Alcoholic Mind Model

What AMM Is

A Simulation Framework, Not a Diagnostic Tool

AMM is not a diagnostic tool. It does not predict individual relapse, does not diagnose addiction, does not replace clinical assessment, and does not provide medical advice. It is exploratory behavioral modeling software intended for simulation, education, systems analysis, and research discussion.

AMM is an interpretable, equation-based computational model that simulates how emotional accumulation, rumination, mental pressure, recovery engagement, social support, and external life stability interact over time in the context of alcohol use disorder and recovery.

The model is grounded in established addiction research literature — including the work of George E. Vaillant and others — and is designed to produce structured, explainable outputs without machine-learning or black-box inference.

Not a cure. AMM does not claim to treat or cure alcoholism.
Not a predictor. AMM does not predict individual outcomes.
Not a black box. Outputs are produced by explicit equations and deterministic state transitions rather than machine-learning inference.
Exploratory. AMM is used to study pattern-level behavior across populations, not to evaluate any single person.

Architecture

Core Model Components

AMM models addiction and recovery through interacting subsystems, each governed by interpretable differential equations with fixed constants and no opaque machine-learning layers.

💧

Emotional Accumulation

Resentment, unresolved emotional load, and guilt build over time and decay slowly. This is the primary input to the rumination loop.

🔄

Rumination Feedback

Unresolved emotional load feeds into a self-reinforcing rumination cycle. Without intervention, this loop amplifies mental pressure.

⚡

Mental Pressure

The accumulated output of emotional load and rumination. High mental pressure drives relapse risk and behavioral-state transitions.

⚠️

Relapse / Consumption Risk

A modeled probability of consuming alcohol on any simulated day, driven by mental pressure, emotional load, and moderated by recovery engagement and external stability.

🌱

Recovery Growth

Recovery stability S(t), resilience buffer B(t), and recovery momentum G(t) grow through engagement, sponsor contact, meeting attendance, and support utilization.

🧠

Behavioral State Transitions

A six-state finite-state machine models discrete behavioral modes. Users transition probabilistically between states depending on internal and external conditions.

🤝

Sponsor / Support Utilization

Sponsor contact, meeting attendance, daily recovery actions, and amends work are modeled as recovery inputs that reduce risk and build stability.

🏡

External Life Stability

Purpose/work stability, family/home stability, and support network density are modeled as independent dimensions that moderate relapse risk and state transitions.

📊

Longitudinal Archetype Reconstruction

Vaillant-informed typologies (stable recovery, chronic relapse, late-onset, etc.) are reconstructed as event timelines and compared against simulated trajectories.

Variables

Core State Variables

Every variable in AMM is tracked continuously over simulated time so modeled behavior can be reviewed, tested, and explained.

ε

Resentment / unresolved emotional load

R(t)

Rumination — self-reinforcing cognition loop

M(t)

Mental pressure — accumulated drive state

C(t)

Modeled pressure index [0–1], not a probability

S(t)

Recovery stability — accrued sober foundation

B(t)

Resilience buffer — moderates emotional impact

G(t)

Recovery momentum — forward engagement trend

P(t)

Purpose / work stability [0–10]

F(t)

Family / home stability [0–10]

N(t)

Support network density [0–10]

ESI(t)

External Stabilization Index [0–100]

State Machine

Behavioral State Model

AMM models behavioral states as a six-state finite-state machine. Users transition probabilistically between states depending on internal pressure, recovery engagement, support utilization, and life stability. State transitions are not deterministic — identical conditions can yield different outcomes, which is consistent with real human behavior.

STABLE_RECOVERY

Sustained engagement in recovery. Low mental pressure, high recovery stability, active support utilization.

Simulated occupancy: 21.2%

DRIFTING

Low engagement, no active use, but little active recovery work. Modeled pressure accumulates gradually.

Simulated occupancy: 17.6%

HIGH_RISK

Elevated mental pressure with active contributing signals. A prompt to reconnect with support, not a forecast of use.

Simulated occupancy: 28.4%

ACTIVE_USE

Actively consuming. The model does not moralize this state — it represents the biological reality of resumed use.

Simulated occupancy: 19.7%

SHAME_COLLAPSE

Post-use shame spiral. Emotional load spikes, isolation increases, recovery engagement drops sharply.

Simulated occupancy: 6.4%

REENGAGEMENT

Re-engagement with recovery after a lapse. Fragile but oriented toward stability.

Simulated occupancy: 6.7%

Simulation

Simulation Results

AMM has been tested with 200 synthetic users across 730 simulated days in multiple experimental runs. These are simulation results, not clinical data. They demonstrate model behavior, not real-world outcomes.

200

Synthetic users

730

Simulated days

69.5%

Relapse rate

Never relapsed

100+ relapses

253d

Avg longest streak

Engagement & Relapse Outcomes

High engagement users showed materially lower relapse pressure and better stabilization than low engagement users in synthetic 200-user simulation runs:

High engagement (47 users): 31.9% relapse rate, avg C=0.32, 68% sustained sobriety
Medium engagement (100 users): 73% relapse rate, avg C=0.72, 27% sustained sobriety
Low engagement (53 users): 96.2% relapse rate, avg C=0.95, 3.8% sustained sobriety

Simulation result, not clinical prediction.

State Occupancy Distribution

Across all simulated users and all days, the aggregate state occupancy reflects the difficulty of sustained recovery without active engagement:

HIGH_RISK: 31.5% — the modal state, reflecting chronic pressure
STABLE_RECOVERY: 33.2% — sustained by active engagement
DRIFTING: 20.9% — disengaged, not using, not recovering
ACTIVE_USE: 9.6% — significant time spent in consumption
REENGAGEMENT: 4.3% — fragile return to recovery
SHAME_COLLAPSE: 0.6% — post-use emotional crisis

Key Findings

Engagement is the dominant differentiator in the synthetic simulation. High-engagement synthetic users show 68% sustained sobriety vs 3.8% for low engagement. Average C (modeled pressure) is 0.32 for high vs 0.95 for low.
Sponsor contact correlates with better simulated outcomes. Synthetic users with regular sponsor contact showed 51.1% relapse vs 90% for no sponsor.
Mental pressure separates early in simulation. M(t) trajectories diverge within the first 30 simulated days across the eventual synthetic outcome groups.
Simulated relapse events are probabilistic, not deterministic. Even highly stable synthetic users retain a small configured event probability.
External stability matters. Adding purpose/work, family/home, and support network dimensions measurably shifted population-level outcomes (avg ESI 65.8, final 72.9).
Vaillant-informed fixture agreement: AMM has 78% retrospective internal agreement across six composite fixtures without retuning core constants. This is not predictive accuracy.
Pre-event temporal audit: Across 3,228 correlated monthly windows from the 11 authored fixtures, AMM showed 81.8% sensitivity but only 37.3% specificity and 60.5% ranking AUC for a relapse event beginning within 30 days. This supports a sustained-pressure interpretation, not event prediction.

Exploratory behavioral model comparison — These are simulation results, not clinical predictions.

External Factors

External Life Stability Layer

Addiction does not occur inside a vacuum. Job stability, meaningful work, housing, safe relationships, and actual support utilization affect relapse and recovery trajectories. AMM models these dimensions explicitly.

P(t) — Purpose / Work Stability

Employment stability, daily structure, sense of meaning. Decays toward baseline when disrupted; grows slowly when stable.

F(t) — Family / Home Stability

Housing security, relationship safety, family support. Negative events (eviction, family conflict) produce sharp drops.

N(t) — Support Network Density

Quality and density of sober social connections. Stronger networks buffer relapse risk and accelerate state recovery transitions.

ESI(t) — External Stabilization Index

Weighted composite [0–100] of P, F, and N. Higher values correlate with lower relapse probability and faster recovery transitions in simulation.

Simulation Impact

Seed-controlled comparison (same 200 users, same 730 days, same RNG seed, external stability layer toggled):

Relapse rate: ~92% without external stability → 68.5% with
Average relapses per user: 97 → 63
Average longest sober streak: 130 days → 266 days

These are simulation results, not clinical predictions. The magnitude of the effect depends on model parameters and synthetic user distributions. The direction of the effect is consistent with existing research on social determinants of recovery.

Internal Checks

Longitudinal Reconstruction Against Vaillant-Informed Typologies

AMM was regression-tested against hand-authored, research-informed composite fixtures derived from George E. Vaillant's longitudinal alcoholism outcome studies (1983, 1995) and extended with findings from other recovery literature. These are not individual records or a held-out validation cohort.

87%

Internal agreement: 26/30 fixture checks

78%

Internal agreement on 6 expanded fixtures

100%

Agreement on one stable-recovery fixture

Fixture Behavior

Chronic relapse-loop behavior reproduced across multiple simulation runs
Long-term stable-abstinence trajectories align with the authored fixture expectations
Chronic HIGH_RISK edge-dweller patterns remain distinguishable in the reconstruction
Sociopathic rapid-onset escalation shows characteristic arrest-before-treatment timing
Recovery-engagement gradients differentiate survivors from relapsers
Mortality-state differentiation aligns with the outcome-informed fixtures
Support-utilization importance emerges endogenously
No core model constants were retuned for the comparison

Divergences & Limitations

Controlled-drinking, late-onset, and chronic stable-abuser profiles converge too much into HIGH_RISK — the model lacks differentiation mechanisms for these pathways
Longest sober streaks may be overestimated because documented real-world relapse timelines are incomplete or underreported
Late-onset full-arc C(t) averages can obscure end-state improvement — survivors who stabilize late show misleading mean values
The model is exploratory and is not a clinical validation study
The 87% and 78% values are retrospective fixture-agreement scores, not predictive accuracy; relapse and sober-interval checks are directly derived from fixture inputs
A stricter pre-event audit found high sensitivity but low specificity for 30-day relapse onset; elevated C(t) often persists through chronic-pressure periods without a new relapse event

Integration

Stillwater Integration

Infrastructure, not clinical validation. The Stillwater integration connects AMM 1.0.0 to the beta application as a server-side analysis pipeline. This is software infrastructure work — it does not constitute clinical validation, individual risk prediction, or a treatment recommendation system.

AMM 1.0.0 is integrated into the Stillwater beta through an authenticated Supabase Edge Function. It processes normalized daily check-in inputs with the same deterministic dynamics as the independent /amm package and returns modeled-pressure output, contributing signals, and suggested protective actions.

📡

Server-Side Analysis Pipeline

The amm-analyze Edge Function accepts normalized check-in values, prior AMM state, journal text for transient deterministic scanning, and days-sober from authenticated check-ins. Analysis runs server-side; derived inputs and modeled-pressure outputs are written through the service role with no direct client write access.

🔒

Authenticated Tier Gating

AMM analysis is gated by authenticated AMM-tier access. The app checks subscription_tier and subscription_status via the user_has_amm_access() PostgreSQL helper. Free-tier users are blocked before any database writes (403). RLS policies enforce read-own-only on output tables.

📋

Model Versioning & Restore Points

Every risk output is tied to model_version_id (FK to amm_model_versions), enabling full auditability of which model version produced which result. The version table stores parameter snapshots as JSONB, deployment timestamps, and rollback references — supporting future model upgrades without breaking historical lineage.

🧪

Synthetic Simulation Tooling

A simulation framework generates synthetic users across 8 recovery archetypes (stable recovery, pink cloud newcomer, chronic relapser, dry drunk, sponsor-supported, isolated user, grief trigger, high-functioning hidden). Runs of 200 users × 90 days produce 18,000 risk observations per run with metrics, archetype validation, and anomaly detection.

Simulation Visualization

The synthetic simulation pipeline generates six chart types — average risk by archetype, risk over time, risk state distribution, volatility, max risk, and individual sample trajectories — as both standalone SVG files and as a self-contained offline HTML dashboard with 7 interactive Chart.js charts.

Recovery archetypes

200

Synthetic users per run

Simulated days per run

18K

Risk observations

Chart types (SVG)

Dashboard charts

Risk Timeline

Multi-line chart showing average risk score per day across all 8 archetypes. Reveals pink cloud late-life escalation, grief-trigger midpoint spikes, and isolated-user upward drift over 90-day windows.

Average Risk by Archetype

Bar chart comparing mean risk scores. Chronic relapser (μ=0.44) and isolated user (μ=0.48) consistently highest; stable recovery (μ=0.09) lowest.

Risk State Distribution

Stacked bar showing days classified as stable, moderate, elevated, or high per archetype. Chronic relapsers spend 91% of days in elevated+ states; stable recovery users spend 100% in stable.

Sample Trajectories

Individual user trajectories (3 per archetype) plotted across full simulation window. Shows per-user volatility not visible in archetype averages — chronic relapsers oscillate rapidly while stable recovery users maintain flat low-risk profiles.

Infrastructure work, not clinical validation. These tools test model behavior and pipeline correctness. They do not evaluate, predict, or treat any individual's recovery outcome.

Visualizations

Simulation Visualizations

Below are selected simulation outputs from the AMM engine. Each chart is generated from synthetic data using internal simulation tooling.

State analysis overview — aggregate behavioral state distribution across 200 simulated users

State transition analysis — flow patterns between behavioral states over 730 days

State occupancy distribution — Sankey diagram of 50k+ state transitions

State transition heatmap — pairwise transition probability matrix

Survivor vs relapser trajectory comparison — C(t), M(t), and engagement

Recovery growth curves — S(t), B(t), G(t) for survivors and relapsers

External stability impact — P/F/N/ESI effects on relapse probability

Behavioral bifurcation — early M(t) separation across synthetic outcome groups

Vaillant-informed archetype comparison — 87% retrospective internal fixture agreement (26/30 checks)

Archetype trajectory comparison — individual Vaillant typology profiles

Engagement vs relapse rate — high engagement users show 31.9% relapse vs 96.2% for low engagement

Engagement vs recovery outcome — survivor, collapse-loop, and chronic HIGH_RISK classification by engagement level

Vaillant typology reconstruction — 78% retrospective internal fixture agreement across six composites

AMM mortality & outcome comparison — Vaillant typology death ages and AMM convergence scores

Boundaries

Ethical & Scientific Boundaries

AMM is exploratory behavioral modeling software. It does not diagnose addiction, predict individual relapse, replace treatment, or provide medical advice. It is intended for simulation, education, systems analysis, and research discussion. No clinical decisions should be made based on AMM output.

This project is guided by the following principles:

Interpretability. Outputs are produced by explicit equations and deterministic transition rules. There are no neural networks, hidden layers, or black-box components.
No clinical claims. AMM does not predict whether any individual will relapse, recover, or follow a specific trajectory. It is a simulation and personal-reflection framework.
Research-informed, not diagnostic. The model is grounded in longitudinal addiction literature but is not a validated clinical instrument.
No machine learning. No clustering, no classification, no prediction models. AMM uses hand-authored differential equations with fixed constants.
No clinical cohort claims. Reported results use synthetic simulations or authored composite fixtures, not a real clinical population.

Computational Behavioral Modeling of Alcoholism & Recovery Trajectories