Math Challenge
More

Avatars, Identity and Progression for Children's Products Under Strict Privacy

mc-43 · Published: · by Math Challenge Research · 3,097 words · 19 cited sources

Executive summary

Serious children's products solve identity the same way: generated aliases (picked, not typed) plus an avatar of pre-made parts (never a photo) — Nintendo's Mii [11], and what the project already decided in D-003 [19]. The Proteus effect (an avatar changes the wearer's behavior) is replicated at a small-to-medium effect size [2][3], justifying treating it as a real lever. The Scunthorpe problem — filters blocking innocent words over an offensive substring — is the central technical risk of generating aliases in five languages [1][10]; the fix that works in production is a maintained whitelist, not a stricter filter. Belgium and the Netherlands ruled in 2018 that randomized reward boxes are gambling even when contents are cosmetic [4] — the line is "randomized-and-paid vs. deterministic," not "cosmetic vs. gameplay." A Tamagotchi-style companion drives real retention but also invented the guilt mechanic that made it famous [5]. Mayer's pedagogical-agent research backs a character-guide with caveats: the effect is modest and depends on social cues, not animation [9]; Sesame Street and Elmo parasocial research show a familiar character teaches better than a new one [6][7][8] — exactly Larry's case, since he already exists in Ignia's canon [18][19]. Roblox is the warning for unrestricted free text: 1,600 moderators plus AI, and inappropriate content still resurfaces [13].

212 words

This document was written in English. It is published here in full, unedited.

Verification status

This document carries no [unverified] flag. Every claim in it is tied to a numbered source below.

[unverified] means the claim is stated in the research but was not confirmed against a primary source in the session that produced it. It is published rather than removed, because a research corpus that hides its gaps is not verifiable.

How this research was produced

The 47 documents were produced on 2026-07-31 by independent agents, each instructed not to invent citations and to flag as [unverified] anything it could not confirm against a primary source. The session's web-search quota ran out mid-way, and later agents worked by direct fetch against primary sources. Several sites (ftc.gov, ico.org.uk) block automated fetching, which is why certain legal claims are flagged on purpose.

Findings

1. Identity without personal data: the industry pattern

Every major children’s/family product converges on the same two-part solution: a generated or curated identifier instead of a typed real name, and an avatar built from a closed set of parts instead of a photograph. Nintendo’s Mii is the clearest case: characters are built by selecting from pre-designed facial features, hairstyles, accessories, and body types; even the optional photo-based creation mode only seeds feature selection and never stores or displays the source photo, preserving “self-representation while maintaining privacy” [11]. Duolingo lets users build an avatar from a fixed wardrobe of seasonal cosmetic pieces rather than an uploaded image [12]. The common engineering shape is: identity = choice among a bounded, pre-vetted set, never free input — exactly what the project already picked in D-003 (generated aliases, no photo, no city, boards segmented by level) [19].

2. The Proteus effect: why the avatar is not just decoration

Yee and Bailenson’s 2007 Stanford research coined the “Proteus effect”: people in virtual environments behave in ways consistent with their avatar’s implied traits, driven by self-perception and behavioral confirmation rather than by what others actually do [2][3]. A meta-analysis of 46 experimental studies found the effect reliable at a “small-but-approaching-medium” size (0.22–0.26) across contexts — more attractive avatars produced more confident behavior, athletic avatars increased exercise, heroic avatars increased prosocial behavior offline [2]. Newer work complicates the embodiment requirement: participants formed attachment to avatars even without strong immersion, suggesting the effect is not purely a VR/embodiment phenomenon [2]. Practical read for Math Challenge: a cosmetic that visually signals “careful, mastery-oriented mathematician” (a scholar’s cap, an explorer’s badge) is not just a reward — it may nudge behavior toward the trait it depicts, arguing for cosmetic categories tied to real learning behaviors (persistence, accuracy, curiosity) rather than arbitrary skins.

3. Safe multilingual name generation and the Scunthorpe problem

The “Scunthorpe problem” is named for a 1996 incident where AOL’s filter blocked the English town of Scunthorpe because its name contains an offensive substring; the same failure mode has hit Craig Cockburn’s surname, “shitake mushrooms,” Google SafeSearch results for the town, and — critically for a five-language product — German compound words that legitimately form an offensive substring at a linking letter, and Chinese place names flagged by character homonymy [1]. The mechanism is structural: naive substring filters “lack contextual understanding” [10], and the failure gets worse the more languages a single filter covers, because a clean word in one language can be an unsafe substring in another — China’s own moderation ecosystem shows the endpoint, with no single authoritative banned-word list even within one language, so every platform runs its own edge-case list indefinitely [10]. The production-proven fix is not a stricter algorithm — it is a maintained, logged whitelist of confirmed-safe strings a naive filter would reject, built from real rejection logs over time [1][10]. The direct implication for word-list construction: EN/ES/FR/PT/DE lists must be authored per language by a fluent reviewer, not machine-translated, because translation is exactly the operation that turns a safe word into an unsafe substring elsewhere (mirrors D-005’s finding that math vocabulary itself cannot be translated word-for-word across these five languages [19]).

4. Moderation of user-generated content — should children get free text at all

Roblox is the highest-investment case study available: mandatory age verification since January 2026, chat restricted to age-banded groups, still “1,600+ moderators” plus AI-driven filtering, and a 2020 investigation described takedown efforts as “whack-a-mole” with sexual content reappearing; a 2024 report tied real predation incidents to “insufficient moderation,” and at least six arrests followed for child exploitation via the platform since January 2025 [13]. This is not a company that under-invested — it is evidence that free text at scale, from children, cannot be fully moderated even with a nine-figure trust-and-safety budget. The honest conclusion for a product at Math Challenge’s scale: no free-text surface for children at all in v1. Alias selection, avatar assembly, and any social interaction should be closed-vocabulary (tap a preset phrase or emoji), never a text box — stricter than most competitors, but the only posture that does not inherit Roblox’s moderation debt.

5. Cosmetic progression vs. gambling: the regulatory line

Belgium’s Gaming Commission ruled in 2018 that loot boxes in FIFA 18, Overwatch, and CS:GO were games of chance under gambling law regardless of cosmetic-only contents — no way to buy the specific item directly makes it chance-based — and the justice minister framed the concern explicitly around children [4]. The Netherlands’ gaming authority reached a similar conclusion the same year for several (not all) titles studied, citing addiction-fostering design even below the legal enforcement threshold [4]. The UK took the narrower position that cosmetic-only, non-tradeable, non-cashable items sit outside licensable gambling, but its own 2022 DCMS review still found elevated “gambling, mental health, financial and problem gaming-related harms” among loot-box players and pushed self-regulation over a law change [4]. Read together, the regulatory center of gravity is: randomization plus payment is the trigger, not the cosmetic/gameplay distinction — a deterministic purchase (buy this specific hat) is safe everywhere these rulings apply; a paid randomized box is contested even when purely cosmetic. For a children’s product this argues for going further than the strictest current ruling: no randomized rewards of any kind, paid or free, since the mechanism regulators worry about (variable-ratio reinforcement) needs no money to work on a child.

6. Companion/pet care loops: retention power and the guilt mechanic

Tamagotchi (1996) is the reference case: three meters (hunger, happiness, training) that decay without attention, with genuine death as the failure state [5]. It produced extraordinary engagement — 40 million units in two years — precisely because neglect had a real, upsetting consequence: children brought devices to school to prevent death mid-lesson, schools banned them for disruption, and press at the time reported genuine grief, including mock funerals for “dead” pets [5]. The retention mechanism and the dark pattern are the same mechanism — the device does not work as a companion without a threat, and the threat produced both the engagement and the backlash. A companion feature can borrow the care loop (feed, dress, interact) without the loss loop: growth stages and cosmetic unlocks tied to the child’s own math practice, no decay state, no death, no notification framed around the companion’s distress — keeping the affection-building mechanic while discarding the guilt mechanic that made the original both beloved and controversial.

7. Pedagogical agent research — why Larry’s format matters, not just his existence

Mayer’s multimedia-learning research on pedagogical agents finds real but modest support: agents help mainly as a “presenter of social cues,” and a recent meta-analysis found only “negligible improvement” from the agent’s mere presence — the gains come from what the agent does (highlighting, personalizing, human-like voice), not from having a character at all [9]. Tutor-role agents show no clear advantage over agent-free lessons; co- learner-framed agents raise self-efficacy more reliably; static vs. animated representation remains an open, contradictory question [9]. This converges with Sesame Street’s own research history: ETS’s 1970-71 studies found heavier viewers learned more regardless of disadvantage, and a 2019 economic analysis (Kearney & Levine) called it “perhaps the biggest, yet least costly, early childhood intervention” [6]. Underneath that is parasocial-bond research: Calvert’s team found 21-month-olds learned a sequencing task better from a familiar character (Elmo) than an unfamiliar one, and that giving children a toy of the unfamiliar character first — so it became familiar — closed the gap; personalized characters deepened engagement further, since “perceived similarities increase children’s interest and investment” [7][8]. This favors Larry as already decided in D-004: he is not a new character but Ignia’s existing orange rhino, so the familiarity Calvert’s research says drives learning is already banked, not something to build from zero [18][19]. The corollary: Larry’s canon rules (never mocks the child, “¡Ya vas!” only on task acceptance) matter more than his visual design — the payoff is in the social-cue behavior, not the character model [9][18].

8. Visualizing progress for children — maps, trees, and the goal-gradient effect

The goal-gradient effect — effort accelerates as perceived distance to a goal shrinks — is well established in the behavioral-economics/marketing literature (Kivetz, Urminsky & Zheng’s “resurrection” of Hull’s original hypothesis, in loyalty-program contexts) [14]. Gamified-learning reviews note visual learners specifically “benefit from… progress bars, game maps, and colorful visuals” for motivation, while cautioning that points/badges/ leaderboards used in isolation are not reliably effective — they work when tied to real competence signals, not as decoration [15]. This matches the project’s own PRIMARY-band research: a “wardrobe effect” study found children build several avatars but converge on using one, meaning the process of customizing carries the motivational value even when the final artifact is narrow [16]; KINDER-band research found young children “expect feedback on every single action” and respond to consistent mascot animation in a way adults explicitly do not [17]. Duolingo’s own history is a cautionary data point on which progress metaphor to pick: it replaced its tree-shaped skill map with a linear path in August 2022, drawing visible, sustained user backlash [12] — evidence the tree/map metaphor itself carried motivational value a straight line did not replace.

Design implications

  1. Alias generator is per-language authored, not translated. Build five independent word lists (EN/ES/FR/PT/DE) with a fluent reviewer per language, combining a category word + a trait word + a randomized (not sequential) two-digit number. Never derive one language’s list by translating another’s — that is the exact operation that creates Scunthorpe-style failures [1][10].
  2. Alias selection is tap-only for ages 3-11 (KINDER + PRIMARY). The child picks from 3-5 generated options; there is no text field to type into, which removes the injection surface for the age bands least able to self-moderate. TEEN (12-17) gets a limited reroll but still never types.
  3. Validate the combined rendered string, not each word alone, and log every rejection/regeneration (words, language, locale, reason). Two clean words can combine into an unsafe one across a language boundary; a logged whitelist of confirmed-safe strings, built from those logs, is the actual production fix for the Scunthorpe problem, not a stricter filter [1]. Rate-limit regeneration (e.g., a small number of rerolls/day) so a child cannot brute-force a combination past the filter.
  4. No free-text field anywhere a child under 13 can reach — no bio, no chat, no custom label. Any child-to-child interaction (reactions, congratulations) is a closed set of preset phrases/emoji, per the honest read of Roblox’s moderation record even at scale [13].
  5. Avatar = pre-made parts only, Mii-style. No photo upload, no camera access, ever. Parts are organized in categories (hair, expression, accessory) unlocked as cosmetics, not typed or drawn.
  6. Cosmetics unlock on a deterministic, published table tied to mastery milestones, streaks, or league placement — e.g., finishing a topic unit unlocks a specific named item. No box, chest, pack, or “mystery” reward of any kind, free or paid — stricter than the UK’s own cosmetic carve-out and clear of the Belgium/Netherlands line entirely [4]. If monetization ever touches cosmetics, it is direct purchase of a named, previewed item only, never a randomized purchase.
  7. A companion/pet feature (if built) has no decay, hunger, or death state, and sends no notification framed around the companion’s sadness or neglect — keep the affection-building loop, discard Tamagotchi’s guilt loop [5]. Likewise, streak loss never erases or visually regresses the progress map — a missed day should not retroactively undo earned progress, avoiding a second guilt mechanic stacked on the first.
  8. Progress visualization is age-banded, not one skin for all ages: KINDER — a physical journey path with the mascot walking forward, no numbers; PRIMARY — a named-topic skill tree/mastery map (not a straight line — Duolingo’s own 2022 switch away from a tree drew visible backlash [12][15]); TEEN — a stats dashboard with an opt-in league; ADULT — plain numeric mastery metrics, gamified skin optional and off by default (per mc-23’s adult/expert findings).
  9. Larry’s presence is age-banded, per mc-37’s existing prompt architecture: KINDER — animated, voice-first, reacts to every attempt; PRIMARY — same character, explains the misconception per the hard rules already set in D-004; TEEN — same rhino, tone shifts toward “patient professor,” fewer animations, opt-out available; ADULT — Larry available on request but backgrounded behind denser feedback UI [9][18].
  10. Larry never comments on the child’s alias or avatar choice. His canon voice stays on math, never on appearance or identity — extends D-004’s “never shames a child” rule to a place shame could otherwise creep in (a bot “complimenting” a name implicitly also can judge one) [18][19].
  11. Cosmetic categories signal learning traits the Proteus-effect research says can influence behavior (persistence, curiosity, care) — e.g., an “explorer” badge line for trying hard problems — rather than arbitrary decoration, since the avatar has a small but real behavioral echo on the wearer [2][3].
  12. Alias/avatar identity is never derivable from the parent account, email, or child’s real name in the UI — a numeric suffix should be randomized, not sequential, else “Bunny07” implies signup order.

Open questions for the project owner

  1. Should the closed-vocabulary reaction set (emoji/preset phrases) exist at all in v1, or is “no child-to-child interaction of any kind” the safer default until there is a moderation budget to support even a closed set?
  2. Is a Tamagotchi-style companion in scope for this product at all, or does the guilt-loop risk (§6) argue for skipping the feature entirely rather than trying to build a “safe” version of it?
  3. For the PRIMARY-band skill tree: should topic order be strictly linear (matches the current curriculum sequencing) or allow branching exploration (matches the “wardrobe effect” finding that children value choice even when they converge on one path)?
  4. Should cosmetic unlock tables be visible to the child in advance (full transparency: “finish fractions to unlock this hat”) or revealed only on unlock (surprise, but closer in spirit to the mechanic regulators scrutinize even without randomness)?
  5. Do TEEN-band aliases get one extra degree of freedom (e.g., a self-chosen number, still validated) versus PRIMARY/KINDER’s fully generated set, or is the same tap-only mechanism kept through age 17?
  6. Should the alias word lists be built in-house per language, or is there budget/appetite to license an existing maintained multilingual profanity/safe-word dataset instead of authoring five lists from scratch?

Sources

  1. Wikipedia — Scunthorpe problem
  2. Wikipedia — Proteus effect
  3. Yee, N. & Bailenson, J. (2007). "The Proteus Effect: The Effect of Transformed Self-Representation on Behavior." Human Communication Research, 33(3). Original study underlying [2]; not re-fetched live
  4. Wikipedia — Loot box (regulatory history: Belgium, Netherlands, UK DCMS)
  5. Wikipedia — Tamagotchi
  6. Wikipedia — Sesame Street research
  7. Wikipedia — Parasocial interaction
  8. Lauricella, A., Gola, A., & Calvert, S. (2011). "Toddlers' learning from socially meaningful video characters." Media Psychology, 14(2). Primary study referenced in [7]; not re-fetched live
  9. Wikipedia — Pedagogical agent
  10. Wikipedia — Profanity filter
  11. Wikipedia — Mii
  12. Wikipedia — Duolingo
  13. Wikipedia — Roblox (safety/moderation record)
  14. Kivetz, R., Urminsky, O., & Zheng, Y. (2006). "The Goal-Gradient Hypothesis Resurrected: Purchase Acceleration, Illusionary Goal Progress, and Customer Retention." Journal of Marketing Research, 43(1). Established literature; not re-fetched live
  15. Wikipedia — Gamification of learning
  16. Internal — math-challenge/docs/research/2026-07-31-mc-21-ui-ages-7-11-primary.md §3 (avatar "wardrobe effect" study, self-determination theory framing)
  17. Internal — math-challenge/docs/research/2026-07-31-mc-20-ui-ages-3-6-kinder.md §5, §7 (mascot/animation preference, feedback-on-every-action finding)
  18. Internal — math-challenge/docs/research/2026-07-31-mc-37-larry-profe-port.md (Larry's hard rules, prompt architecture, age-band tone)
  19. Internal — math-challenge/docs/decisions.md D-003 (generated aliases), D-004 (Larry), D-005 (five languages)

Open questions this document leaves for the owner

These are unanswered on purpose. They are listed, not resolved — turning them into a FAQ would mean inventing answers the document does not contain.

One of 51 research documents, 168,346 words in total, counted at build time from the files themselves. Read this document in the repository