01Verified
Batter probability weights
Six quantized weights are derived from AB + BB. Each rate is truncated separately; the six weights are not a normalized outcome distribution.
Read supporting analysisCOMPUTATIONAL BASEBALL RESEARCH / EST. 2026
The MicroLeague
Project.
Reverse Engineering Baseball Simulation
A computational investigation of probability, strategy, and decision-making in the 1984 MicroLeague Baseball engine.
Phase 1 activeResearch & reconstruction
Static arithmeticReviewed and consistency-tested
Conditional state modelsCoverage varies by starting state
Native validation pendingNo Atari execution or debugger trials
01 / RESEARCH PROGRAM
Recover the rules.
Expose the assumptions.
A simulation is a sequence of decisions. Our aim is to make each one inspectable.
- Recover and document the original simulation algorithms.
- Reconstruct the probability engine mathematically.
- Analyze offensive and defensive strategy.
- Validate predictions against original game behavior.
- Investigate improvements for a future simulation engine.
02 / RESEARCH FINDINGS
What the evidence supports
Local instruction-level findings are distinct from modeled probabilities and runtime proof.
Verified Reviewed instructions + relevant static checks; scoped, not native validation.
Modeled Mathematical reconstruction under stated assumptions.
Preliminary Strong evidence; incomplete confirmation.
Open Question Unresolved behavior requiring investigation.
02Verified
Pitcher adjustments
Hits, walks and strikeouts use stored whole innings pitched and two stages of fixed-point division. Original fractional-innings import rules remain unknown.
Read supporting analysis03Verified
Team-specific baselines
The batting team supplies the normalization header. Of 25 stored teams, 23 share one pattern; the 1927 Yankees and 1968 Tigers each use a distinct pattern. Their historical derivation is unresolved.
Read supporting analysis04Verified
Integer arithmetic & quantization
Intermediate truncation, byte wrap and first-crossing thresholds materially affect initial selections. Continuous real-number formulas can give different answers.
Read supporting analysis05Modeled
Defensive positioning
Normal, corners-in and infield-in alter initial weights and later transitions. Strategy rankings depend on the objective, runners and supplied fielding state.
Read supporting analysis06Modeled
Runner advancement
Joint base/out/scoring kernels retain runner identities. Some paths require cached speeds and geometry; first-occupied completed appearances can retain unresolved mass and a continuation tail.
Read supporting analysis07Verified
Random-number generation
The initial selector reads the POKEY hardware RANDOM register. Uniform independent bytes are a calculation assumption; real timing and correlations require runtime validation.
Read supporting analysis08Preliminary
Crowd-animation investigation
Shared hardware RNG confirmed; predictive timing remains unverified. The crowd routine’s identification is strongly supported statically, but has not been visually confirmed in Altirra.
Read supporting analysis03 / PROBABILITY & STRATEGY
From a matchup to a distribution
The object of study is the resulting game state, not just the name of a play.
Batter + Pitcher + Baseline + Game State
+ Offensive Strategy + Defensive Strategy
↓Distribution of Resulting States
+ Offensive Strategy + Defensive Strategy
↓Distribution of Resulting States
Preserve the whole probability space
Joint transitions track runs scored, outs recorded, batter reaching base, runner advancement, double plays and resulting base occupancy. Completed plate appearances also require integrating retries and continuations.
Resolved states, unresolved mass and any nonterminal tail must together sum to 100%. Unknown mass is disclosed separately and never silently redistributed.
Modeled · initial selection only
A documented 256-byte example
Ten initial classes, before downstream runner and fielding logic. The table is a complete initial-selection distribution, not final plate-appearance rates.
| Initial selection | Byte count | Conditional probability |
|---|---|---|
| 0 · Walk | 16/256 | |
| 1 · Strikeout | 109/256 | |
| 2 · Single family | 29/256 | |
| 3 · Double family | 10/256 | |
| 4 · Triple family | 1/256 | |
| 5 · Home-run family | 14/256 | |
| 6 · Mixed out | 64/256 | |
| 7 · HBP / interruption | 4/256 | |
| 8 · Error family | 6/256 | |
| 9 · Groundball | 3/256 | |
| Total | 256/256 | 100.000000% |
Inputs, assumptions & limits
Documented demonstration inputs (not verified season statistics): batter AB 626, H 175, BB 83, SO 187, 2B 31, 3B 8, HR 45; pitcher IP 174.1, H 154, BB 37, SO 174. Store IP as 174 under an explicit floor policy. Baselines H/BB/SO/X = 246/94/157/20; X = 0; offense 0; normal defense; supplied tail cutpoints 243/247/253.
Uniform sampling across bytes 0–255 is assumed. Initial-selection unknown mass: 0%. Final-state unknown mass: not calculated by this table. Live hardware timing, original fractional-IP import and baseline provenance remain unverified. Full derivation and limits ↗
04 / OPEN QUESTIONS
The work still ahead
An expandable research backlog. Each answer must earn its evidence.
Q01Can every legal starting state be resolved?
Open Question Account for all final states without unknown mass. First-occupied continuations, missing native caches and incomplete event categories remain explicit boundaries.
Q02How does runner speed change defensive strategy?
Open Question Extend conditional comparisons across starting states and objectives; avoid transferring third-only results to first-occupied situations.
Q03How do position-specific fielding ratings influence outcomes?
Open Question Separate effective ratings, physical positions and lineup lookup behavior, including nonstandard cases.
Q04What baselines should custom teams use?
Open Question Recover the original normalization generator and test editor behavior. A modern reference population is a modeled design choice.
Q05How can reconstructed distributions be validated in Altirra?
Open Question Capture native state and RNG samples, compare deterministic transitions, then investigate timing-aware frequencies. No native validation has yet been performed.
Q06Can every action fit a general probability matrix?
Open Question Map offensive and defensive choices into joint resulting states, retaining unresolved and nonterminal mass rather than normalizing them away.
Q07Can animation and gameplay be faithfully reproduced?
Open Question Investigate timing, presentation, assets and rights before a future implementation. This publication includes no game assets or playable engine.
Q08Can crowd patterns predict subsequent random outcomes?
Open Question Shared hardware RNG confirmed; predictive timing remains unverified. A controlled timing experiment is needed before any predictive claim.
05 / DEVELOPMENT ROADMAP
Research first. Then implementation.
PHASE 1 · ACTIVE
Research and
Reconstruction
Document original algorithms and create a validated probability engine. Static reconstruction is underway; native validation remains pending.
PHASE 2 · FUTURE OBJECTIVE
Experimental
Simulation
Develop a modern implementation with carefully designed strategic choices while preserving the original game’s visual character.
PHASE 3 · FUTURE OBJECTIVE
Multiplayer and
Tournament
Investigate browser games, simultaneous hidden decisions, online drafts and a possible inaugural tournament of approximately 20 participants.
06 / RESEARCH LOG
A record of the investigation
All listed milestones are dated 24 October 2026 in the research documents. The order follows research progression; no within-day timestamps are asserted.
Arithmetic and evidence checks
The initial validation checkpoint documents 21 passing static tests, including byte arithmetic and evidence integrity. This is a historical checkpoint, not a current full-suite count.
Research record ↗Baseline and matchup reconstruction
Stored team references and the batter–pitcher calculation are documented, with explicit limits on input conversion.
Research record ↗Plate-appearance transition model
Initial classes are composed into conditional transitions; full general closure remains incomplete.
Research record ↗Runner advancement and defensive sensitivity
Later research corrects earlier branch interpretations and compares strategy objectives under documented assumptions.
Research record ↗