OPEN NOTEBOOK
The wing science notebook.
Our wing science combines published tasting scores, provisional sensory descriptions, editorial coverage, public-review evidence, cultural history, search demand and a limited social-visibility pilot. Different questions use different models; we show which sources and assumptions enter each score.
Jon’s seed covers 44 states with 50 provisional candidates. No state has ten candidates in his list. States with no Jon seed candidates: California, Iowa, Kansas, Nevada, New Jersey, Ohio. The expanded national census contains 1,380 records across all 50 states plus D.C., combining Jon’s seed, Sam’s visits and community suggestions. These are research leads, not 1,380 verified branches.
Ways to Compare
Jon’s List — default: ranks the expanded census using Jon estimates for Sam-scored visits and historical scores for Jon-only records. Reset returns to this list. The original Wings 50 records are retained as historical data.
Sam Reid’s list — unchanged: all 87 ranked entries plus two N/A visits from his All Restaurants tab. Names, scores, source order and explicit tie-breaks are retained. The repeated Local visit remains a separate unranked row. Source snapshot: 2026-09-15 01:48:48 UTC. Search/state filters show subsets; they do not change Sam’s national ranks.
Sam Reid’s list — as a weight: retains only Jon’s 50 candidates. Each list’s rank becomes a 0–100 index: 100 × (list size − rank) / (list size − 1). Default blend: 50% Jon index + 50% Sam index. The slider requests a new ranking with Sam’s share from 0–100%. At 0%, the order is exactly Jon’s; at 100%, matched candidates use Sam’s index while unmatched candidates still retain Jon’s index.
There are 21 reviewed name/alias-and-state associations and 29 unmatched candidates. Unmatched candidates keep their Jon-only rank index; their absolute rank can still move as others change. Matches are provisional at restaurant/market level because the sheet provides no branch addresses. Blend ties retain Jon’s prior order. This is a rank aggregation heuristic across different candidate sets, not a calibrated taste probability or a combined tasting score.
Sam’s five ten-point categories are appearance, meatiness, texture, sauce and overall experience. His notes describe a subjective grading scale and personal tie-breaks. These source observations are kept separate from the model-coded sensory factors.
The Expanded Sam Census
All six spreadsheet tabs are preserved in the source snapshot. The master suggestions contain 1,975 nonempty rows, grouped conservatively into 1,403 name/city/state groups. Eleven groups lack a state and remain outside state rankings. Repeated suggestions are not independent votes. Names and cities can still overlap across branches or aliases.
Sam’s State vs. State preset preserves his separate 51-representative ordering. The National Census preset combines sourced visits and suggestions with Jon’s seed; suggestion-only records remain unranked until usable evidence exists. State pages distinguish scored candidates from research leads rather than inventing a verified top ten.
Sam’s meatiness score maps directly to meatiness on a 100-point scale. His texture, sauce, appearance and overall-experience scores have separate controls. Texture does not prove crispiness or juiciness; sauce does not establish heat, tang or sweetness. Regional awards and recalled heat challenges retain their original scope. Suggestion comments are not converted into recurring negative-review themes.
Jon Estimates and Public Review Coverage
The Jon’s List preset assigns a model-derived Jon score to every scored Sam visit across all 50 states and D.C. It uses a declared preference prior and a tested calibration to historical Discovery Scores, not actual tasting labels. Original lists remain unchanged.
Custom taste presets
The Open-minded mix and other sensory presets use the original adjustable engine. Editing a sensory slider leaves any fixed-list mode and switches to Your own mix. Jon’s historical aggregate scores are excluded in custom modes. Sam’s measured categories contribute only through the corresponding enabled controls. Sliders are disabled in Sam’s unchanged view to protect his order.
The arithmetic
Every enabled factor has a weight from 1 to 5. Zero removes it entirely. Descriptive seed signals are manually coded: 75 for a clear positive/intensity signal, 100 for an explicitly extreme signal. Unmentioned attributes are unknown. A 0 is only an explicit low intensity or traditional style, never a blank.
Heat, sweetness and style use match = 100 − |seed intensity − your target|. Other factors reward stronger signals. No personal like/dislike labels are used.
50 is a disclosed neutral prior, not an observed rating. This pulls thin evidence toward the middle rather than treating missing data as bad wings. With no known active factors, a candidate stays unranked. With every factor off, there is no ranking.
Coverage is the share of active weight with inputs. The missing-factor sensitivity range lets unknown inputs vary from 0 to 100. It is not a confidence interval: coded descriptions, review selection, correlated factors and location ambiguity create additional uncertainty. Close ranks are not proof of quality differences. Exact ties share a rank and display alphabetically.
Public ratings & confidence
The seed has no verified platform/location/rating/count pairs. Historical numbers without that context are not imported. Public-only mode therefore leaves all 1,380 census records unranked. The engine supports future location-verified observations, using (n × rating + 100 × 3.5) ÷ (n + 100), then converting from five stars to 100. The 3.5-star prior and 100-review strength are declared prototype choices, not a fitted empirical-Bayes model. Platforms are averaged equally after shrinkage. Confidence uses capped log review volume only when enabled; it does not measure taste.
Contrarian means evidence, not invention
Most profiles have no auditable negative-review corpus. They say so. Wingnutz includes a narrowly scoped counterpoint from two indexed Reddit discussions, with source links and access limitations; it is not a representative Google/Yelp analysis or a ranked set of worst reviews. Positive-versus-negative taste tensions are labeled separately as hypotheses.
One shared calculation service
Rankings, model estimates and wingonomics conversions are calculated centrally from the shared database. Change a ranking slider and WingMath requests a new result using the same rules. The browser displays the results, draws charts and keeps your controls for the current page visit; it does not run a separate ranking model. An internet connection is required. If the service is unavailable, the page asks you to retry instead of inventing a replacement score.
Database records preserve source dates and earlier observations. Shared results may be cached for up to 30 seconds. This keeps repeated requests efficient; it does not turn saved market quotes or restaurant reviews into a live feed. Cards retain clearly labeled photo placeholders and links to original photo sources.
Continuing the Census
The data model separates stable candidate IDs, state/market, location verification, raw observations, source records, photos with rights, and scoped review themes. Future work should split chains into verified branches, record sauce and visit conditions, retain independent sources, and evaluate predictions on held-out tastings. Discovery Alpha is not shown because a defensible public-rating baseline is missing.
ATL 10: wing science with an Atlanta accent
The Atlanta model compares 19 contenders using Infatuation, Eater, two Reddit threads, dated Google and Yelp snapshots, documented cultural history, search demand and a social-visibility pilot. Editorial list order is respected: an unranked geographical list is not converted into a preference order. Restaurant branches and unresolved brand matches are disclosed.
The base blend is 40% editorial evidence, 15% Reddit recommendations, 20% public reviews and 25% documented cultural history. Search demand starts at 20% of the final score, leaving default final weights of 32%, 12%, 16% and 20% for those four components. Social visibility starts at 0%. The controls expose the final weights after each change; removing review scores redistributes the remaining base weights.
Search demand uses each restaurant’s highest supplied average monthly keyword value, not a sum of overlapping aliases. The signal is scaled as 100 × log(1 + searches) ÷ log(1 + the largest selected value). The supplied United States export covers September 2025–August 2026. We use its reported values as supplied; broad brand terms are not branch-level Atlanta search counts.
The social pilot counts matching, deduplicated units in small Google, YouTube, X and news samples. It has 64 of 76 searches observed and complete four-channel scores for 14 of 19 contenders. Blocked searches stay missing. When social weight is enabled, an incomplete contender retains its pre-social score for that component. This measures sampled visibility, not food quality, unique people or the preferences of offline Atlantans. Cultural history is a documented rubric; it is not a substitute for an in-person population survey.
ATL 10 shrinks each platform toward a four-star prior with 100 reviews of strength, caps observed review weight at 1,000 and maps one-to-five stars to 0–100. A missing platform uses that declared prior. The national model uses a separate 3.5-star prior and stricter branch eligibility. Those are distinct wing science choices, not interchangeable review datasets. Inspect Atlanta’s formulas and source ledger →
Wingonomics: one wing, several lenses
The Wing Index uses the arithmetic mean of all eligible menu unit prices in the active database edition. Each unit is a menu-listed wing piece, before tax, tip and fees. It is a sample benchmark, not a representative national average. Purchasing-power and per-capita comparisons use their disclosed state denominators; the census of 1,380 records does not create additional price observations.
Market Data divides each saved share price by that benchmark. Crispy Crypto does the same for the top 25 source-listed crypto assets by market capitalization, including stablecoins and tokenized assets. Dollar changes use the same fixed denominator; quote timestamps and retrieval dates remain visible. These are wing equivalents, not investment-value scores or predictions.
Wingflation covers 1980–2025 using annual-average U.S. CPI and a restaurant-price proxy anchored to the menu sample. It shows estimated historical wing prices, inflation’s cost in wings, and the change in a consumer basket after repricing both years in their own wings. Historical wing prices are model estimates, not observed annual wing menus. Donation tiers use the shared wing benchmark to express proposed dollar gifts; payments remain disabled.
Wingonomics price comparisons
Both reorder the national census, cheapest first, without a taste-score multiplier. Gross parity wing price = (menu order price ÷ wing count) ÷ (state regional price parity ÷ 100). BEA’s 2024 all-items index uses U.S. = 100. It is a broad state purchasing-power proxy, not a wing-specific price index or income measure.
Wing price per capita = (menu order price ÷ wing count) ÷ state population. We display the result multiplied by 1,000,000: dollars per wing per million residents. This does not estimate average customer spending or a population-weighted national average. Larger populations lower the index. Within one state, this preserves raw unit-price order.
Population is Census July 1, 2025 (Vintage 2025). Menu observations were retrieved September 15, 2026; these different vintages are disclosed, not inflation-adjusted. The Wing Index reports the current count of sourced menu observations and census coverage. Missing values remain unranked. Missing price, count, or state denominator means unranked. Exact metric ties share a rank.
Prices use regular bone-in orders closest to ten menu-listed pieces, before tax, tip, delivery fees or optional extras. Included dips and sides can differ; pieces are not standardized by meat weight. Brand and representative-branch observations are explicitly scoped. No bulk offers, combos, old menu guesses, or restaurant dollar-sign bands are converted into prices. These are saved menu observations, not guaranteed checkout prices. Additional sourced menu observations can expand coverage.