Chameleon · marketing source sheet · not linked from the app

Every number here is computed, not generated.

Thirteen product surfaces from Chameleon's insight and analytics system, rebuilt as standalone Marginalia sections for landing-page use. Each one shows what it does, what data it is computed from, and a ready-to-use landing line.

To lift a section: copy its <section> + the shared .sc styles. Tokens are 1:1 the app's --marg-* set; inside the app delete the :root block and dark mode works for free. Chart palettes are validator-passed (see the head comment).
Group A

The headline instruments

Readiness · exam-weighted

One honest number for "am I ready?"

Not a completion bar. Readiness weights every knowledge atom by how critical it is and how likely it is to be examined, then compares you against the pace you would need to hit your target by exam day.

READINESS — EXAM-WEIGHTED 30 DAYS TO EXAM · 72 CARDS REVIEWED You: 52% exam-weighted readiness YOU 52 PACE 27 TARGET 85 Ahead of pace. Hold the line and the target takes care of itself.
Meter · you vs pace-ramp vs target
What it showsCurrent exam-weighted readiness (YOU) against the linear ramp you would need today (PACE) and your configured target. The verdict line is deterministic: ahead / on track / behind, with a ±3 pt tolerance band.
Computed fromFSRS memory state per flashcard × atom criticality (CORE/EXTENDED/PERIPHERAL) × exam relevance. Quiz questions count toward coverage, and missed quiz questions become new cards in the deck, so test performance flows into the number too. Sample size travels with it.
"Most apps show progress. Chameleon shows readiness, weighted by what your exam actually tests."
Readiness v2 · evidence-blended

A readiness score that doesn't trust your self-grades

Flashcard apps ask you to grade yourself, then believe you. Chameleon measures each student's self-grading bias and corrects for it, blends in objective quiz evidence (keyed answers, AI-graded written work), prices in the whole syllabus, and projects the result to exam day.

READINESS — EVIDENCE-BLENDED EXAM TYPE: MIXED → WEIGHTS .40 / .25 / .35 MEMORY (FSRS, BIAS-CORRECTED) RECOGNITION (KEYED QUIZ, VERIFIED ANSWER KEYS) WRITTEN (AI-GRADED PRODUCTION) Memory channel: 54% after a measured −11pt self-grading correction · 156 reviewed cards Recognition: 66% recency-weighted keyed-answer accuracy Written: 16% — the AI-graded free-text answers are weak; a mixed exam weights this at .35 54% −11 PTS BIAS 66% 16% TESTED 59% WHOLE SYLLABUS 51% EXAM DAY 21% 22 APPLY-LEVEL TOPICS CAPPED — RECALL-ONLY EVIDENCE · READ-BUT-UNTESTED FLOORS AT 15%, NEVER HIGHER
Three evidence channels · two denominators · decayed to exam day
What it showsThree evidence channels blended by exam format: an open-answer exam trusts written work most, a multiple-choice exam trusts recognition. The memory channel is shrunk by the student's own measured overconfidence (their predicted-vs- actual recall record). "Whole syllabus" prices in every core topic, tested or not; "exam day" decays each channel forward on its own forgetting curve.
Computed fromFSRS states × the calibration ledger × verified-key quiz answers × AI rubric scores × Bloom-level atom tags × reading dwell (as a floor only — looking at a card is never knowing it). Deterministic; ships behind a flag until its backtest beats the memory-only model on real exam-answer data.
"It knows the difference between what you'd recognize, what you'd recall, and what you could write from scratch on exam day."
Forgetting forecast · FSRS

See the forgetting before it happens

Every card carries a live memory model. Chameleon projects each one forward and warns you which topics you know today will have faded below target by exam day.

85% 0% EXAM DAY TODAY A cramming-only card: fades within days Reviewed twice: holds ~2 days above target Reviewed three times: ~3 days above target Well-spaced card: ~4 days above target First topic to fade: crosses your target in under a day without review A card FSRS scheduled well: still near target at exam day 42 topics you know today cross the line before exam day · the first within ~3 days
Retrievability projection per card · target line at 85%
What it showsEach curve is one card's projected memory decay from its current FSRS stability. Cards that cross your target before the exam-day marker are counted and surfaced as "N topics will fade… the first within ~K days".
Computed fromts-fsrs stability + elapsed days per card, the plan's exam date, and the retrievability threshold. Pure math, no model calls.
"Your brain has a forgetting curve. Chameleon schedules the exact day to interrupt it."
What stands out · 10 deterministic rules

Insights that show their evidence

Ten behavioral rules read your study exhaust and only speak when the sample size clears a threshold. Every insight carries its evidence and an honesty badge, and low confidence is labeled a hypothesis, never a fact.

11 cards are "leeches", you keep forgetting them. They're eating your review time; tackling them directly is the highest-leverage fix.

Evidence: "What is a positive statement…" · 7 lapses · "Why does diminishing marginal product…" · 6 lapses
HIGHreview these ☞

Your accuracy drops off in long sessions: 57% in 45-minute-plus sittings vs 75% in your shorter ones. Shorter, more frequent sessions keep retention higher.

Evidence: long sessions 57% over 8 · shorter sessions 75% over 16
HIGHtry shorter sessions ☞

Your written answers on "Income and Substitution Effects" are coming up short, averaging 1.3/10. That usually points to an understanding gap rather than a memory one.

Evidence: avg 1.3/10 across 3 written answers · from your own AI feedback: "you recognized that income is rising and tried to reason about…"
LOW — A HYPOTHESIS, NOT A FACTrevisit this concept ☞
The full rule setLeeches · overdue debt · read-vs-test balance · importance×weakness · forgetting forecast · time-of-day · session fatigue · cramming vs spacing · you-now-vs-then · written-answer struggle.
Computed fromReview logs, FSRS state, study sessions, quiz answer records, and AI grading feedback. No LLM writes these numbers.
"No vague AI advice. Every insight names its evidence, and tells you when it's only a hypothesis."
Vs last week · you against you

The only baseline is your own past self

No leaderboards, no peer pressure. Chameleon snapshots readiness weekly and reports the delta against you, and only you.

Exam readiness · this week
52%
▲ +9 PTS VS LAST WEEK · 3 WEEKS OF HISTORY
Stat tile · weekly analytics snapshot
What it showsThis week's exam-weighted readiness, the signed delta vs the prior weekly snapshot, and the history depth so the trend is never claimed on thin data (two weeks minimum before a delta is stated).
Computed fromWeekly analytics_snapshots per plan. The sparkline is the raw snapshot series.
"No peer comparison, anywhere. The product philosophy is a feature."
Group B

The diagnostic panels

Calibration · predicted vs actual recall

Find out if you know what you know

Before every review, the memory model predicts your recall odds. Chameleon keeps score of prediction vs outcome, so overconfidence shows up as a measurable gap, not a feeling.

PERFECT CALIBRATION +15 PT OVERCONFIDENT Predicted 35% → actually recalled 20% Predicted 45% → actual 40% Predicted 55% → actual 40% Predicted 65% → actual 60% Predicted 75% → actual 60% Predicted 85% → actual 70% Predicted 95% → actual 80% PREDICTED RECALL → ACTUAL → Below the line = surer than you should be · RMSE 0.13, tracked weekly
10-percentile prediction buckets · identity line = perfectly calibrated
What it showsEach dot is a prediction bucket: what the model predicted vs how you actually did. Below the diagonal = overconfidence. RMSE and log-loss trend week over week as your data accumulates.
Computed fromA (predicted R, outcome) pair logged at every grade, binned by decile. This is the standard FSRS calibration methodology, per student.
"The difference between feeling ready and being ready, measured."
Recognition vs recall · by question type

Recognizing is not the same as knowing

Multiple choice flatters you. Chameleon splits accuracy by question type and exposes the gap between recognizing an answer and producing one.

Multiple choice True / false Numeric input Multi-select Fill in blank Open-ended Multiple choice: 90% (9/10) True/false: 86% (6/7) Numeric input: 63% (5/8) Multi-select: 60% (3/5) Fill in the blank: 50% (2/4) Open-ended: 43% (3/7) 90%86% 63%60% 50%43% 47-PT GAP A 30+ point recognition→production gap is the classic comprehension flag.
Accuracy per question type · seven types generated per chapter
What it showsAccuracy across the seven generated question types, sorted. Recognition formats up top, production formats below, and the gap between them called out as the diagnostic.
Computed fromEvery quiz answer record joined to its question type. Free-text answers are AI-graded against a rubric; everything else is keyed.
"Quizlet tells you your score. Chameleon tells you which kind of knowing you're missing."
When you study best · time of day

Your sharpest hour, from your own record

Chameleon aggregates accuracy by hour of day and finds your peak window, then the insight engine suggests scheduling your hardest material inside it.

7 AM: 71% · n=22 8 AM: 79% · n=64 9 AM: 88% · n=108 — your peak 10 AM: 82% 11 AM: 76% 12 PM: 73% 1 PM: 70% 2 PM: 72% 3 PM: 74% 4 PM: 72% 5 PM: 69% 6 PM: 67% 7 PM: 65% 8 PM: 63% 9 PM: 64% 10 PM: 61% 9 AM · 88% 7A12P 5P10P Mornings run ~20 points sharper than your late evenings.
Emphasis form · peak window highlighted, context in gray
What it showsReview accuracy bucketed by hour across the last 30 days. The peak hour is highlighted; the rest stays as quiet context. Only speaks with ≥3 sessions and ≥20 reviews per band.
Computed fromStudy-session start times × items reviewed × items correct.
"Chameleon found my study sweet spot: 9 AM, 88% recall. My 10 PM sessions were basically decorative."
Diminishing returns · session length

The 45-minute cliff

Accuracy by how long you've been sitting. When your long sessions measurably underperform your short ones, the fatigue insight fires with the numbers.

70% Under 15 min: 74% 15–30 min: 76% 30–45 min: 72% 45–60 min: 60% 60+ min: 57% −12 PTS PAST 45 MIN <15M15–30 30–4545–60 60+
Single series · accuracy vs sitting length, 30-day window
What it showsWhere your personal fatigue point sits. The companion insight only fires when short and long sessions both have enough data (≥3 sessions, ≥20 reviews each) and the gap clears 12 points.
Computed fromSession duration × accuracy, bucketed.
"Study smarter isn't a platitude when the app can show you your own 45-minute cliff."
Consistency · 30-day heatmap

Spacing you can see

The research says spaced beats crammed (d≈0.85, the strongest effect in learning science). The heatmap makes your spacing, or your cramming, impossible to miss.

Mon: no reviews Tue: 8 reviews Wed: 19 reviews Thu: no reviews Fri: 6 reviews Sat: no reviews Sun: 15 reviews Mon: 9 reviews Tue: no reviews Wed: 26 reviews Thu: 7 reviews Fri: 14 reviews Sat: no reviews Sun: 11 reviews Mon: 17 reviews Tue: 8 reviews Wed: 24 reviews Thu: 16 reviews Fri: 10 reviews Sat: 13 reviews Sun: 31 reviews — heaviest day Mon: 18 reviews Tue: 9 reviews Wed: 22 reviews Thu: 15 reviews Fri: 8 reviews Sat: 21 reviews Sun: 14 reviews Mon: 29 reviews Tue: 16 reviews Wed: 7 reviews Thu: 23 reviews Fri: 12 reviews — today LESS MORE CURRENT STREAK 12 days BEST 18 · 214 REVIEWS THIS MONTH SPACING BEATS CRAMMING, d ≈ 0.85
Sequential single-hue ramp (validated) · streak from the review log
What it showsReview volume per day for the last 30 days plus the live streak. The cramming-vs-spacing insight reads the same distribution: if half your reviews land on your two busiest days, it says so.
Computed fromThe append-only flashcard review log.
"Streaks you can defend to a learning scientist."
Interleaving · mixing score

Are you mixing, or grinding one chapter?

Interleaved practice strengthens long-term retention. Chameleon scores how much you mix chapters within sessions against the research-backed 40–60% zone.

Interleaving score: 52% of adjacent reviews switch chapter TARGET ZONE 40–60% YOUR MIX 52% ROUND-ROBIN QUEUE DOES THE MIXING FOR YOU
Meter with target band · Bjork's interleaving zone
What it showsThe share of adjacent reviews that switch chapters, against the optimal band. The review queue itself round-robins across chapters, so the default behavior lands you in the zone.
Computed fromChapter transitions inside each study session, 14-day window.
"The queue interleaves for you. The meter proves it's working."
Group C

The knowledge map underneath

Weakest concepts · exam-weight × weakness

What to study next, ranked honestly

Not your lowest scores. Chameleon ranks by exam weight × weakness, so a shaky core concept outranks a shakier footnote, and tells you what has never been tested at all.

WEIGHT 95 WEIGHT 95 WEIGHT 95 WEIGHT 90 WEIGHT 80 Utility, marginal utility, and substitution Consumer optimization and bundle selection Price elasticity of demand Income and substitution effects Market equilibrium 22% mastery · 8 atoms assessed 31% mastery · 6 atoms assessed 38% mastery · 7 atoms assessed 44% mastery · 5 atoms assessed 58% mastery · 9 atoms assessed 22%31%38% 44%58% † 67 CORE POINTS NOT YET TESTED BY ANY CARD — SURFACED SEPARATELY, NEVER COUNTED AS "WEAK"
Do-first at top · weight chips in Spline Sans Mono
What it showsTop five concepts by exam-weight × (1 − mastery). Untested material is listed as a coverage gap, never mislabeled as weakness, honesty the panel states in its own footer.
Computed fromConcept→atom mastery rollup: FSRS retrievability per covering card, criticality-weighted per atom.
"Ranked by what moves your grade, not by what's easiest to show."
Per-source mastery · your files

It knows which of your PDFs you're weak on

Everything is generated from the student's own uploads, so mastery can be attributed back to each file. A low-mastery source is a re-read candidate.

MICRO-TEXTBOOK.PDFBOOK LECTURE-NOTES-W1-6.PDFNOTES SLIDES-WEEK3.PPTXSLIDES 45% · 40 of 94 cards mastered 62% · 33 of 53 cards mastered 71% · 17 of 24 cards mastered 45% · 40/94 62% · 33/53 71% · 17/24
Single hue · mean retrievability across each file's cards
What it showsMean card retrievability grouped by the uploaded file each card was generated from, with the mastered-count fraction.
Computed fromCard→source attribution recorded at generation time (exact on single-source chapters), joined to FSRS state.
"Your textbook, your notes, your slides, and a mastery bar for each."
Atom mastery · the knowledge graph

Mastery at the resolution of a single fact

Chameleon decomposes your material into typed knowledge atoms (definitions, procedures, computations…) and tracks each one: mastered, shaky, or never yet tested. This is the substrate every other number rolls up from.

DEFINITIONCONCEPT PROCEDURECOMPUTATIONAL RELATIONEXAMPLE Definitions mastered: 18 Shaky: 7 Untested: 6 Concepts mastered: 14 Shaky: 9 Untested: 8 Procedures mastered: 9 Shaky: 11 Untested: 5 Computations mastered: 6 Shaky: 8 Untested: 12 — the gap the coverage footnote points at Relations mastered: 11 Shaky: 6 Untested: 4 Examples mastered: 12 Shaky: 5 Untested: 7 18 · 7 · 614 · 9 · 8 9 · 11 · 56 · 8 · 12 11 · 6 · 412 · 5 · 7 MASTERED SHAKY NEVER TESTED
Typed atoms × status · 2px page gaps between segments
What it showsEvery knowledge atom the pipeline extracted, grouped by kind, in three honest states. "Never tested" is kept separate from "weak", a distinction most apps collapse.
Computed fromThe atom graph (kind, criticality, exam relevance) × card coverage × FSRS retrievability. Computational atoms in math courses are additionally SymPy-verified at generation time.
"Flashcard apps track cards. Chameleon tracks the 300 individual facts inside them."