# Credibility Map — Research Files Audit

**Date**: 2026-04-11
**Method**: 1 verification attempt per claim via WebSearch/WebFetch. Budget ~8 min/claim. Breadth over depth.
**Scope**: 5 files, ~62 distinct empirical claims checked.
**Auditor**: Research integrity pass following retraction of unsourced material in `competitor-analysis.md`.

---

## Summary

| Status | Count | % |
|---|---|---|
| VERIFIED (URL cited or direct match) | 28 | 45% |
| PARTIAL (close but not exact, or partially attributed) | 13 | 21% |
| UNVERIFIED (no source found in 1-2 searches) | 14 | 23% |
| RETRACT (contradicted, fabricated, or misattributed) | 7 | 11% |

**Top 3 load-bearing risks** (if wrong, thesis breaks):

1. **"1,500 micro-skills" — VERIFIED**, but the only public source is a Medium post by Leta Capital (Novakid's investor) and a TEFL Academy / Teast.co reproduction. No primary source from Novakid itself, no academic validation, no independent audit. The number is real but the citation chain is thin. Thesis stands but disclose source quality.

2. **"FTC 2026 Click-to-Cancel Rule — legally mandatory" (dataviz-patterns.md line 86) — RETRACT**. The rule was VACATED by the U.S. Court of Appeals for the Eighth Circuit on July 8, 2025, days before its compliance deadline. The FTC restarted rulemaking via ANPRM on March 11, 2026. As of April 2026 there is no enforceable click-to-cancel rule. This is a factually wrong claim presented as legal fact.

3. **"40-30-20-10 Space Rule (Linear)" (dataviz-patterns.md line 44) — RETRACT**. Linear's actual blog "How we redesigned the Linear UI (part II)" does NOT contain any 40-30-20-10 percentage allocation or any space rule with these proportions. This appears to be a fabrication from Improvado's marketing dashboard blog that is then attributed to Linear. The principle itself may be reasonable, but the attribution is false.

**Top 3 safe wins** (strong verification, can cite confidently):

1. **Novakid official metrics (940K parents, 24M+ lessons, 1 in 11 teachers, $41.4M funding, Inc 5000 #105 in 2023)** — all verifiable on novakidschool.com/about-us, Crunchbase, and multiple secondary sources. Cite directly.
2. **CampusESP +8.3% retention (2024 study, 12,374 students at 9 universities)** — verified verbatim on campusesp.com/blog. Citable as primary research.
3. **Chegg $7.5M FTC settlement (2025)** and **ABCmouse $10M FTC settlement (2020)** — both directly verified on FTC.gov with case files. Citable as legal facts.

---

## File 1 — parent-psychology.md

### Claims extracted

| # | Claim (verbatim or paraphrased) | Status | Source / Note | Recommended action |
|---|---|---|---|---|
| 1 | "EdTech retention: 4-27%" | PARTIAL | Already disclosed in trustpilot-systematic.md as "PARTIALLY VERIFIED": 4% from Alchemer 2021, 27% upper bound has no primary source | KEEP+DISCLOSE: "as low as 4% annual (Alchemer 2021); industry estimates up to 27% in premium segments" |
| 2 | "17-20x lower than top subscription categories" | UNVERIFIED | No source found for this specific multiplier in 1 search | DROP or replace with "significantly lower than entertainment/news subscription categories" |
| 3 | "Most need 6-15 months to break even, average lifespan ~4 months" | VERIFIED | loyalty.cx/edtech-retention-problem/ — exact match | KEEP+CITE loyalty.cx |
| 4 | "Cancellation drivers: low usage (37%), perceived value (35%), cost (35%)" | UNVERIFIED | Specific %s not found in loyalty.cx or RevenueCat in 1 search; the file lists "RevenueCat" as source but no direct URL | KEEP+CITE explicitly as RevenueCat State of Subscriptions report (find URL) OR drop the percentages and keep the categories |
| 5 | "Parents feel guiltier about unused learning subscriptions than any other type" | UNVERIFIED | Search for this specific framing returned only Lingokids' 2025 "Parent Guilt on Trial" campaign and screen-time guilt research (74% feel guilt about screen time). No source for "guiltier than any other type" | RETRACT or rephrase as "parents express guilt about unused subscriptions (Lingokids 2025 campaign frames this as a known phenomenon)" |
| 6 | "CEFR framework gives structure: 'follow a path instead of guessing'" | UNVERIFIED | Quote not attributed; could be a paraphrase from a parent or research. Generic enough to be defensible | KEEP as paraphrase, drop quote marks |
| 7 | "26 concrete milestones covering vocab/grammar/communication for ages 1-6" | RETRACT (misattributed) | A real peer-reviewed instrument (Visser-Bochane et al. 2019, ScienceDirect) identifies 26 milestones for Dutch children ages 1-6 — but it's a developmental language-acquisition study, NOT part of CEFR. CEFR is A1-C2 for adults/teens and has no "26 milestones" structure. The file frames this as part of the CEFR framework, which is factually wrong. | RETRACT the CEFR framing. Either cite the actual study (Visser-Bochane 2019) as separate developmental research, or drop the claim |
| 8 | "85% of parents give low marks to cross-app progress tracking" | UNVERIFIED | No source found in 1 search. This statistic does not appear in any indexed material | RETRACT or find source. If kept, must cite specific study |
| 9 | "80% of parents report educational anxiety" | PARTIAL | Already disclosed in trustpilot-systematic.md as China-specific (Family Education Blue Book 2024). The Frontiers Wu et al. 2022 paper is real but studies 259 parents and does not contain 80% figure | KEEP+DISCLOSE: "80% of Chinese parents report educational anxiety (Family Education Blue Book, 2024)". Do NOT use globally |
| 10 | "Loyalty.cx: EdTech retention problem" (source) | VERIFIED | URL exists, content matches | KEEP |
| 11 | "Frontiers in Psychology: parental educational anxiety" (source) | PARTIAL | Wu et al. 2022 paper exists at the cited DOI but does not contain the headline 80% claim | KEEP citation but qualify what it actually says |
| 12 | "EdSource: parents in the dark about language progress" (source) | UNVERIFIED | Listed as a source without URL or specific claim. Cannot verify what was used from it | Find specific URL or drop |
| 13 | "Baymard Institute: SaaS subscription UX" (source) | UNVERIFIED | Generic citation, no specific finding referenced | Find specific Baymard study or drop |
| 14 | "NNGroup: dashboard usability" (source) | UNVERIFIED | Generic citation. NNGroup has many dashboard articles. Which one? | Find specific NNGroup article URL |
| 15 | "RevenueCat: subscription churn reasons" (source) | UNVERIFIED | Generic citation. RevenueCat publishes annual State of Subscriptions, but specific report not linked | Cite specific RevenueCat report URL |

**Overall verdict on this file**: MIXED. The narrative arc is plausible and the framework (notification cadence, info hierarchy, anxiety reduction) is defensible as design opinion. But the empirical scaffolding is weak: the 6 "Sources" section at the bottom lists organization names without URLs or specific findings, multiple unattributed statistics float in the body, and the "26 milestones / CEFR" claim is factually wrong. The file reads as if every claim has research backing when in practice ~40% are unsourced or misattributed.

**Top fix needed**: Fix the "26 milestones in CEFR framework" misattribution (it confuses Visser-Bochane developmental research with CEFR), and either add real URLs to the Sources section or rewrite as design heuristics with the few sources that are real.

---

## File 2 — dataviz-patterns.md

### Claims extracted

| # | Claim (verbatim or paraphrased) | Status | Source / Note | Recommended action |
|---|---|---|---|---|
| 1 | "Giorgia Lupi's Data Humanism: data should be used 'to become more humane and connect at a deeper level'" | PARTIAL | Lupi's actual writing says data can be used "to become more humane and to connect with ourselves and others at a deeper level" — close paraphrase. Verified on giorgialupi.com and Pentagram | KEEP, fix exact quote |
| 2 | "Apple Activity Rings: strict design rules (always consistent)" | VERIFIED | Apple HIG explicitly says "Never change the look of the rings by using filters, changing colors, or modifying opacity" | KEEP+CITE HIG URL |
| 3 | "4 concentric rings: Speaking, Reading, Writing, Listening" — adaptation | N/A (design proposal) | This is the file's own adaptation, not a claim about Apple. Apple Activity Rings has 3 rings (Move, Exercise, Stand) | KEEP, but clarify "adaptation, not Apple's actual implementation" |
| 4 | "Small Multiples (Tufte): 'Figure out the key only once'" | PARTIAL | Tufte's actual concept exists but exact quote not found. Multiple paraphrases: "explain a graph once, then rely on that knowledge" | KEEP concept, drop or rephrase quote |
| 5 | "ISOTYPE (Otto Neurath)" | VERIFIED (framework) | Real, well-known visualization framework | KEEP |
| 6 | "Sparkline Comparison (Tufte)" | VERIFIED (framework) | Tufte invented sparklines | KEEP |
| 7 | "Eye tracking: users fixate 2.6x longer on larger items" | PARTIAL | Statistic appears in multiple 2026 design blogs (SaaSFrame, Landdding, Inkbot) but no peer-reviewed citation. Possibly originates from one bento-grid blog and was repeated | KEEP+DISCLOSE: "design-blog rule of thumb (2026)" or find primary source |
| 8 | "40-30-20-10 Space Rule (Linear)" | RETRACT | Linear's actual blog has NO such rule. This is from Improvado's marketing dashboard blog, falsely attributed to Linear | RETRACT attribution. The principle can stay but rename to "Improvado's recommended dashboard ratio" or remove |
| 9 | "Chegg: $7.5M settlement for hidden cancellation" | VERIFIED | FTC.gov press release Sept 2025, confirmed | KEEP+CITE ftc.gov |
| 10 | "ABCmouse: FTC case for difficult cancellation" | VERIFIED | FTC.gov, $10M settlement Sept 2020 | KEEP+CITE ftc.gov |
| 11 | "76% of subscription sites use dark patterns" | VERIFIED | FTC/ICPEN/GPEN July 2024 study of 642 sites, 76% used at least one dark pattern | KEEP+CITE ftc.gov press release |
| 12 | "FTC 2026 Click-to-Cancel Rule — legally mandatory" | RETRACT | The rule was VACATED by the 8th Circuit on July 8, 2025. As of April 2026 there is no enforceable rule. FTC announced ANPRM on March 11, 2026 to restart rulemaking. The file presents a non-existent rule as legally binding | RETRACT and replace with: "FTC restarted negative-option rulemaking in March 2026 after the 8th Circuit vacated the 2024 Click-to-Cancel Rule. Compliance is currently voluntary but enforcement actions continue under existing ROSCA authority" |
| 13 | "Apple HIG: Activity Rings" (source) | VERIFIED | URL exists, content matches | KEEP |
| 14 | "Google Design: 6 Principles for Any Chart (Manuel Lima)" | UNVERIFIED | Generic citation, no URL. Manuel Lima is real, may have written this for Google Design, but not verified | Find URL or drop |
| 15 | "Linear: UI Redesign Part II" (source) | VERIFIED (URL exists) but MISUSED | The URL is real but does not contain the 40-30-20-10 rule attributed to it | KEEP citation, fix the claim it supposedly supports |
| 16 | "Enabling Insights: 11 Principles for EdTech Dashboards" | UNVERIFIED | No URL, vague source name | Find URL or drop |
| 17 | "Nadieh Bremer: visualcinnamon.com", "Nicholas Felton: feltron.com" | VERIFIED (people exist) | These are real data viz designers, but no specific claim is attached | Acceptable as inspiration sources, not citations |

**Overall verdict on this file**: SUSPECT. Two outright fabrications (40-30-20-10 attribution to Linear, FTC click-to-cancel as binding law) plus the design-blog "2.6x" claim that has no primary source. The frameworks (Tufte, Lupi, ISOTYPE, Apple HIG, Bento Grid, Spotify Wrapped) are real and the file's design philosophy is defensible, but the empirical hooks ("eye tracking shows", "Linear's rule", "FTC mandates") are the weakest part and the most quoted in design defense.

**Top fix needed**: Drop or correct the Linear 40-30-20-10 attribution and the FTC click-to-cancel claim. These are the two most legally/factually verifiable items and both are wrong — the rest of the file's credibility hangs on them.

---

## File 3 — platform-current-state.md

### Claims extracted

| # | Claim (verbatim or paraphrased) | Status | Source / Note | Recommended action |
|---|---|---|---|---|
| 1 | "AI tracks 1,500 micro-skills internally (grammar, vocabulary, speech, engagement)" | VERIFIED | Medium/Leta Capital case study, TEFL Academy, Teast.co all cite "1,500 micro-skills" | KEEP+CITE: Leta Capital Medium post (https://medium.com/letavc/how-ai-helped-to-become-a-leading-online-english-learning-school-for-kids-in-europe-novakid-case-fdbd05a63e00). CAVEAT: only Novakid-affiliated/derived sources, no independent audit |
| 2 | Curriculum table (Pre-K through Level 5, lesson counts) | UNVERIFIED in this audit | Specific lesson counts (67, 49, 49, 63, 83, 80) not cross-checked. Likely from Novakid programs page | Verify against novakidschool.com/programs/ |
| 3 | "Lesson format: 5 min pre-lesson + 25 min live + 10 min follow-up" | VERIFIED | TEFL Academy and Teast.co both confirm 25-min lesson format | KEEP |
| 4 | "Brand colors #6d46fc, #4221b8, #ffe606..." | UNVERIFIED in this audit | Likely correct (taken from live site CSS) | Verify by inspecting novakidschool.com |
| 5 | "Mikado font (rounded child-friendly sans-serif)" | UNVERIFIED in this audit | Plausible, not checked | Verify via DevTools |
| 6 | "Logo: Rocket ($100K redesign by BBDO/Proximity 2022)" | PARTIAL | $100K and BBDO/Proximity verified via BusinessWire 2021 press release. Year is 2021 NOT 2022 | KEEP, fix year to 2021 |
| 7 | "Tagline: 'The Right Place to Start'" | VERIFIED | Confirmed in 2021 BusinessWire rebrand announcement | KEEP |
| 8 | "Old Me vs. New Me transformation ads = best performing (CAC -64%)" | VERIFIED | Theperformers.io case study confirms "Old Me vs New Me" reduced CAC by 64% in 6 weeks | KEEP+CITE theperformers.io |
| 9 | "940K parents" | VERIFIED | novakidschool.com/about-us official claim. Note: cumulative, not active | KEEP+DISCLOSE cumulative |
| 10 | "3,700+ certified teachers" | VERIFIED | novakidschool.com homepage "3,700+ certified, caring teachers" | KEEP+CITE homepage |
| 11 | "Only 1 in 11 pass" | VERIFIED | novakidschool.com homepage "Only 1 in 11 passes our 10-step hiring process". Self-reported marketing claim | KEEP+CITE, mark as self-reported |
| 12 | "Europe 80% (Poland, Spain, Italy, Germany, Turkey), MENA 15%, APAC 5%" | UNVERIFIED in this audit | Specific market split not found in 1 search | Find source (likely Tracxn or company press) or qualify |
| 13 | "50+ countries" | VERIFIED | novakidschool.com About Us | KEEP |
| 14 | "Standard $9-12/lesson, Premium $17-20/lesson" | UNVERIFIED in this audit | Pricing varies by region and plan length; not cross-checked | Verify on live site or qualify "approximate, varies by region" |
| 15 | "Family discount: 10% for 2+ kids" | UNVERIFIED in this audit | Not checked | Verify on novakidschool.com/pricing |
| 16 | "Revenue ~$80M+/year" | PARTIAL | Per TheBrandHopper Feb 2026 (secondary source). Tracxn reports $35M (older). Novakid is private | KEEP+DISCLOSE: "estimated ~$80M (secondary sources, 2025)" |
| 17 | "Profitable since 2023" | UNVERIFIED in this audit | Plausible but not directly checked | Verify against TheBrandHopper or Novakid press |
| 18 | "Total funding: $41.4M (Series B)" | VERIFIED | Crunchbase $41.5M, Tracxn $41.4M | KEEP |
| 19 | "24M+ lessons delivered" | VERIFIED | novakidschool.com About Us official | KEEP |
| 20 | "Inc. 5000: #105 overall, #1 in education" | VERIFIED for 2023 list specifically | Already disclosed in trustpilot-systematic.md | KEEP, specify year (2023) |
| 21 | "Trustpilot 4.4/5, 2,527 reviews" | VERIFIED | Currently 2,528 reviews, 4.4/5 — essentially exact | KEEP |
| 22 | Pain point list (billing, support, teacher booking, technical) | VERIFIED qualitatively | Matches systematic Trustpilot read in trustpilot-systematic.md | KEEP |

**Overall verdict on this file**: MOSTLY SOLID. This is the cleanest of the 5 files. Almost every empirical claim about Novakid's company, brand, and product is verifiable against primary or near-primary sources (novakidschool.com, BusinessWire, Crunchbase, theperformers.io). The 1,500 micro-skills claim — the keystone of the entire research thesis — is verified in 3 independent sources (Medium, TEFL Academy, Teast.co), though all trace back to Novakid-affiliated material.

**Top fix needed**: Fix "BBDO redesign 2022" → "2021" (small but factual). Add inline citations to each claim, especially for 1,500 micro-skills, since this is the keystone.

---

## File 4 — trustpilot-systematic.md

### Claims extracted

| # | Claim (verbatim or paraphrased) | Status | Source / Note | Recommended action |
|---|---|---|---|---|
| 1 | Trustpilot domain table (~12,148 total across 7 domains) | VERIFIED (spot-checked 4) | novakidschool.com 2,527→2,528, novakid.es 2,504, novakid.pl 2,243, novakid.it 2,061, novakid.com.tr 1,570 — all match | KEEP |
| 2 | Rating distribution novakidschool.com (82/11/3/1/3%) | VERIFIED | Currently 82/11/3/1/3 — exact match | KEEP |
| 3 | "Topic distribution from 200+ reviews" with %s | PARTIAL (methodology) | The %s are explicitly "estimated from systematic read" — author's qualitative coding, not statistical | KEEP+DISCLOSE: "qualitative estimates, not statistical sample" |
| 4 | All 15 named-quote reviews with dates and pages | UNVERIFIED in this audit (would take 15 individual checks) | Plausible since URLs are specific. Spot-check 2-3 manually before submission | Spot-check 2-3 quotes via direct Trustpilot URL |
| 5 | GoStudent: $779M funding, $3.2B valuation Jun 2024 | PARTIAL | Tracxn confirms $779M and $3.21B Jun 2024. BUT Prosus has since devalued GoStudent to ~€900M (~$970M), losing unicorn status | KEEP+DISCLOSE: "$3.2B peak (Jun 2024); subsequently devalued by Prosus to ~$970M" |
| 6 | "GoStudent Pricing GBP 21.99-30.49/lesson" | UNVERIFIED in this audit | Plausible, not directly checked | Spot-check on gostudent.org/pricing |
| 7 | "GoStudent AI Lesson Summaries: 99% find useful" | UNVERIFIED in this audit | Quote attributed "per GoStudent" — self-reported marketing claim | Verify on gostudent.org/lesson-summaries (URL is in file). Mark as self-reported |
| 8 | Lingokids features (Progress Center, badges, weekly reports) | UNVERIFIED in this audit | URLs are listed (help.lingokids.com), plausible | Spot-check the help center URLs |
| 9 | IXL features (6 reports, SmartScore, Diagnostic) | UNVERIFIED in this audit | URLs listed, plausible | Spot-check ixl.com/analytics |
| 10 | Prodigy, SplashLearn, Cambly Kids feature lists | UNVERIFIED in this audit | URLs listed, plausible | Spot-check before submission |
| 11 | "Duolingo 7-day streak = 3.6x course completion" | VERIFIED | blog.duolingo.com "How the Duolingo Streak Builds Habit" — exact match | KEEP |
| 12 | "EdTech retention 4-27%" | PARTIAL (already disclosed) | 4% from Alchemer 2021, 27% no primary source | KEEP+DISCLOSE as already done in this file |
| 13 | "80% parents educational anxiety" | PARTIAL (already disclosed) | China-only (Family Education Blue Book 2024) | KEEP+DISCLOSE as already done |
| 14 | "Novakid revenue ~$80M+" | PARTIAL (already disclosed) | TheBrandHopper secondary source | KEEP+DISCLOSE |
| 15 | "940K parents", "1 in 11", "24M+", "$41.4M", "Inc 5000 #105" | VERIFIED | All cross-checked against novakidschool.com, Crunchbase, TheBrandHopper | KEEP |
| 16 | "92% of students show improvement within 3 months" | RETRACT (already done in file) | Already correctly retracted as conflated SWOT-analysis claim | Already handled |
| 17 | "Magic Academy 39% improvement (BusinessWire Oct 2022, n=30,390)" | VERIFIED via BusinessWire URL | Already cited correctly | KEEP — this is a strong substitute for the retracted 92% |
| 18 | "TEFL Academy: after-class report < 2 minutes" | VERIFIED | theteflacademy.com/blog/teaching-on-novakid-what-you-need-to-know — exact match | KEEP |
| 19 | "1,500 micro-skills (grammar, fluency)" | VERIFIED | Medium/Leta Capital | KEEP+CITE primary URL |
| 20 | Glassdoor 4.2/5, 375 reviews; English Teachers 3.6/5, 86 reviews | UNVERIFIED in this audit | URLs listed, plausible. Glassdoor often paywalled | Spot-check before submission |
| 21 | Teacher complaints (Heart System, $16/hr, sudden termination) | UNVERIFIED in this audit | Quoted from Glassdoor/Indeed reviews | Spot-check 2-3 quotes |

**Overall verdict on this file**: STRONGEST of the 5 files. This file already has self-disclosed methodology and explicit "VERIFIED" / "PARTIALLY VERIFIED" / "REMOVE" labels for most major statistics (sections 1-10 in P0-5). The author has clearly already done a verification pass. Remaining risks are in the competitor feature lists (which would benefit from a spot-check pass) and the GoStudent valuation framing.

**Top fix needed**: Add the post-Jun-2024 GoStudent devaluation to the funding section. Spot-check 2-3 of the 15 named Trustpilot quotes and 2-3 of the competitor feature claims before submission.

---

## File 5 — formal-ux-audits.md

### Claims extracted

| # | Claim (verbatim or paraphrased) | Status | Source / Note | Recommended action |
|---|---|---|---|---|
| 1 | Nielsen's 10 Heuristics framework | VERIFIED (framework) | Real, well-known | KEEP |
| 2 | Norman's 7 Principles framework | VERIFIED (framework) | From The Design of Everyday Things | KEEP |
| 3 | Cognitive Walkthrough methodology | VERIFIED (framework) | Standard HCI evaluation method | KEEP |
| 4 | Each H1-H10 finding with severity rating | VERIFIED (grounded) | Each finding cites specific screenshots (`02-dashboard-main-homework-feed.jpg`, etc.) and quotes specific UI text. This is direct observation, not desk research | KEEP — strongest evidence type in the entire research corpus |
| 5 | Each DN1-DN7 finding | VERIFIED (grounded) | Same pattern: specific evidence from screenshots | KEEP |
| 6 | "AI internally tracks 1,500 micro-skills — zero of which are surfaced" (H1) | VERIFIED (cross-file) | The number is the same keystone claim as platform-current-state.md, verified above. The "zero surfaced" part is grounded in the screenshot review | KEEP |
| 7 | "1,500 micro-skills" (DN/synthesis sections) | VERIFIED | Same as above | KEEP |
| 8 | Cognitive Walkthrough Tasks 1-5 with step-by-step user actions | VERIFIED (methodology) | Each task shows what user would see based on screenshots. Walkthrough is a documented HCI method | KEEP |
| 9 | "5 of 5 tasks fail (Task 2 partial)" completion rate | VERIFIED (derived from audit) | Conclusion follows from the evidence presented | KEEP |
| 10 | "research personas" reference (Task 4 severity, Task 2) | VERIFIED | References persona interviews in `../02-research-inputs/persona-interviews/` — LLM-simulated design probes, explicitly labelled as generative input (not empirical evidence). Methodology disclosed in `synthesis.md` | KEEP |
| 11 | "89% of personas described wanting to share progress" (Task 4) | RETIRED | Specific % came from the previous persona methodology with pseudo-statistics. Replaced by corpus-anchored claim: "shareable artefacts pattern across 8 persona interviews and corpus 'what we tell grandma' framing" | SUPERSEDED |
| 12 | Recommendations (rebuild conceptual model, surface progress, etc.) | DESIGN OPINION | Not empirical claims | KEEP |

**Overall verdict on this file**: STRONGEST EVIDENTIARY BASE of the 5 files. This is direct first-person observation of the live product with screenshot evidence for every finding. The Nielsen and Norman frameworks are correctly applied (not misattributed). The risk is NOT in this file's audit findings — it's in the 2-3 places where it cross-references statistics from the other (weaker) research files: "45 personas", "89% of personas wanted sharing", and the "1,500 micro-skills" keystone (which is verified, but its only source is Novakid-derived).

**Top fix needed**: Add inline disclosure where it imports from synthetic persona work ("45 personas" and "89%"). Otherwise this is the cleanest file in the corpus and the strongest evidence for the redesign thesis.

---

## Cross-file patterns

1. **The 1,500 micro-skills claim appears in 3 files** (platform-current-state.md, formal-ux-audits.md H1 + DN, trustpilot-systematic.md P1-9) with consistent framing. All three trace to the same root: Medium/Leta Capital case study, reproduced by TEFL Academy and Teast.co. **Good news**: consistent across files. **Bad news**: single root, all Novakid-friendly sources, no independent audit. The thesis still stands but the citation should disclose source quality.

2. **The "80% educational anxiety" claim appears in parent-psychology.md without disclosure but in trustpilot-systematic.md WITH disclosure** that it's China-only. This is a cross-file inconsistency: one file uses the headline, the other corrects it. The corrected version should propagate back to parent-psychology.md.

3. **The "EdTech retention 4-27%" range appears in parent-psychology.md and trustpilot-systematic.md.** Both should harmonize to the disclosed version: "4% annual (Alchemer 2021)" and qualify the 27% upper bound as estimate.

4. **Generic source citations without URLs**: parent-psychology.md and dataviz-patterns.md both end with "Sources" sections that list organization names ("NNGroup", "Baymard", "RevenueCat", "Google Design", "Enabling Insights") without URLs or specific findings. This is the same pattern that compromised competitor-analysis.md. These should be expanded to specific URLs or removed.

5. **The two clearest fabrications are both in dataviz-patterns.md**: the Linear 40-30-20-10 rule and the FTC 2026 click-to-cancel "legally mandatory" claim. Both look authoritative but neither survives a 1-search check. This file should get the most aggressive editing pass.

6. **The formal-ux-audits.md file does NOT exhibit these patterns** because its evidence is screenshots and direct observation, not desk research. Where it imports stats from other files (45 personas, 89%), those imports should be disclosed.

---

## Recommended fix priority

| Priority | File | Action | Estimated time |
|---|---|---|---|
| P0 | dataviz-patterns.md | RETRACT "FTC 2026 Click-to-Cancel Rule legally mandatory" — replace with current legal status (vacated 8th Cir Jul 2025, ANPRM Mar 2026) | 10 min |
| P0 | dataviz-patterns.md | RETRACT Linear 40-30-20-10 attribution — drop or rename to "Improvado dashboard ratio" | 5 min |
| P0 | parent-psychology.md | RETRACT "26 milestones / CEFR" misattribution — CEFR has no such structure | 10 min |
| P0 | parent-psychology.md | DROP or source the unverified "85% parents low marks" and "guiltier than any other type" claims | 10 min |
| P1 | platform-current-state.md | Fix BBDO logo redesign year 2022 → 2021 | 2 min |
| P1 | platform-current-state.md | Add inline citation to the "1,500 micro-skills" claim with disclosure that source is Novakid-affiliated | 10 min |
| P1 | trustpilot-systematic.md | Add post-Jun-2024 GoStudent devaluation note to funding section | 5 min |
| P1 | parent-psychology.md | Expand the "Sources" section: replace org-name list with specific URLs OR delete unsourced claims | 30 min |
| P1 | dataviz-patterns.md | Same: expand "Sources" section with URLs, drop unverified specific stats (eye-tracking 2.6x without primary source) | 20 min |
| P2 | trustpilot-systematic.md | Spot-check 2-3 of the 15 named Trustpilot quotes against the cited URLs | 20 min |
| P2 | trustpilot-systematic.md | Spot-check competitor feature lists (Lingokids, IXL, Prodigy, SplashLearn) against cited URLs | 30 min |
| P2 | formal-ux-audits.md | Add inline disclosure where "45 personas" and "89%" are imported from synthetic persona work | 10 min |
| P3 | parent-psychology.md | Harmonize "80% educational anxiety" across files — propagate the China-only disclosure | 5 min |
| P3 | parent-psychology.md | Harmonize "4-27% retention" disclosure across files | 5 min |

**Total estimated cleanup time**: ~3 hours.

---

## Method notes

- WebSearch was used for initial discovery, WebFetch for direct content checks (Trustpilot pages, Linear blog).
- "Verified" means the claim was confirmed in 1-2 searches against a primary or strong-secondary source. Not all verifications were independent — many trace back to the same root (e.g., 1,500 micro-skills traces to Leta Capital).
- "Unverified" means 1-2 searches returned no source matching the specific claim or numeric value. Some unverified items may still be real but require deeper search.
- "Retract" means an active contradiction was found (e.g., Linear's actual blog does not contain the 40-30-20-10 rule; the FTC click-to-cancel rule was vacated in July 2025).
- The formal-ux-audits.md file was treated as direct observation, not desk research, so its findings were spot-checked for framework attribution and cross-file imports rather than for empirical sources.