Kidoio Methodology

Behavioural Measurement for Children's Digital Content

Studio age ratings describe what children's content forbids. They do not describe what it teaches, what behaviour it models, what it rewards, or which developmental band the content is genuinely appropriate for. Kidoio is a measurement layer that produces, for any video intended for child audiences, a structured behavioural profile and a recommended developmental age band — anchored only on frameworks with strong empirical support, with contested frameworks explicitly excluded from the scoring engine.

1. Purpose and scope

1.1 What Kidoio measures

Kidoio analyses children's digital content — initially video, with games and interactive media to follow — and produces:

• A structured behavioural profile across nine observable dimensions: empathy and emotion-naming, aggression, cooperation versus dominance, conflict-resolution style, reward framing, cognitive engagement, self-regulation modelling, adult/parenting voice modelled, and body & autonomy framing.
• A content-maturity layer tracking adult themes that are not behavioural patterns: real (live-action) violence, on-screen death, drug use, alcohol and tobacco, sexual content, nudity, profanity, graphic imagery, disturbing themes, crime glamorisation, political and religious content.
• A recommended developmental age band — one of seven evidence-aligned bands (infancy, toddlerhood, preschool, early school, late childhood, early adolescence, late adolescence) — derived from the behavioural profile, the maturity layer, and the cognitive complexity of the content.
• Per-finding on-screen evidence: every flag links to a timestamp in the source content.

1.2 What Kidoio does not measure

Kidoio is a measurement of content, not children. It does not:

• Predict, infer, or claim individual neurological, behavioural, or developmental outcomes for viewers.
• Diagnose, replace, or supplement clinical assessment.
• Recommend or prescribe content. The recommended age band is descriptive, not prescriptive.
• Produce per-child profiles, biometric data, or individual-level behavioural prediction.
• Make claims about the long-term effects of any specific viewing on any specific child.

These boundaries are not stylistic; they are load-bearing. Crossing them would put Kidoio in regulatory and methodological territory it is not built for.

1.3 Audience

Three audiences are served simultaneously: platforms (streaming services, content libraries, kids' EdTech) for pre-release content audit and compliance documentation; regulators and policy bodies as measurement infrastructure for emerging child-safety obligations under DSA Article 28, the UK Online Safety Act, COPPA reauthorisation, and child provisions of the EU AI Act; and researchers as structured stimulus metadata for downstream developmental, neuroscientific, and educational research.

2. The measurement gap

Existing classification systems for children's content fall into three categories. Each is necessary; none is sufficient.

Studio and statutory age ratings (PEGI, ESRB, BBFC, MPAA, Common Sense Media age levels) describe forbidden content with reasonable accuracy: graphic violence, sexual content, profanity. They do not describe what behaviour the content models, what reward structure it uses, or how it paces cognitive load. A 7+ animated series can pass its age rating while explicitly rewarding physical aggression with status elevation in every conflict. A 3+ preschool show can pass its age rating while consisting of rapid-cut sequences with near-zero narrative dependency. Both situations occur in widely-distributed contemporary children's content.

Editorial reviews (Common Sense Media, IMDb Parents Guide, parent-led blogs) provide context that ratings do not. They are valuable, slow, expensive per title, and depend on reviewer judgement that varies between titles and across time. They do not scale to a streaming catalog of tens of thousands of titles, and they cannot document compliance with regulation that requires platform-level content stewardship.

Trust & Safety tooling (Bark, Qustodio, Net Nanny, Kidas, SuperAwesome) addresses the runtime side: filtering, monitoring child usage, blocking inappropriate apps. These are valuable. They do not analyse what content is teaching — they treat content as opaque.

The missing layer is structural measurement of content itself: what does this video model? what does it reward? how is it paced? for what developmental band is its profile appropriate? Kidoio is that layer.

3. Taxonomy v1

3.1 Behavioural dimensions

Each dimension is described below alongside its Tier 1 anchor — the framework that load-bears the corresponding scoring rule.

3.1.1 Empathy and emotion-naming

Observable moments where characters acknowledge another's emotional state, express concern about harm, apologise, or engage in emotional repair. Particular attention to whether emotions are named (vocabulary) and whether mixed or self-conscious emotions are handled appropriately for the content's apparent target band.

Tier 1 anchor: RULER (Brackett, Yale Center for Emotional Intelligence). RULER has cluster RCT evidence in 62 schools / 3,824 students [Hagelskamp et al., 2013] and is listed as evidence-based by CASEL. Gottman emotion coaching reinforces the construct.

Orienting frame in copy only: Selman's perspective-taking stages.

3.1.2 Aggression

Observable instances of physical force, verbal threats, intimidation, and destructive action. Cartoon-stylised aggression and live-action aggression are distinguished, because the latter belongs in the content-maturity layer. Within the behavioural layer, aggression is evaluated on three sub-questions: how much is present, how the narrative reacts to it (rewarded, neutral, punished), and what alternative responses are modelled.

Tier 1 anchors: Bandura's social learning theory (one of the most reliable findings in developmental psychology); Tremblay's developmental aggression curve, which calibrates expectations — physical aggression peaks at 24-36 months and declines, against common assumption.

3.1.3 Cooperation versus dominance

For each conflict, the resolution mode is classified — collaboration, hierarchical authority, domination, or unresolved. The aggregate distribution characterises whether the content is cooperation-dominant, balanced, or dominance-dominant.

Tier 1 anchors: Bandura modelling reinforced by Gottman's longitudinal work on conflict patterns and child outcomes.

Orienting frame in copy only: Kohlberg's moral reasoning stages.

3.1.4 Conflict resolution style

The methods by which conflicts conclude: dialogue, hierarchical authority, physical force, avoidance, or no resolution. Particular attention to the presence of repair attempts after rupture — whether characters acknowledge harm and reconnect.

Tier 1 anchors: Gottman emotion coaching and the related "four horsemen" framework (criticism, contempt, defensiveness, stonewalling); Greene's Collaborative & Proactive Solutions [multiple RCTs reducing externalising behaviour, ~50% of treated youth diagnosis-free post-treatment] for the framing of behaviour as lagging skill rather than moral failure.

3.1.5 Reward framing

What behaviour does the narrative reward? Praise, success, status elevation, material gain, social approval — and whether these follow prosocial behaviour, neutral behaviour, or aggressive/dominance-based behaviour. Particular attention to whether reward is for process (effort, strategy, persistence) or for outcome (winning, talent, innate ability).

Tier 1 anchor: Bandura modelling — reinforcement schedules visible to observers shape acquisition.

Drop: Dweck's "growth mindset" intervention construct. Macnamara & Burgoyne (2023) meta-analysis of 63 studies / ~98,000 participants reported overall effect d = 0.05 (n.s. after publication-bias correction); the highest-quality subset returned d = 0.02. The behaviour of praising process is preserved in copy and feature design (it does no harm and aligns with self-determination theory), but mindset science is not cited as a scoring anchor.

3.1.6 Cognitive engagement

The structural characteristics of pacing and narrative dependency. Four sub-signals are evaluated: stimulus density (rate of audiovisual change), repetition, narrative dependency (whether later scenes require understanding of earlier ones), and cognitive demand (markers of cause-effect, intent tracking, problem-solving). A specific structural pattern — high stimulus density combined with low narrative dependency — is labelled as a brain-rot pattern and treated as distinct from intentional simplicity in early-learning content.

Tier 1 anchors: Vygotsky's zone of proximal development; AAP (2016, 2023) and WHO (2019) screen-time and digital-media guidance for the precautionary thresholds.

Orienting frames in copy only: Piaget's progression direction (literal stage ages explicitly not used as gating logic, because modern developmental psychology has revised those ages downward and made them more domain-specific). Greenspan's functional emotional milestones as reference frame.

3.1.7 Self-regulation modelling

Whether characters demonstrate strategies for emotion regulation — distraction, reframing, requesting support, naming the feeling — or whether they rely on pure willpower or external restraint.

Tier 1 anchor: Gottman emotion coaching's documented effects on regulatory outcomes.

Drop: Mischel's marshmallow-test predictive narrative. Watts, Duncan & Quan (2018) replication with larger and more representative N showed the predictive correlation reduced to roughly half the original size, and reduced by two thirds when controlling for SES, family background, and early cognitive ability. The strategies (distraction, reframing) remain valid as feature patterns; the predictive narrative around delay-of-gratification as a marker of future success is excluded from scoring.

3.1.8 Adult-voice / parenting style modelled

Where adult characters are present, what parenting voice is modelled: authoritative (warm + structured + reasoning), authoritarian (controlling, low warmth), permissive (warm, low structure), uninvolved, or shaming. The dimension is descriptive of the adult voice the content offers as an ambient model.

Tier 1 anchor: Baumrind's parenting-styles typology. Authoritative parenting is consistently linked to better child outcomes across multiple cultures and decades.

3.1.9 Body and autonomy framing

For content that addresses bodies, growth, sexuality, or physical autonomy: whether the language is anatomically correct or euphemistic; whether bodily autonomy is represented; whether forced sharing or forced apology is modelled as moral teaching; whether consent is depicted appropriately for the content's apparent band.

Tier 1 anchor: AAP guidance and Planned Parenthood clinical recommendations for age-appropriate body language and consent education.

3.2 Content-maturity layer

A separate layer tracks the kind of content present — distinct from how content models behaviour. Maturity themes: real (live-action) violence, on-screen death, drug use, alcohol and tobacco, sexual content, nudity, profanity, graphic imagery, disturbing themes (torture, abuse, kidnapping), crime glamorisation, political content, religious content. Each instance is timestamped, intensity-graded, and tagged with the age bands for which it is inappropriate.

This layer does not attempt to model what such content "does" to viewers. It documents what is present.

3.3 Recommended age band

The recommended developmental age band is one of seven:

Erikson's framework provides the labels; scoring does not depend on Erikson's predictive claims. Bands are also rendered as a single-number equivalent (9+, 12+, 16+) for compatibility with legacy rating systems.

4. Evidence-quality tiering

This section is the crux of the methodology and the principal artefact for advisory review.

4.1 The three-tier evidence framework

Frameworks are tiered by the quality and quantity of empirical evidence supporting their constructs and predictions, not by popularity, theoretical elegance, or familiarity to general audiences. This distinction is load-bearing: many widely-cited frameworks in popular parenting culture (growth mindset, popular EQ, marshmallow predictive) have not survived rigorous replication, and Kidoio does not depend on them.

Tier 1 — Strong evidence

Frameworks with multiple RCTs, large meta-analyses, or robust cross-cultural replication. Used as load-bearing anchors in the scoring engine.

Tier 2 — Moderate evidence

Useful theoretical frames with mixed empirical record. Used as orienting frames in copy and parent-facing language; not used as scoring gates.

Tier 3 — Contested or weak evidence

Frameworks whose strongest empirical claims have not survived rigorous replication, or whose constructs have known measurement problems. Not used as scoring anchors. Where useful behavioural patterns from these frameworks survive, the patterns are preserved without the framework branding.

4.2 Why this matters

A measurement methodology cited by regulators, platforms, and researchers must be defensible against reviewers who know the evidence-quality literature. Anchoring scoring on Tier 3 frameworks would invite avoidable methodological challenge. Anchoring on Tier 1, citing Tier 2 only as orienting language, and explicitly excluding Tier 3 from the scoring engine produces a methodology that survives scrutiny.

This tiering is not static. As the evidence base evolves — new RCTs, new failed replications — frameworks will move between tiers, and the methodology version will be incremented accordingly.

5. The scoring engine

5.1 Architecture

Content analysis flows in three stages:

1. Extraction. A multimodal foundation model (currently Gemini 1.5/2.5) processes the source content end-to-end and produces a strictly-structured JSON payload enumerating observable instances within each behavioural and maturity dimension. Each instance includes a timestamp, a description, and metadata fields appropriate to the dimension. The model is constrained to enumerate, not to judge: prompts explicitly forbid evaluative or diagnostic language.
2. Scoring. A deterministic rule engine maps the extraction payload into categorical scores per dimension and an overall recommended age band. Rules are documented, versioned, and traceable: every score derives from a stated rule applied to a stated count or pattern. Scores do not "emerge from a model"; they are computed.
3. Explanation. A plain-language summary is generated, citing the specific instances that drove each score. The explanation may be model-generated or template-rendered; in both cases it is constrained to refer only to instances present in the extraction payload.

This separation between extraction and scoring is load-bearing. The extraction model can be replaced as the foundation-model state of the art evolves; the scoring rules remain stable and continue to produce comparable scores across models.

5.2 Auditability

Every score in a Kidoio profile is traceable to: the taxonomy version under which the dimension was defined; the prompt version used to extract observations from source content; the scoring-rules version applied to those observations; the specific observations (with timestamps) that drove the score; the Tier 1 framework anchoring the rule.

A reviewer who disagrees with a score can examine the underlying rule and the underlying observation, and adjudicate at either level. There is no "model said so" component in the scoring layer.

5.3 Determinism

The same extraction payload, scored under the same rule version, will produce the same scores. The extraction step (depending on a stochastic foundation model) is not fully deterministic, but its output schema is strict and its output is fully inspectable; reproducibility is empirical (run-twice agreement) rather than guaranteed by construction. This is acknowledged limitation and is a focus of ongoing measurement (see §7).

6. Recommended age bands

6.1 Derivation

The recommended age band is the lowest band for which the content's joint behavioural and maturity profile is appropriate, where "appropriate" is determined by the band-specific guidance synthesised from Tier 1 anchors and Tier 2 orienting frames as applied to each band.

The derivation is not a simple lookup against the maturity layer. A piece of content with no mature themes can still be recommended for an older band if its cognitive complexity, conflict patterns, or moral framing is mismatched with younger bands. Conversely, content with low cognitive complexity and no problematic behavioural patterns can be recommended for a younger band even if some mature elements (e.g. brief parental conflict shown with repair) are present, because the framing is appropriate for that band.

6.2 Band-by-band synthesis (summary)

A full per-band reference is published separately. In summary:

• Infancy (0–18 mo). AAP / WHO: no solo screen content. Kidoio rates any content as "not recommended for solo viewing in this band" by default.
• Toddlerhood (18 mo – 3 yr). Soft cap on session length. No violent or scary media, including stylised. No sexual content. No public shaming as moral teaching. Tremblay's curve: physical-aggression coaching framed as developmental, not moral failure.
• Preschool (3–6 yr). Strict media curation. No chat or social features. Bandura: what characters model is what children acquire. Gottman: adult-voice modelling matters most here.
• Early school (6–9 yr). Concrete-operational thinking dominant. Effort-based feedback. No public profiles or algorithmic recommendations. Closed-circle messaging only.
• Late childhood (9–12 yr). Relational-aggression peak. Coping-skills framing. Default-off algorithmic feeds. No public counts. Strong report flow.
• Early adolescence (12–15 yr). Identity formation; high peer salience; immature impulse control. Hard guardrails on pornographic content (Owens et al. 2012 plus subsequent reviews), self-harm content, extremism rabbit holes, gambling-style mechanics.
• Late adolescence (15–18 yr). Adult-style consent UX. Friction on late-night use. Algorithm transparency. Direct paths to local crisis, DV, sexual-assault resources where any clinical surface is present.

6.3 Output presentation

The recommended band is presented as both the band label ("late childhood 9–12 yr") and the single-number equivalent (9+) for legacy compatibility. Where a comparison rating is provided by the user, a comparison line is rendered: "Studio rated 7+; Kidoio recommends 9-12 yr (9+)." No comparison is rendered when no studio rating is provided; the recommendation stands on its own.

7. Limitations

7.1 Empirical limitations

• Reproducibility of extraction. The extraction model is stochastic. Two runs on the same content can produce different observation counts, different timestamps within tolerance, and occasionally different categorical labels for borderline instances. We measure inter-run agreement and report it; we do not claim deterministic extraction.
• Validation gap. The scoring rules are theoretically anchored. Whether scores produced by Kidoio correlate with independent expert ratings of the same content is the subject of ongoing validation work and is not yet demonstrated. A small validation study (40-60 episodes, 5-10 expert reviewers, inter-rater agreement against Kidoio scores) is planned.
• Cultural variation. Most of the Tier 1 evidence is dominated by Western, English-language samples. Frameworks vary in how well they generalise. The methodology will be revised as non-Western evidence accumulates.

7.2 Methodological limitations

• Contested boundaries. Some flag categories sit at the edge of Tier 1 / Tier 2. Reasonable methodologists would draw the line in slightly different places.
• Foundation-model dependence. Extraction depends on a multimodal foundation model. Performance will track that frontier; large changes in foundation-model behaviour can affect extraction quality. The scoring layer is robust to this only because it is rule-based; the upstream remains a dependency.

7.3 What we explicitly do not claim

• That a Kidoio recommendation for a given band predicts any individual child's developmental outcome from viewing the content.
• That higher Kidoio scores on a dimension cause better child outcomes.
• That the methodology is a substitute for clinical assessment, parental judgement, or editorial review.
• That Kidoio scores are calibrated against any specific regulatory threshold (although they are designed to be useful inputs to regulatory analysis).

8. Versioning & governance

8.1 Version pins

Three version numbers travel with every Kidoio profile:

• Taxonomy version (current: v1) — the set of dimensions measured.
• Prompt version (current: v0.5.0) — the exact text of the extraction prompts.
• Scoring rules version (current: v3) — the deterministic rules applied to extraction output.

A profile produced under a given version pin remains valid against that version regardless of subsequent changes. Old profiles are not silently re-scored. Where a buyer requires up-to-date scores, content is re-analysed under the current versions and the audit trail records both runs.

8.2 Governance of changes

Changes to taxonomy, prompts, or scoring rules go through internal review against the Tier 1 / Tier 2 / Tier 3 evidence framework; external advisory review; publication of the change with rationale, version increment, and effective date; and audit-trail update. Backward-incompatible changes that would alter scores under earlier versions require a major taxonomy version bump.

9. Acknowledgements

The Tier 1 / Tier 2 / Tier 3 evidence framework was developed for this methodology drawing on a structured reference compiled by the founder, with synthesis of the contemporary developmental-psychology literature. Errors of synthesis are the founder's; the underlying scholarship is the cited authors'.

This methodology is currently in advisory-review draft. Named advisors will be added in a subsequent revision following formal advisory engagement. Researchers interested in advisory engagement are invited to contact hello@kidoio.com.

10. References