The Science Behind 95% Accuracy

How CLT and Backward Reconciliation Deliver Research-Grade Nutrition Tracking

2.4-12x fewer tracking errors. The difference between hoping and knowing.

What Does 95% Accuracy Even Mean?

Understanding confidence intervals and error bars

The Statistical Definition

When we say "95% accuracy", we mean:

"95% of the time, the true value is within ±5% of our estimate"

This is a 95% confidence interval (CI) or 2σ (two sigma) in statistical terms.

Error Distribution (Gaussian/Normal)

68% CI (1σ):

±2.5% error bar

68% of measurements fall within this range

95% CI (2σ):

±5% error bar

95% of measurements fall within this range

Example: If Eatomate estimates you ate 2,000 kcal today at 95% accuracy:

  • True intake is between 1,900-2,100 kcal (±5%)
  • We're 95% confident the true value is in this range

Our Conservative Baseline Assumptions

Home-Cooked Meals (Pre-Reconciliation)

Baseline: 70% accuracy (±30% error at 2σ)

Why 70%? Recipe databases have inherent uncertainty:

  • Generic "chicken breast" varies ±10-20% by brand/cut
  • Your cooking method differs from database assumptions
  • Oil/seasoning amounts vary per recipe

This is conservative — professional databases are typically 75-80% accurate, but we use 70% to ensure our claims hold even with worst-case database quality.

Restaurant Meals (With Smart Scale Weight)

Baseline: 70% accuracy (±30% error at 2σ)

Why 70%? Even with precise weight measurement:

  • Unknown ingredient composition (hidden oils, sauces)
  • Database match may not reflect actual recipe
  • Commercial cooking uses more fat/sugar than home recipes

No reconciliation possible — no pantry tracking for restaurant ingredients.

Restaurant Meals (AI Estimation Only)

Baseline: 50% accuracy (±50% error at 2σ)

Why 50%? AI (Mistral) estimates average portion sizes from just meal item names:

  • No weight data — relies on generic average portions
  • Portion sizes vary wildly by restaurant and individual
  • Unknown cooking methods and ingredient quantities
  • Hidden ingredients (sauces, oils, butter) not captured

Without weighing or pantry data, the AI can only guess at average portions.

Multi-Ingredient Packaged Products (Barcode Labels)

Baseline: 90% accuracy with 1.1x systematic inflation

FDA Regulation (21 CFR 101.9(g)): Allows manufacturers to understate calories by up to 20% without penalty. This creates systematic bias.

Why 1.1x inflation factor?

  • FDA rule: Actual calories cannot exceed 120% of labeled value
  • Manufacturers systematically understate labels to stay within legal limits
  • Average actual value is 1.1x (midpoint between 1.0x and 1.2x)
  • We multiply barcode values by 1.1x to correct for systematic understatement

Remaining error sources (why not 99%?):

  • Batch-to-batch variation (±3-5%)
  • Rounding errors on label (e.g., "10g protein" could be 9.5-10.4g)
  • Manufacturing tolerances

Not reconcilable — pre-packaged products don't go through pantry (consumed directly), so no backward reconciliation possible.

Why These Conservative Estimates Matter

By using worst-case baseline accuracy (70% for recipes, 50% for visual estimation), our improvement claims are defensible even in unfavorable conditions.

If actual baseline accuracy is better (e.g., 75% instead of 70%), our reconciliation improvements are even more impressive. This is scientific conservatism — we'd rather under-promise and over-deliver.

Accuracy Improves Over Time

Physics-based reconciliation learns your cooking patterns and corrects historical estimates

From Guessing to Knowing

Error reduction isn't about percentage points — it's about control

Reconciliation Isn't About Percentage Points — It's About Error Reduction

Going from 88% to 95% accuracy sounds like a modest 7 percentage point improvement. But the real story is in the remaining error:

12% → 5%
Error drops by 58%
2.4x
Fewer tracking errors

That 2.4x error reduction is the difference between your 500 kcal deficit being eaten down to 260 kcal vs staying at 400 kcal. It's the difference between losing 6 lbs and 10+ lbs over 12 weeks.

Typical Usage (10% restaurant meals)
Before reconciliation:88% accuracy
12% error
↓
After reconciliation:95% accuracy
5% error
2.4x
Fewer Errors
58% error reduction
Cook at home →
Best Case (100% home cooking)
Before reconciliation:88% accuracy
12% error
↓
After reconciliation:99% accuracy
1% error
12x
Fewer Errors
92% error reduction

The more you cook at home, the more powerful reconciliation becomes.
Restaurant meals can't be reconciled (no pantry tracking), but home-cooked meals achieve research-grade 99% accuracy.

The Hidden Danger: When Tracking Errors Eat Your Deficit

You're trying to lose 1 lb/week with a 500 kcal/day deficit. But what if your tracking is wrong?

Without Reconciliation

88% accuracy

Your planned 500 kcal deficit
Worst case: Tracking overestimates by 240 kcal

Planned deficit:
500 kcal
Tracking error:
-240 kcal
Actual deficit:
260 kcal

48% of your deficit was eaten by errors!

Real weekly weight loss: 0.5 lbs instead of 1.0 lbs

After 12 weeks:
  • Expected: -12 lbs
  • Actual: -6 lbs

"I'm doing everything right but the scale won't budge. Nutrition tracking doesn't work for me."

With Reconciliation

Typical — 95% accuracy

Your planned 500 kcal deficit
Worst case: Tracking overestimates by 100 kcal

Planned deficit:
500 kcal
Tracking error:
-100 kcal
Actual deficit:
400 kcal

Only 20% of your deficit lost to errors

Real weekly weight loss: 0.8 lbs (on track!)

After 12 weeks:
  • Expected: -12 lbs
  • Actual: -9.6 lbs

"I'm in control. If I'm not losing fast enough, I adjust my plan — I don't question my tracking."

Best Case

99% accuracy — 100% home cooking

Your planned 500 kcal deficit
Worst case: Tracking overestimates by 20 kcal

Planned deficit:
500 kcal
Tracking error:
-20 kcal
Actual deficit:
480 kcal

Only 4% of your deficit lost to errors

Real weekly weight loss: 0.94 lbs (research-grade precision!)

After 12 weeks:
  • Expected: -12 lbs
  • Actual: -11.3 lbs

"Research-grade precision. Every week is predictable within normal body weight variance."

How We Achieve 95%+ Accuracy

Two mathematical techniques working together

Central Limit Theorem (CLT)

Recipe databases have ~70% accuracy (±30% 2σ error per ingredient). But you don't eat single ingredients — you eat meals with multiple ingredients.

The Magic: Independent errors partially cancel out through quadrature sum (RMS).

For N independent recipe groups:

Combined error = σ / √N

  • Week 1: 6 recipe groups → 30% / √6 = 12.2% error at 2σ (88% accuracy)
  • Partial cancellation: Not random chance — it's statistics
  • Meal grouping: Same recipe = correlated errors (tracked as one group)

Backward Reconciliation

CLT gets you to 88% accuracy, but that's still ±240 kcal/day uncertainty. Reconciliation uses physics to lock in ground truth.

The Physics: Mass conservation — weigh your pantry, measure actual consumption.

When milk carton empties:

consumed = initial + purchased - final

  • Ground truth: Smart scale measures actual grams consumed (±0.5% precision), or barcode scans already tell us true weights
  • Work backward: Correct all historical meals that used that ingredient
  • Your recipes: System learns your cooking patterns, not generic database

CLT + Reconciliation = 2.4-12x Fewer Errors

CLT gives you a strong baseline (88% accuracy). Reconciliation corrects it to research-grade precision (95-99% accuracy).

70%
Recipe DB alone
88%
+ CLT (6 groups)
95-99%
+ Reconciliation

This Is the Difference Between Hoping and Knowing

Without Reconciliation

  • •"I hope I logged everything accurately"
  • •"I hope the database is right"
  • •"I hope this will work"
  • •"Why am I not losing weight?"

With Reconciliation

  • •Physics-based verification of actual consumption
  • •2.4-12x fewer tracking errors
  • •Predictable, controllable outcomes
  • •"I know exactly where I stand"

95% accuracy isn't just a number.

It's the confidence to trust your plan.

It's the ability to make data-driven adjustments.

It's the difference between frustration and results.

Experience Research-Grade Tracking

No credit card required. Cancel anytime.

GDPR Compliant
Your data stays private

Technical Details

Monte Carlo Simulation Methodology▼

All accuracy claims are based on Monte Carlo simulationswith 100-1,000 trials to eliminate variance from individual "lucky" runs.

Simulation parameters:
  • Recipe database: Gaussian distribution (mean=1.0, σ=15%) → 70% accuracy at 2σ (±30% error)
  • Meal grouping: 6 independent recipe groups per week
  • Restaurant meals: 10% of meals, stdDev=25% (50% error at 2σ)
  • Smart scale precision: ±0.5% (matching consumer scale specs)
  • Confidence intervals: 2σ (95% CI) for all reported values
Gaussian Error Propagation (Quadrature Sum)▼

Independent error sources combine via quadrature (RMS) rather than simple addition, allowing partial error cancellation per the Central Limit Theorem.

For N independent sources with error σ:

σ_combined = √(σ1² + σ2² + ... + σN²) = σ / √N

Example: 6 recipe groups at 30% (2σ) error each:

σ_combined = 30% / √6 = 12.2% (2σ)

Accuracy = 100% - 12.2% = 87.8% ≈ 88%

Why Restaurant Meals Lower Accuracy▼

Restaurant meals cannot be reconciled because there's no pantry tracking (you don't buy restaurant ingredients). This limits their accuracy to AI estimation (~50%).

Impact on overall accuracy:

  • 0% restaurant: 88% → 99% (12x error reduction)
  • 10% restaurant: 88% → 95% (2.4x error reduction)
  • 30% restaurant: 88% → 90% (1.2x error reduction)

Recommendation: Keep restaurant meals to 2-3 per week for optimal accuracy

Read the full mathematical methodology→