NHL 95 (Genesis) — Decoded

A reverse-engineering deep-dive into NHL 95 (USA, Europe).gen, the 1994 Sega Genesis classic — full static (Ghidra) + live (BlastEm debugger) analysis of its data formats and game logic.

Browsable site: https://breakablehoodie.github.io/nhl95-decoded/ — the same docs below, plus a search page.

New here? Start with docs/OVERVIEW.md — what we found, in plain English, no assembly required. Want the proof? docs/FINDINGS.md is the full technical write-up — every claim backed by ROM addresses, disassembly, and live-debugger evidence. Unfamiliar term anywhere (what's a "nibble"?)? docs/GLOSSARY.md explains every piece of jargon in plain English.

Why should you care?

This started from one real bug report and grew into answers to questions the NHL 95 community has argued about for 30 years:

• Is "hot/cold streaks" real, or just flavor text? It's real — fully traced end to end, from the exact RNG algorithm (a 32-bit LCG seeded once per boot off the Genesis's V-counter hardware) through to which player gets picked hot/cold each game, confirmed live against the actual on-screen announcement. See §5.
• What's the actual formula behind a player's Overall Rating? Solved two independent ways: statistically (a fixed linear combination of 10 specific nibbles out of the game's 7-byte attribute block, live-validated to within ~2 points of the ROM's own output), and then confirmed a second time by decoding the ROM's own UI-widget bytecode directly — the exact set of nibbles the formula uses is bit-for-bit identical to a parameter found sitting in the ROM itself, not just inferred from outside data. Every named stat (Agility, Shot Power, Checking, etc.) is mapped the same way, for both skaters and goalies. See §6.
• The bug report that started this whole project: Boston's Bryan Smolinski shows up cloned at two positions at once in the Line Editor. Root-caused to a specific stale-data condition — and checking all 208 line/team combinations in the game confirmed it's the only one, a genuine one-off 1994 shipping bug, not a general glitch. See §3.
• The full 7-line system (Sc1/Sc2/Chk/PP1/PP2/PK1/PK2) mapped and confirmed byte-for-byte against a live penalty kill — including why the in-game Line Editor sometimes shows only one line and sometimes shows all seven, a real UI-behavior gotcha this project hit before it understood the setting driving it. Mainly useful if you're building tools against this ROM (a save editor, another stats tracker) and need to know which internal slot is actually the penalty-kill unit.

What's in this repo

• docs/OVERVIEW.md — the plain-English version of every finding below, no ROM addresses or assembly required.
• docs/FINDINGS.md — the living document. Every finding, with the reasoning and evidence behind it, root-cause first.
• docs/full_roster_database.json — every player's name, jersey number, and attribute bytes, extracted from the ROM.
• docs/external_sources/ — our derived comparison data (ROM-predicted vs. third-party stats) and the fitted formula models themselves, used to validate the formulas above. Raw third-party scrapes are excluded — see below.
• tools/ — the actual scripts and live-debugger tooling used to do this work: a Ghidra data-dump script, a persistent BlastEm debugger-console daemon (nhl95_daemon.py/nhl95ctl.py) for scripted live tracing over SSH, an unattended-game instrumentation script (nhl95_monitor.py) that polls RAM during CPU-vs-CPU games and logs state changes to CSV instead of hand-driving screenshots, a static ROM string/pattern scanner (rom_scan.py), and the statistical correlation scripts behind the rating-formula work.

Not included, on purpose: the ROM file itself (Sega's copyrighted binary), the Ghidra project database (it embeds analyzed ROM bytes), and two raw third-party data scrapes used only for cross-validation. None of those are ours to redistribute. If you want to reproduce this work, get your own legally-obtained copy of the ROM (US/Europe, no header, 2MB, product ID T-50856) and re-run the Ghidra import described in CLAUDE.md.

Method, if you're doing similar work

Two techniques here generalize well beyond this one ROM:

1. External-dataset correlation to crack an opaque data format. Rather than trying to derive the Overall Rating formula from disassembly alone, fitting the ROM's raw attribute bytes against an independent, already-published stat list (from a game FAQ) turned an opaque 7-byte block into a fully solved, live-validated formula in one session.
2. Live-vs-static verification as the actual bar for "solved." Several findings here were fit or hypothesized statistically first, then proven (or revised) by reading real values off the running emulator via a scripted debugger console — see docs/FINDINGS.md for more than one case where the live check caught a wrong hypothesis before it got written down as fact.

Credits

This project drew on several existing community resources to get started, and is grateful for all of them: nhl-95.com's ratings data and the GameFAQs community FAQ this project also used were both valuable, good-faith starting points that made the correlation work in FINDINGS.md possible in the first place. Sega Retro's NHL 95 page was an independently useful developer/manual-sourced reference throughout.