# Separate penalty frequency — 2026-10-31-r8 ## Menu and rule discovery The last start-menu row is **Penaltys: 100% / 50% / Off**, default 100%. The pre-existing combined Penalties option is renamed **Rules**. Its three original values remain Off, On, and On, Except Off-sides. It still controls native foul eligibility and whether offside and awarded penalty shots are active. The percentage is independent. 100% retains the existing frequency under the chosen Rules; it does not turn every check into a foul. Rules Off still blocks the native calls it already blocked. Icing keeps its own separate On/Off row. The [initial discovery](https://dela.jdc.se/nhl94/notes.html#PENALTYS_RESEARCH.md) is retained, with the original source locations and explanation. 50% uses one native random draw for each eligible ordinary foul, retaining approximately half. Off admits none of those new ordinary calls. A free queue slot and no already pending offender are required before drawing. Duplicate, blocked and full-queue requests retain their native behavior without rerolling. The same setting applies to both teams, CPU and human players, at every difficulty. Already queued penalties finish normally; this menu is not a cancellation tool. Ordinary events are elbowing `0x12` and charging `0x16`, slashing `0x18`, roughing `0x1A`, cross-checking `0x1C`, hooking `0x1E`, tripping `0x20`, interference `0x22` and holding `0x24`, in hex. An injury-associated elbowing call suppressed by the percentage becomes native zero-duration event `0x14`, the existing injury-without-penalty report. The injury stoppage remains. Injury and checking physics are not changed. Fighting must remain one incident: 50% gates an entire eligible new fight after the original fighting-rating and HV-counter chance checks. Off prevents new fights even when the independent Fighting choice remains saved as On. 100% keeps the existing Fighting setting and chance. Fighting `0x26/0x28` and instigator `0x2A` awards are never filtered individually, so paired resolutions retain their original code. A fight already underway can finish normally. Icing `0x0C`, offside `0x10`, faceoffs, other zero-duration events and negative-duration penalty-shot variants are not percentage-filtered. P1–P4 awarded-shooter ownership is unchanged. All five percentage/Height/Speed/Mode/Cone choices save independently for each Full/Light/Mini season and lock after its first completed match. The lock survives reboot, slot changes and playoffs. Confirmed restart resets only that season; standalone preferences remain independent and editable. Controller mappings, pass behavior and the always-switch-nearest-other policy retain their code. ## Native hooks and storage | ROM offset | r7 bytes | r8 bytes | Purpose | |---|---|---|---| |11F78|067800C8B89C|4EF9002EC800|High-event admission, after existing native guards| |2004F0|284B7005267CFFFFB04A|4EF9002EC8044E714E71|Whole-fight commit after original HV chance| |2EC800–2ECFFF|Free|Bounded helper|Exports+0 event admission,+4 whole fight| |2E1088|610000DC4A40|40E7007C0700|Protect the human owned contact-to-shot transition| |2E1260|Penalties label|Rules label|Same native D056 option and values| |2E1310 /2E1484|Free|Penaltys label /100%,50%,Off strings|Eighteenth row, logical34| The helper preserves registers and SR around classification. Accepted calls replay native ADDI#200,B89C and rejoin11F7E; suppressed ordinary calls unwind the existing native frame at11FEC. Injury-only reports use event `0x14` and that native continuation. 100% and invalid values draw no filter RNG.50% calls native11086 with bound2 once for eligible events or a whole fight. Existing sound, queue, delayed-penalty flags, duration, clock, PIM and scoring paths run for admitted calls. Native Rules and penalty-shot checks, original fight eligibility/HV read, icing and pass bodies remain byte-exact. Extra instructions can affect later native HV-based decisions. No general animation/controller/frame hook, rating or physics change is added. RAM `F006` is one active setting word. There are no additional SRAM bytes. The existing complemented two-byte pass preference record is extended: per-season logical `1EEE–1EEF` and standalone `4D20–4D21` in Light's excluded shared-profile gap. Legacy `50–53` and r7's `80–9F` tags load100%. Bits 0–4 retain Height/Speed/Mode/Cone; bits 5–6 encode rate 0=100%, 1=50%, 2=Off. Tags `E0–FF` reserve rate 3 and default all five choices on load; a bad complement also defaults. Returning to legacy combinations writes their original encoding. Existing mapped put/load/writer ABI, checksums, adjacent season snapshot and Full goalie data are tested. Build normally with `--version 2026-10-31-r8`. `--legacy-penalty-rate --version 2026-10-31-r7` reconstructs the exact released r7 image. The website exposes named hooks, exact guarded bytes, the complete base diff and line-addressable sources. ## Evidence and limits The complete `penalty_rate` regression requires seven sections: - 564 actual native admission contexts cover both teams, every event class, default/invalid values, paused/stopped/native Rules-Off, duplicate and full-queue guards, real queue/actor/delayed flags, complete register/SR/stack ABI and RNG. - 54 series of 2000 evolving native event decisions each:108000 controlled draws exactly match a separate original-native-RNG oracle, including the final seed. Every series stays within 45–55% acceptance. This is not 108000 matches. - 48 paired whole-fight commit contexts cover both teams, CCR extremes, all values, real RNG and start/reject endpoints. Admitted fights retain native register output; rejected fights do not partially transfer actor registers or start one fighter. - 48 actual native 1446A injury-caller cases compare against the original Rules-Off injury-only oracle, including keeper/skater, both teams, reports and stoppage. The audio sink and final bench continuation are stopped after the tested outputs. - 1216 exhaustive tag/complement/five-choice/load-ABI contexts and 40 actual five- setting writers verify independent companions, native preferences, locks and adjacent-byte preservation. - 30 real native menu/lifecycle cases cover both controller configurations, rendered 100%/50%/Off text, slot circles/reboot, legacy/corrupt data, separate standalone preferences, Full goalie storage, restart and actual first-match save/lock. A declared final-goal puck/clock fixture lets native scoring/save complete. - Three cold games driven by actual emulator controller input exercise all settings and P1–P4, with actual native foul decisions and queues observed. Only observer counters are initialized: no game positions, fouls, queue entries, scores or RNG are injected. Private preserving probes may alter native HV timing. Final counts are bound to the release hash in PENALTYS_RESULT.json; these are not physical human-controller recordings or a competitive balance claim. The untested first prototype hooked all event entries; it was retained before moving the filter past low event handling. A later working prototype placed its helper at2DA000, which collided with existing private reset-driver probes. The partial shared report and exact ROM/source/proofs are retained; no failed check was accepted. The final helper uses the declared free 2EC800 reservation and all 89 shared suites run freshly again against its exact ROM/source hashes. Early host failures remain recorded too: menu rendering was initially checked before the bottom row was visible, one global-record expectation was calculated incorrectly, and the first short input run saw only two half-rate decisions, both rejected. The final driver extends all three games and sends real Start input through native intermissions, without picking seeds or forcing outcomes. An initial AI comparison used the wrong pre-game state and failed its baseline; the standard immutable trail fixture is used in final shared QA. All 89 suites execute: 88 pass, only the retained finisher_hold geometry failure from jdc's selected 38-win base; the strict gate remains FAILED. The current 40scene comparison records all exact frames and every divergence against fresh r7 controls; those controls retain every earlier strict qualifier. Full numeric results are in QA_STATUS.json and PENALTYS_RESULT.json. Native event output/ABI at100% is separately paired; no universal unchanged-gameplay, new Hard win-rate or improved pass-completion claim is made. Independent rebuild, exact r7 rollback, matched ablations and original-base IPS reconstruction pass. ## Buffered-shot regression guard The first complete 89-suite r8 run exposed a human buffered-shot gap: contact was accepted at the correct skating hotspot, but the shot observer later saw a different sprite hotspot 42.047592 pixels away from the same puck/body positions. The exact pre-fix ROM is retained as an independent negative control. A late accepted-branch observer confirms valid contact before setup; extra probe timing changes the interleaving, so both the original and SR-aware observers are retained in tests. The owned buffered-contact branch now saves SR and masks interrupts from contact validation until native shot completion, restoring the original upper SR and the actual post-shot CCR. The wait path restores both too. This does not change the 4-pixel contact radius, puck position, native shot physics or CPU shooting code. The module remains bounded below the menu labels at 2E1200. The pre-fix 42 px release reproduces with the original observer; the corrected original-observer run and independent SR-aware mash/hold runs stay within 4 pixels. 192 paired frame contexts cover all 12 actors, near/wait branches and all 8 initial IRQ-mask levels, comparing full registers, CCR, restored IRQ mask, stack, pending request and shot outcome with the native pre-fix contract. SR-aware actual shot entries must observe mask level 7. The original r9 negative control and all 432 contact boundary cases remain. Build `--legacy-contact-atomic` to reconstruct the exact private pre-fix r8 control. `--legacy-penalty-rate` also restores the previous unprotected code for exact r7 rollback. Current QA binds this corrective change to the same released image; the failed whole run and all prototypes/source remain retained. ## Delivery The dated r8 replaces current Downloads, canonical and website ROM/IPS with fresh game progress, per jdc's standing policy. Controller/config/history hashes are preserved. A running RetroArch stays open; progress cleanup waits for normal exit. Bodin's existing playlist item receives the same ROM when reachable. Queued offline work is recorded as pending until independent verification. No RetroArch launch/closure, forum message or Hard-strength search is performed.