Mortality Roulette (Finnish: Kuolleisuusruletti) is an educational and entertainment statistical simulation of human mortality, built around real population-level life-table and cause-of-death data.
It rolls a simulated life year by year using age-, sex-, country- and model-specific mortality probabilities and, when death occurs, can continue into cause of death, detailed cause trees and seasonal timing.
⚠ This is a statistical concept project, not an individualized medical prognosis or healthcare decision tool. This project presents mortality probabilities and causes of death directly and may be unsettling to some readers.
Mortality Roulette is an educational and entertainment statistical simulation and concept project. It works with population-level mortality data; it is not a medical risk calculator or an individualized prognosis tool.
The software and bundled datasets are provided AS IS. Output is not medical advice, diagnosis, prognosis, treatment guidance, or a basis for healthcare decisions. Do not use it as a substitute for professional medical assessment.
Just because you can fly an airliner in Microsoft Flight Simulator does not make you a pilot. Likewise, running Mortality Roulette does not make you a physician, epidemiologist, actuary, or fortune-teller.
Requirements:
- Python 3.10+
- no third-party Python packages are required for the main program
- internet access is optional for bundled default datasets; refresh/download paths and non-bundled data can still use the network
Clone the repository:
git clone https://github.com/FlyingFathead/mortality-roulette.git
cd mortality-rouletteJust run it:
python mortality_roulette.pyIn an interactive terminal, Mortality Roulette asks for the basic choices that were not supplied on the command line, including sex and the present-day mortality model.
Choose sex: (m)ale, (f)emale, (r)andom:
Choose mortality model:
(s) age-graduated official baseline [recommended for simulation]
(o) official raw period table [literal published single-age qx]
(l) original legacy Mortality Roulette
Selection [s]:
The default/recommended smoothed model is still derived from the bundled official period table. It uses a transparent age-domain graduation to reduce single-calendar-year sawtooth noise; it does not rewrite or replace the raw official dataset.
For added morbid education about how sustained lifestyle hazards can modify an already-unfriendly mortality baseline, load up Finland and Canada with the Boozehound/Wino scenario and let the cause-hazard model do its thing.
Quick fun run example:
python mortality_roulette.py \
--deathmatch fi ca \
--sex m \
--mortality-model official \
--boozehound-wino \
--alcohol-model cause-hazard-prototype \
--cause-hazard-weight-model evidence-v4-cancerOr, less academically: load up the contestants with booze and see what the statistics do to them.
A plain baseline Deathmatch, with no alcohol exposure, is simply:
python mortality_roulette.py --deathmatch fi ca --sex mPrint annual mortality odds without rolling a life:
python mortality_roulette.py --printout --gender mChoose a mortality model explicitly:
python mortality_roulette.py --mortality-model smoothed
python mortality_roulette.py --mortality-model official
python mortality_roulette.py --mortality-model legacy--legacy-mortality remains available as a backward-compatible alias for --mortality-model legacy.
For all options:
python mortality_roulette.py --helpsmoothed — age-graduated official baseline (recommended for simulation)
- starts from the bundled official single-age period-table
qx - converts
qxto annual hazards - leaves age 0 untouched
- applies a centered five-age triangular smoother with weights
1,2,3,2,1 - from exact age 30 onward, applies nondecreasing isotonic graduation (PAVA)
- keeps the official source data untouched and separately selectable
This is intended to expose the underlying age-loading of the mortality die without letting random single-calendar-year bumps dominate adjacent ages.
official — raw published period table
Uses the literal published single-age qx values. Annual sampling noise is preserved, so adjacent ages can occasionally move downward even though the overall age trend rises strongly.
For Finland the bundled official 2024 StatFin table supplies one-year qx through exact age 99. The age-100 row is terminal/open-ended and has no published 100→101 qx; the program therefore labels age 100+ as an explicit tail model. StatFin reports remaining life expectancy at age 100 separately (1.85 years for males and 1.80 for females).
For Canada the bundled 2024 Statistics Canada complete life table supplies exact-age qx through age 109. Any modeled tail starts only after the last available exact-age value.
legacy — original legacy Mortality Roulette
The original baked mortality schedule from early versions is retained in the code for historical comparison and exact reproducibility. It is not represented as the current official StatFin table and is not silently substituted for the new age-graduated model.
Present-day model selection does not remove any existing data plumbing: bundled datasets are preferred for normal offline use, while explicit cache paths and --refresh-statfin, --refresh-statcan and the other refresh/download options remain available.
--birth-year YEAR can follow the mortality conditions that applied as a simulated person aged through calendar time rather than applying one modern period table to the whole life. For a person born in 1947, age 0 uses 1947 qx, age 1 uses 1948 qx, and so on.
Mortality Roulette can read Human Mortality Database (HMD) 1x1 period life tables directly from locally downloaded country ZIPs. HMD is optional: the normal present-day game does not need it. For national historical birth-cohort runs, an installed HMD archive is preferred automatically for its long historical coverage. If a bundled canonical national source contains a newer observed year than the installed HMD archive, Mortality Roulette appends that newer observed year rather than discarding it or silently treating an older HMD year as current. If HMD is not installed, Canada falls back to its bundled Statistics Canada history and Finland can use the open Statistics Finland 12ap history (cached/downloaded on demand for pre-2024 cohort years). Province-specific Canadian runs continue to use province-specific Statistics Canada mortality rather than substituting national HMD data.
After registering/logging in at the Human Mortality Database, download the desired country archive from the relevant HMD country page:
Place the downloaded archive under the Git-ignored local data directory:
local-data/hmd/FIN.zip
local-data/hmd/CAN.zip
local-data/hmd/USA.zip # parser-ready for future U.S. geography support
The program reads only the HMD-created STATS/*ltper_1x1.txt period life tables required for qx. It does not consume HMD InputDB material. Complete HMD country downloads are deliberately not bundled in Mortality Roulette releases; users obtain their own current copy from HMD. local-data/ is optional and is not created merely by launching the base program. When HMD is actually used, the historical run/printout identifies the HMD country source page; there is no HMD nag in ordinary present-day startup.
--hmd-dir PATH can override the default location and accepts a directory containing country ZIPs, an extracted HMD country tree, or a direct country ZIP path for a single-country run.
Deathmatch is population-baseline/no-alcohol by default. Add --boozehound or --boozehound-wino explicitly to apply alcohol exposure.
--country fi
--country ca
--canada # shorthand for --country caUse --ca-province to select a Canadian province for mortality and monthly timing data:
python mortality_roulette.py --country ca --ca-province bc --sex m
python mortality_roulette.py --country ca --ca-province ontario --sex fSupported single-year complete-life-table provinces:
nl Newfoundland and Labrador
ns Nova Scotia
nb New Brunswick
qc Quebec
on Ontario
mb Manitoba
sk Saskatchewan
ab Alberta
bc British Columbia
Prince Edward Island is deliberately rejected for now because Statistics Canada does not publish it in the same single-year complete life-table series used by the exact-age engine. The territories likewise use a different abridged life-table product.
Canadian provincial mode currently uses:
- province-specific annual
qxfrom Statistics Canada 13-10-0837-01 - province-specific monthly timing from Statistics Canada 13-10-0708-01
- national Canadian WHO cause-of-death distributions, clearly labelled as national when a province is active
The preferred scalable contestant syntax is repeatable --player. Each player carries its own country, optional Canadian province, sex, and optional birth year:
python mortality_roulette.py --player ca:on:m --player fi:fCompact player format:
fi:m Finland, male
fi:f Finland, female
fi:r Finland, random sex
fi:m:1947 Finland, male, born 1947
ca:m national Canada, male
ca:m:1980 national Canada, male, born 1980
ca:on:m Ontario, male
ca:on:f:1962 Ontario, female, born 1962
ca:bc:f British Columbia, female
--player must currently be supplied exactly twice. :r is resolved independently for each player using its own deterministic RNG stream when --seed is supplied. Province and sex are part of the player spec, so --ca-province and --sex/--gender are intentionally rejected when --player is used. This compact contestant object is the foundation for adding further per-player exposures later without proliferating parallel --foo-1 / --foo-2 switches.
Birth years can be embedded per player or supplied as a match-level convenience override:
# Per-player years
python mortality_roulette.py --player fi:m:1979 --player fi:f:1985 --mortality-model official
# Override both embedded/player years left-to-right
python mortality_roulette.py --player fi:m --player fi:f --birth-years 1979 1985 --mortality-model official
# One value applies to both contestants
python mortality_roulette.py --deathmatch fi ca --sex m --birth-years 1947 --mortality-model officialWhen both birth years are known, Deathmatch defaults to a shared calendar timeline. The earlier-born contestant starts first; the later-born column displays WAITING TO BE BORN... until its birth year arrives, after which both lives advance through the same calendar years. In this mode the long winner is the contestant who dies in the later calendar year (the last one alive), not necessarily the contestant with the greater lifespan in years.
Use --deathmatch-timeline independent for the alternative age-for-age comparison. Both players then start at age 0 immediately, but each column advances through its own calendar years (for example age 40 may mean 2019 for a 1979-born player and 2025 for a 1985-born player). --deathmatch-timeline calendar forces the shared-world form and requires birth years for both contestants. With no birth years, Deathmatch behaves exactly as the ordinary present-day mode always has.
Historical layers are calendar-gated independently. Annual mortality uses the player's year-specific cohort source. Cause/detail/seasonality data are used only when the death calendar year falls within that source's historical coverage; pre-coverage deaths remain explicitly N/A/unavailable rather than borrowing later distributions. Years after the newest observation retain the existing explicitly labelled future-hold behavior. This prevents anachronisms such as assigning a modern cause distribution to a death that occurred before that cause/data classification existed.
Deathmatch presentation identifies the two sides consistently before their geography, for example PLAYER 1: 🇨🇦 CANADA (ONTARIO) and PLAYER 2: 🇫🇮 FINLAND. The final two-column card now uses explicit semantic subsections when applicable: 🚗 CRASH CONTEXT for downstream traffic impairment/intoxication context, 💊 SUBSTANCES for poisoning-agent context, and 🍸 ALCOHOL for the configured lifetime exposure. Cumulative beverage conversions are labeled 🍷 WINE EQUIV. and 🥃 VODKA EQUIV. to make clear that they are descriptive equivalents of the same ethanol total, not additional consumption.
The older --deathmatch interface remains backward-compatible and keeps its historical shared-sex semantics while using the same PLAYER 1 / PLAYER 2 presentation.
Two countries with the legacy interface:
python mortality_roulette.py --deathmatch fi ca --sex mSame country, two independent players:
python mortality_roulette.py --deathmatch fi --sex mCanadian provincial Deathmatch:
python mortality_roulette.py --deathmatch ca ca --ca-province bc on --sex mWin condition:
--deathmatch-win long # longevity; last contestant standing wins (default)
--deathmatch-win short # brevity; first contestant to tap out winsDeathmatch players always receive independent mortality, cause, detail and timing RNG streams. --seed N still makes the entire match reproducible.
Normal single-run roulette defaults to the full presentation stack:
cause-of-death roulette: ON
cause detail: TREE
seasonal death timing: ON
Controls remain available:
--causes / --no-causes
--cause-detail broad
--cause-detail specific
--cause-detail tree
--seasonality / --death-month
--no-seasonality
--no-who-detailBatch mode remains lean by default (broad/optional cause work and no monthly timing unless explicitly requested), and --printout remains qx-only. Finland uses Statistics Finland cause data and Canadian cause selection uses Canada's WHO Mortality Database submission. Seasonal timing is rolled only after death and does not alter annual mortality or cause selection.
Finnish detailed-cause data use the vendored datasets/finland/causes/statfin_cause_detail_2024.json as an immutable read-only seed. If a legitimate StatFin detail cell is missing and must be fetched lazily, the new cell is written to ~/.cache/mortality_roulette/statfin_finland_cause_detail.json (or an explicit --detail-cache) rather than modifying the tracked bundled dataset.
v0.13.8 adds a sparse CAUSE NOTE annotation layer backed by datasets/cause_notes/cause_note_model_v1.json. Notes are deterministic explanatory metadata applied only after an underlying cause has already been resolved; they do not consume RNG or alter mortality, cause, detail, PLACE, seasonality, substance or crash-context probabilities. The initial rule covers Finnish M17 gonarthrosis from StatFin 11be, where the public table records the underlying cause but not the complete death-certificate chain of immediate/intervening causes. The table is intentionally extensible so similarly non-obvious underlying-cause results can receive precise notes later without hard-coding prose into the roulette engine.
If the resolved underlying cause is suicide (X60-X84 / Y87.0), the simulator adds one further STATISTICAL REASON roll. Finland and Canada use separate sex/age evidence models; unsupported future countries fall back to an explicitly labelled Finnish-Canadian reference distribution. This is a probabilistic context model, not an assertion of an individual's proven motive.
v0.13.2 added two narrow conditional detail rolls where published evidence supports them: X80 (intentional self-harm by jumping from a high place) receives a broad statistically weighted site type, and X41 accidental psychotropic/antiepileptic poisoning can receive a broad evidence-weighted drug class. In v0.13.7 poisoning output is normalized under a dedicated 💊 SUBSTANCES section: X40, X42, and X43 expose the broad agent category already encoded by ICD; X41 retains its existing evidence-weighted class roll; and X44 explicitly represents other/unspecified or multidrug poisoning context instead of leaving the agent field blank. No doses or molecule-level lethality ranking are modeled.
v0.13.7 also adds an independent 🚗 CRASH CONTEXT roll after a resolved road-traffic death. The model deliberately distinguishes who the statistic describes. For Finland, OTI 2015–2024 investigation-board data provide deceased-person intoxication distributions for pedestrians and cyclists; motorcyclist and other motor-vehicle deaths use a crash-level at-fault-driver impairment reference and explicitly state that this does not establish that the simulated decedent was the impaired driver. For Canada, Transport Canada's 2022 fatal-collision contributing-factor statistic is used as a broad crash-level reference and likewise does not assign impairment to the decedent. Unknown/unavailable Finnish statuses are retained as explicit outcomes rather than silently treated as sober. Statistics Finland table 11b2 confirms that death-certificate alcohol/drug intoxication is available by external cause, age, sex and year; those exact cross-tabs are a future refinement and are not reverse-engineered from marginal age/sex percentages here. The traffic roll has its own RNG stream and cannot change mortality, cause, detail, SUBSTANCES, PLACE, or seasonal outcomes.
v0.13.8 tightens that boundary for railway and explicitly nontraffic transport deaths. A realized railway-train/vehicle detail is no longer fed into the generic road-setting distribution: ICD wording resolves Railway tracks / premises for nontraffic events, Railway crossing / public road for traffic events, and Railway tracks / crossing when traffic status is unspecified. Railway collisions and exact nontraffic V-codes also skip the generic road/fatal-motor-vehicle impairment roll because the bundled impairment sources do not provide a defensible rail-event denominator.
v0.13.3 generalizes the existing X80 location machinery into a public PLACE layer without replacing the older evidence. PLACE is a separate downstream RNG stream and is emitted only when a matching country/cause model exists. Bundled coverage now includes Finnish and Canadian drowning settings, Finnish sex-specific homicide scenes, Finnish and Canadian fatal-road settings, Finnish and Canadian cancer place of death, and Finnish neurodegenerative place of death. Explicit ICD environment information constrains the roll first: for example, bathtub/pool drowning codes resolve that setting directly, while natural-water codes restrict the subsequent statistical PLACE roll to compatible water categories. Existing X80 probabilities and its independent RNG stream are preserved exactly and now render through the same PLACE presentation. Unsupported cause/country combinations remain blank rather than receiving invented scenery. See datasets/README.md for exact source populations, derived residuals, caveats and provenance.
v0.13.5 extends PLACE to selected fatal poisoning and fire contexts. Finnish drug-poisoning PLACE is age-limited to 15–29 because the bundled THL forensic study covers under-30 deaths; Canada uses sex/life-stage national accidental acute-toxicity event-place data. Fire PLACE is likewise source-specific: Finnish building-fire detail is used only when ICD has already resolved a building/structure fire (X00/X01), while Canada has a broader national unintentional-fire residence/property distribution. No missing or suppressed category is silently reconstructed. In v0.13.7 compact output separates 🚗 CRASH CONTEXT, 💊 SUBSTANCES, 📍 PLACE, and 🍸 ALCOHOL into distinct semantic blocks.
v0.13.10 added an opt-in downstream modeled postmortem layer. In v0.13.11, Deathmatch moved those boxes below the final comparison table; v0.13.12 tightened relevance. v0.13.13 makes the doctor factual-first: every supported country first receives a narrative assembled from already-realized simulation fields (age/sex, cause/detail, PLACE, crash/substance context and timing), then exact ICD knowledge and country-specific evidence are added where applicable. A missing specialist pathway no longer erases information the simulator already rolled. The postmortem layer never changes mortality, cause/detail selection, PLACE, timing, or any upstream RNG result.
# Generic switch; currently the evidence-backed pathway model is Canadian
python mortality_roulette.py --country ca --sex m --postmortem
# Branded Canada-specific showcase preset
python mortality_roulette.py --deathmachine-ca --sex m
# Works with Canadian Deathmatch contestants too
python mortality_roulette.py --deathmachine-ca \
--player ca:on:m --player ca:bc:fThe presentation uses a randomly selected doctor emoji (👩⚕️, 👨⚕️, or 🧑⚕️) from its own independent downstream RNG stream. The doctor is a renderer for the model, not an assertion that an individual medical record was observed. Realized simulation facts are narrated first; evidence-backed pathway material remains explicitly labeled as modeled population context. If a Canadian pathway rule has no match, the same factual doctor summary remains and receives a short No additional Canada-specific pathway evidence... footer. Finland uses the same factual composer; the politician/💸 gag is now a true last resort only when the realized stack itself cannot yield a meaningful summary.
The initial Canadian public-data model covers selected cause families where Statistics Canada exposes useful aggregate structure: snowmobile and ATV fatalities, motorcycle and passenger-vehicle fatal-event classes, pedestrian fatality context, unintentional fire context, fatal falls, Alzheimer's multiple-cause associations, and the multiple-cause burden recorded with diabetes deaths. These source products do not all provide the same kind of evidence. The model preserves that distinction: event-type marginals stay event-type marginals, multiple-cause co-recording is not promoted to temporal causation, and qualitative evidence summaries are not converted into fake probabilities.
--deathmachine-ca is a Canada-only showcase alias. If no country/player mode is otherwise supplied it selects national Canada and enables the postmortem layer. Use the generic --postmortem switch for mixed-country Deathmatch. v0.13.13 also adds a code-indexed ICD annotation layer (datasets/who/icd10/icd10_annotations_v1.json) over the bundled WHO title catalog. Runtime genre/attribution decisions are made from normalized ICD codes and explicit parent inheritance, never title keywords. The first large annotation block covers all 872 bundled V01–V99 transport entries, so an exact transport code can contribute victim/vehicle role, collision counterpart and traffic/nontraffic status without inventing details; broad ranges such as V40–V79 remain broad. The same table already tags direct-alcohol families such as K70.* and G31.2, plus dementia/Alzheimer families. Optional scope metadata can restrict future annotations/enrichments by country/region/sex/age without suppressing the universal ICD meaning.
Finnish Alzheimer Europe prevalence context is code-gated to realized dementia/Alzheimer ICD families; old age alone does not inject dementia into unrelated deaths such as stroke. If the realized code itself establishes dementia (F00/F01/F02/F03/G30), the doctor prints DEMENTIA: YES; Alzheimer-specific F00/G30 also prints ALZHEIMER: YES. Population prevalence remains a separately labelled benchmark, not a diagnosis probability. For Finnish players age 85+, Vantaa 85+ is shown separately as an observed oldest-old cohort benchmark (214/553, 38.7% dementia at baseline) and is never averaged with Alzheimer Europe. Proxy-only boozehound weights such as the current I63 fallback no longer create an alcohol-themed postmortem; direct alcohol-coded outcomes (for example K70.1 or G31.2) and genuinely evidence-backed alcohol enrichments can still add alcohol context. The postmortem layer is per-life presentation and is intentionally unavailable in --runs batch mode and qx-only --printout mode.
The bundled datasets/pathways/ca_postmortem_pathway_model_v1.json contains only project-authored parameters derived from Statistics Canada products covered by the Statistics Canada Open Licence. datasets/pathways/fi_postmortem_context_v1.json is a separate background-context model; its Alzheimer Europe rule stores 75+ age/sex dementia prevalence but is trigger-gated to dementia/Alzheimer-relevant realized cause stacks and never converts prevalence into a new diagnosis. Alcohol-model context is derived at runtime from the already-selected boozehound cause/detail fields and the active alcohol weight model; it adds no new mortality/cause roll. CIHI workbooks may be used locally as research/reference material under local-data/cihi/, but they are not runtime dependencies, are not committed, and are not redistributed in releases. See datasets/README.md for source-by-source provenance and limitations.
--boozehound
--boozehound-wino
--alcohol-start-age AGE # default: 18
--alcohol-end-age AGE # optional; stop age itself is alcohol-free
--alcohol-model legacy|cause-hazard-prototype
--cause-hazard-weight-model proxy-v1|evidence-v1|evidence-v2-popnorm|evidence-v3-popdist|evidence-v4-cancer--boozehound-wino models one 750 mL bottle of 12% ABV wine per day, including hazard-based all-cause mortality adjustment, duration-aware cause weighting, cumulative survival accounting and descriptive ethanol/wine/vodka equivalents. Exposure starts at age 18 by default; --alcohol-start-age changes the start age and --alcohol-end-age optionally ends modeled intake. Start/stop ages are printed in run headers and exposure summaries.
The current cessation implementation preserves alcohol exposure accumulated before the stop age and its already-compounded survival effect, but returns the current alcohol hazard multiplier and alcohol cause reweighting to baseline after cessation. A disease-specific post-cessation residual-risk decay curve is not yet calibrated, so the program states this limitation explicitly when relevant.
v0.13.4 clarifies that the printed g/day exposure is an ongoing modeled / annualized habitual exposure, not a literal day-by-day drinking schedule. It also fixes explanatory ICD context for Finnish StatFin alcohol rows such as F10: public StatFin mortality tables publish the underlying cause at 3-character level, so the simulator does not invent an F10.x probability roll, but it can display the legitimate WHO clinical-state children and Finnish ICD-10 withdrawal/delirium refinements (including seizure-status subcodes) as non-probabilistic taxonomy. Statistics Finland documents that its source register is coded at the most accurate ICD-10 level even though public underlying-cause statistics are published at 3 characters; exact F10.x mortality weights therefore remain a data-access task rather than a classification gap.
v0.13.0-dev1 added an opt-in --alcohol-model cause-hazard-prototype engine for present-day Finland. It splits each annual all-cause hazard using StatFin broad cause shares, applies the existing duration/dose-aware broad alcohol cause weights as provisional hazard multipliers, recombines them, and derives the annual death probability from that total. This is an architecture sensitivity prototype, not a validated gameplay model. The existing broad cause weights were originally conditional cause-shape proxies, and Finnish population background alcohol exposure is not yet deconvolved. legacy remains the default through the single top-level DEFAULT_ALCOHOL_MODEL constant.
The selected alcohol engine is printed in single-run, batch and Deathmatch headers whenever a boozehound preset is active. The prototype loads the active country's cause data internally even when visible cause reporting is off: StatFin 11az broad causes for Finland and WHO complete-ICD cause cells for Canada. This does not by itself enable visible cause-of-death roulette. Birth-cohort mode remains unsupported for the prototype.
v0.13.0-dev2 adds an A/B-selectable broad-cause batch sampler. --cause-batch-sampler fast-grouped is the default: it resolves each StatFin sex/age/year cause cell once and performs grouped draws, avoiding the previous per-death distribution rebuild. --cause-batch-sampler reference-slow preserves the original one-death-at-a-time implementation for validation. Fast grouped sampling currently applies only to broad causes without seasonality; specific/tree detail or seasonal timing falls back explicitly to the reference sampler. The top-level DEFAULT_CAUSE_BATCH_SAMPLER controls the default.
v0.13.0-dev3 introduced A/B cause-hazard weights. proxy-v1 exactly preserves the original prototype and evidence-v1 replaces only the directly alcohol-coded hazard with the Carr et al. 2024 AUD-mortality dose-response. v0.13.0-dev4 added evidence-v2-popnorm, which divides that direct-alcohol RR by a provisional population mean-dose RR because the observed period-table hazard already embeds background drinking. v0.13.0-dev8 adds evidence-v3-popdist: instead of pretending the population mean dose is a representative person, it estimates population E[RR(D)] with the WHO/Rehm/Kehoe Gamma exposure model for current drinkers plus an abstainer point mass. All heterogeneous non-direct-alcohol broad groups remain explicit proxy fallbacks. None of the evidence models is the global default yet.
v0.13.0-dev9 adds evidence-v4-cancer. It preserves v3's population-normalized Carr direct-alcohol hazard and adds disease-specific cancer subhazards from Dai et al., Nature Health (2026), Table 3. The mapped ICD-10 partitions are C00-C08 (lip/oral cavity), C09-C10 and C12-C14 (other pharynx; C11 nasopharynx excluded), C15 (oesophagus), C16 (stomach), C18-C21 (colorectal), C22 (liver), C25 (pancreas), C32 (larynx), C50 (female breast) and C61 (prostate). The published mean-RR table is used at 10 g/day intervals through 100 g/day and held flat above 100 g/day rather than extrapolated beyond published support. Each site RR is normalized by the same country/sex WHO-style Gamma E[RR(D)] exposure distribution used by v3.
For Finland, DEV9 also removes the old flat neoplasm broad-hazard shortcut when v4 is active. The annual StatFin 04-22 neoplasm hazard is rebuilt from the exact same 11be C00-D48 child distribution used by specific cause roulette. Suppressed/missing child mass is preserved as an explicit unresolved residual at multiplier 1.0, and an over-inclusive child partition is rejected. This is a hard consistency invariant: the broad annual cancer hazard and the visible detailed cancer roulette cannot silently use different cancer mass. Non-cancer/non-direct alcohol-sensitive mappings remain explicit proxy-v1 fallbacks.
For evidence-v3-popdist, the current-drinker distribution is Gamma with sex-specific SD/mean ratios (1.171 men, 1.258 women), uses 80% of APC in accordance with WHO burden-of-disease exposure methodology, and is normalized on 0–150 g/day. Carr reports the AUD-mortality curve through 100 g/day; inside the population denominator only, dev8 holds the Carr RR flat above 100 g/day rather than inventing an exponential continuation through the 100–150 g/day tail. Finland uses OECD 2025 sex-specific APC plus THL 2023 abstainer shares (20–69: men 10%, women 12%). Canada currently uses a WHO GISAH 2020 sex-neutral APC fallback plus Statistics Canada CCHS 2023 past-year nondrinker prevalence as an explicitly labeled sex-neutral fallback. These source/age/year mismatches are printed in diagnostics rather than hidden.
The alcohol calibration harness also reports an unfitted cause-hazard Wood check when an existing StatFin 11az cache is available. This is deliberately different from the Wood-smooth curve: the cause-hazard prototype is not fitted to Wood, so its lifespan-loss row is an out-of-sample architecture diagnostic.
--runs 100000
--batch-engine fast
--batch-engine step
--birth-year 1980 # single run; in Deathmatch one shared year for both
--birth-years 1979 1985 # Deathmatch per-player override; one value = both
--deathmatch-timeline auto|calendar|independent
--hmd-dir PATH # optional local HMD ZIP directory/tree/archiveBatch summaries include a fixed-bucket terminal death-age histogram by default, with counts and shares. Use --no-histogram to suppress it. The histogram is presentation-only and uses the already simulated death ages; all pre-existing batch summary figures and sections remain unchanged.
Other useful options:
-v, --version # print the central VERSION value and exit
--sex m|f|r, --gender m|f|r # aliases; same parser destination
--printout, --odds-table # deterministic qx/1-in-X table, no RNG
--end-age N # inclusive final interval for --printout
--seed N
--delay 0
--start-age N
--exceptional-tail # Finland: allow model-only survival beyond observed record ceilingRun python mortality_roulette.py --help for the full argument list.
v0.12.10 added a standalone validation utility without changing roulette mortality semantics.
v0.12.11 extended that same calibration-only harness with a smooth Wood-calibrated
candidate curve. v0.12.12 adds an independent high-dose holdout panel; the main
roulette still uses the unchanged legacy alcohol model:
python tools/alcohol_calibration.pyThe harness analytically integrates the existing Finnish period table and current legacy alcohol hazard model, then compares its age-40 life-expectancy differences with the Wood et al. (2018) current-drinker benchmark. It also prints a diagnostic single-RR value required to reproduce 4.0, 4.5 and 5.0 years of life-expectancy loss under the existing simplistic architecture. Those diagnostic values are explicitly not proposed alcohol relative risks.
The v0.12.11 candidate solves architecture-specific RR knots that reproduce the midpoint of Wood's reported lifespan-loss bands at the reported group mean intakes, then connects them with a dependency-free monotone cubic interpolation in log(RR). Above Wood's highest observed group mean (~52.4 g/day), the tool uses conservative log-linear extrapolation and labels those outputs LOW CONFIDENCE. The 71 g/day BOOZEHOUND-WINO comparison is therefore a sensitivity analysis, not a proposed epidemiological estimate.
v0.12.12 adds an independent holdout rather than another fit: published male all-cause RR points from Wang et al. (2014) at 50/75/90/100 g/day are printed beside the Wood-effective multiplier. Their reference populations are different, so the comparison is explicitly diagnostic rather than an apples-to-apples pass/fail test. Cirrhosis mortality dose-response estimates and hospitalized Nordic AUD mortality/life-expectancy ranges are shown only as severity/context bounds. None of these values change gameplay.
The tool is intentionally calibration-only: it does not alter the main CLI, mortality probabilities, alcohol presets, cause selection, RNG streams or batch simulation.
The 0.12.x series begins a gradual modularization of the original single-file prototype.
The public entry point remains mortality_roulette.py; low-risk support code can move into
mortality_roulette_core/ without changing CLI usage.
Run the bundled dependency-free regression suite with:
python -m unittest discover -s tests -vThe first tests pin CLI version reporting, terminal rendering primitives, Canadian province assignment, hazard-space alcohol mortality adjustment, and reproducible/independent Deathmatch RNG streams.
Release history is maintained in CHANGELOG.md.
Smoking is planned as an independent first-class lifestyle exposure using the 🚬 indicator, given its substantial impact on long-term morbidity and mortality trajectories. Contestants should eventually support at least these states independently per player:
- neither exposure
🍷alcohol only🚬smoking only🍷 + 🚬alcohol + smoking- former smoker, with explicit start/stop ages and time since cessation
The smoking model should track intensity and accumulated exposure (for example cigarettes/day and pack-years), plus cessation-related risk decay where evidence supports it. Combined alcohol + smoking must use cause-specific interaction terms where epidemiological evidence supports synergy, especially upper aerodigestive cancers; the simulator must not blindly multiply one global alcohol RR by one global smoking RR. Single-exposure modes remain available for clean A/B comparisons.
Primary statistical/data sources:
- Statistics Finland — Finnish life tables, causes of death and seasonal timing. Bundled StatFin open-data snapshots are attributed to Statistics Finland and distributed under CC BY 4.0.
- Statistics Canada 13-10-0837-01 — complete single-year life tables for Canada and supported provinces.
- Statistics Canada 13-10-0708-01 — deaths by month and place of residence. Bundled Statistics Canada snapshots are redistributed under the Statistics Canada Open Licence with source attribution.
- WHO Mortality Database — Canadian civil-registration cause-of-death data and Finnish deep-detail support where available.
- WHO ICD-10 2019 terminology — code-title presentation metadata. WHO retains copyright/licensing control over ICD-10; this material is not represented as Statistics Finland/Statistics Canada open data.
- Public Health Agency of Canada — Canadian Chronic Disease Surveillance System (CCDSS) data tool — current Canadian surveillance source for dementia including Alzheimer disease; age-specific estimates are available nationally in five-year groups. CCDSS methods explain case definitions, prevalence measures, age grouping and data-quality limits. Canadian dementia benchmarks are kept geographically distinct from Alzheimer Europe estimates.
- Vantaa 85+ cohort — Hall et al. 2019 — Finnish population-based oldest-old cohort context; 214/553 clinically examined participants had dementia at baseline in 1991, with later diagnoses retained as longitudinal context rather than folded into the baseline prevalence.
- Human Mortality Database (HMD) — optional local historical 1x1 period-life-table input for national birth-cohort simulation. Mortality Roulette supports HMD country ZIP imports without bundling complete HMD downloads. Current country pages: Finland, Canada, and parser-ready United States. HMD-created estimates are CC BY 4.0; separately supplied input data retain their providers' distribution licences.
- Suicide statistical-reason evidence: Finnish nationwide psychological-autopsy studies (Heikkinen and colleagues) plus Canadian coroner/medical-examiner studies from Alberta and Montréal; exact references and model provenance are in
datasets/README.md. - Conditional external-cause context evidence: Toronto/Swiss/Taipei jumping-site studies for
X80; Finnish poisoning studies plus Public Health Agency of Canada coroner/medical-examiner toxicology data forX41; WHO ICD-10 multidrug-poisoning coding semantics plus PHAC accidental acute-toxicity substance-count data for the explicitly labelled CanadianX44context reference; OTI fatal-road investigation data for Finnish traffic intoxication/impairment context; and Transport Canada fatal-collision contributing-factor data for the Canadian traffic reference. Exact references and modeling limitations are indatasets/README.md.
See datasets/README.md and datasets/manifest.json for the bundled-file provenance map.
Alcohol-risk calibration and disease-specific scenario references are documented next to the relevant constants/functions in mortality_roulette.py.
No software license has been granted for Mortality Roulette at this time. If you are interested in the project, collaboration, or licensing, please contact FlyingFathead.
Third-party datasets remain subject to their respective source licences and terms. See datasets/README.md and datasets/manifest.json for details.
Project / code / design: FlyingFathead With special thanks to ChaosWhisperer for development and research assistance.
Mortality Roulette was inspired in part by the work of Sheldon Solomon and colleagues on terror management theory: the curious predicament of an animal intelligent enough to understand that it is going to die, and inventive enough to spend much of its life trying not to think about that.
Memento mori