# Pressure waves: code findings and a small observational probe

jdc reports that the CPU sometimes sustains unusually strong pressure in waves,
possibly when one or two goals behind. The score association is a hypothesis.
This investigation has not changed the released game.

## Confirmed in published 2026-10-28-r9

ROM SHA256 811c912c42d2aacb6a7326c8131090df71972deb63a597d0298b049c05277cea.

- `experimental/crowd.s`: shots on goal add 200 and body checks 150 to the team's
  side of one shared momentum value, clipped to ±1000. Each game-clock second it
  decays toward zero by M/128 (at least one unit). Normal and Hard use this
  dynamic system. Current player/keeper attributes are scaled within ±10%,
  with native integer rounding/caps; lower-is-better timing attributes are
  treated in the opposite direction. Weight, roster identity, jersey and flags
  are retained. Goals/score difference are not inputs to this momentum module.
- Native CPU line choice compares team scores during period 3 and chooses a
  different priority table for tied/leading/trailing. The actual published ROM
  comparisons are at F7E2 and F852. Available strength, energy and special-team
  situations still matter; a changed priority table is not a universal boost.
- Native CPU goalie pulling at F746..F78C requires period 3, at most 60 seconds
  remaining, a deficit of 1 or 2 goals and the relevant puck/faceoff location in its
  attacking half. That extra skater can create stronger late pressure. The
  actual v5.1W comparison exits on deficits above 2, unlike an older reference
  assembly that allowed only exactly 2. This is inherited behavior, not a newly
  authored fix or a general all-period comeback switch.

These mechanisms can contribute to the reported feeling. No special universal
“berserk at 1–2 goals behind” switch has been confirmed. No claim that the CPU
receives guaranteed goals or outcomes is supported by these observations.

## Read-only native observation, not a human playtest

Six already completed private r9-Hard-control CPU-vs-CPU games 471..476 were
replayed with identical ROM/commands/batteries, three home per club. Extra host
logging reads scores, SOG, checks, momentum, actual rating percentages, line
numbers, on-ice individual energy, manpower, ownership and puck coordinates.
Every replay's entire result JSON, goal frames and sampled native trajectory
bytes equal its original reference. The observer adds no gameplay writes.

The corrected probe reads PCTH F70E and PCTA F72A; equal/opposite percentages
hold throughout. An initial observer accidentally read the tile tick F728 as
home percent; that dataset is superseded and not used for percentage claims.

Both clubs combined spent 43.8% of sampled live time in their attacking zone
when momentum was at least +500, versus 29.0% at -500 or below. This is observational:
shots/checks drive the meter, states have unequal exposure and six games do not
establish causation. The sample does not establish a 1–2 goal comeback effect.
Actual energy varies, and many shifts use PP/PK units. Player fatigue, line
changes, manpower and momentum should be tracked together when investigating
pressure, rather than assuming one hidden cause.

Private evidence: work/hard-strength-20261028-231416/momentum-probe-corrected/
(results.json, analysis.json, per-match momentum.json and exact-equivalence
checks). See AI_STRENGTH_WIP.md for the separate unfinished Hard-strength work.