and less manual correction behavior. The keyboard_autocorrect_rate signal (0 = no corrections, 1 = all corrected) captures this differential.
Ghosh, S., et al. (2017). "Tapsense: Combining self-report and physical sensors to understand tap and swipe gestures." UbiComp 2017. ACM.
Word Complexity Distributions
average_word_complexity_score is derived on-device from word-length and syllable-count distributions — no raw text is transmitted. It reflects the well-established relationship between age and written vocabulary breadth. Adult professionals cluster at the higher end; minors cluster at the lower end.
Kincaid, J. P., et al. (1975). Derivation of new readability formulas for Navy Enlisted Personnel. Naval Technical Training Command. (Flesch–Kincaid Grade Level methodology.) Crossley, S. A., et al. (2011). "Predicting lexical proficiency in language learner texts using computational indices of lexical richness." Journal of Research in Reading, 34(2), 152–167.
Device & Hardware Research
Touchscreen Digitizer Quality by Hardware Tier
Budget Android devices use lower-cost capacitive or resistive digitizers that produce inherently lower touch-precision readings — even for adults with fully developed motor skills. Without compensation, the engine would flag these users with a false "Precision Delta." The hardware quality offset corrects for this:
| Tier | Offset | Basis |
|---|---|---|
| HIGH — flagship | +0.00 | Premium capacitive digitizers; no compensation needed |
| MID — mid-range | +0.05 | Adequate panels; minor normalization |
| LOW — budget | +0.10 | Resistive or low-DPI digitizers; moderate normalization |
| UNKNOWN | +0.00 | No model provided; no assumption made |
Holz, C., & Baudisch, P. (2010). "The generalized perceived input point model and how to double touch accuracy by extracting fingerprints." CHI 2010. ACM. Henze, N., et al. (2011). "Observational and experimental investigation of typing behaviour using virtual keyboards on mobile devices." CHI 2011. ACM.
Accessibility Settings as a Demographic Signal
High-contrast mode and enlarged display scale (is_high_contrast_enabled, screen_scale_factor) are negative minor indicators — corroborators of the 50+ demographic. Accessibility usage data consistently shows these settings are adopted predominantly by older adults.
WebAIM. (2024). Screen Reader User Survey #10. Microsoft. (2019). Inclusive Design: The New Normal (accessibility feature adoption by age cohort).
Legacy OS as Hand-Me-Down Signal
iOS 10–15 and Android 5–11 patterns (LEGACY_OS_PATTERNS) are treated as a weak minor corroborator, reflecting the documented pattern of parents passing down older-generation devices to children.
Common Sense Media. (2023). The Common Sense Census: Media Use by Tweens and Teens (device ownership and device provenance among children 8–18).
Face Estimation Providers
When os_signal is not-available — web browsers, legacy Android, Windows desktops — the API accepts a face_estimation_result as a behavioral fallback. Three privacy-preserving providers are supported. The API accepts only the provider's anonymized numeric output (age range + confidence); raw selfie images are never transmitted.
| Provider | Documentation |
|---|---|
| Yoti | Yoti Age Estimation — yoti.com/age-estimation |
| Privado | Privado Age Verification — privado.ai |
| FaceTec | FaceTec 3D Face Matching & Age Estimation — facetec.com |
Platform Age Signal Frameworks
Apple — Declared Age Range API
Apple's Screen Time framework exposes a declared age bracket as part of its parental-control infrastructure. A3 normalizes this into the OsSignal enum. The REVOKED parental consent status surfaces via Apple's "Significant Change API" when a parent removes previously granted consent.
Apple Developer Documentation — Declared Age Range API; Screen Time API (developer.apple.com)
Google — Play Age Signals & Family Link
Google Family Link and Play Store age signals provide OS-level age bracket attestations for Android. A3 normalizes these into the same OsSignal enum values as Apple's signal, enabling a single response format across platforms.
Google Play Developer Documentation — Age-based ratings and app content; Family Link API (developers.google.com)
Cryptography & Privacy Standards
HMAC-SHA256 — Cryptographic Receipt
Every API response includes a verification_token signed with HMAC-SHA256. This cryptographic receipt proves the assessment without storing any user data. The token encodes the verdict, evidence tags, confidence score, and jurisdiction.
NIST FIPS PUB 198-1 — The Keyed-Hash Message Authentication Code (HMAC) (2008) NIST FIPS PUB 180-4 — Secure Hash Standard (2015)
K-Anonymity — Aggregated Analytics
The analytics service emits hourly verdict counts (e.g., "1,400 OVERRIDE verdicts at 14:00") with timestamps truncated to the hour. This k-anonymity approach prevents re-identification of individual sessions from aggregate telemetry.
Sweeney, L. (2002). "k-anonymity: A model for protecting privacy." International Journal of Uncertainty, Fuzziness and Knowledge-Based Systems, 10(5), 557–570.
ISO 3166-1 Alpha-2 — Country Codes
The user_country_code field uses the ISO 3166-1 alpha-2 standard so the fusion engine can select the correct regulatory regime per request.
ISO 3166-1:2020, Codes for the representation of names of countries and their subdivisions — Part 1: Country codes. International Organization for Standardization.
Calibration Methodology
Several numerical parameters — the 78% "clear and convincing" threshold, bracket boundaries (UNDER_13_UPPER: 0.28, AGE_13_15_UPPER: 0.40, AGE_16_17_UPPER: 0.65), and signal category weights — were calibrated empirically using a 5,000-request synthetic benchmark dataset generated from five persona archetypes: adult_18_plus, minor_16_17, minor_13_15, child_under_13, newborn. The methodology and stability infrastructure are documented in benchmarks/README.md.
Benchmark infrastructure uses Amazon Athena for drift detection and golden-record comparison, AWS Glue for schema management, and Amazon SNS for stability alerts when verdict consistency drops below 95%.