Docs  /  MathNN Coach + Safety training loop

MathNN Coach + Safety training loop

Module: _MathNNCoachTrain · MathNNCoachTrain.cpp

Connects:

Role System
AI Coach _DomainLLMBridge (curriculum, distill, explain, gaming coach)
Safety _EthicsGood + _LegalShield (+ _GamingEthicsGate for gaming)
Baby net MathNN / _RLAgentNet via distill + _MathNNLearning prefer/align

The LLM is not the 20 ms policy. It proposes curriculum and coaching text; MathNN takes the supervised/RL steps. Safety can drop proposed problems before they touch weights.

Pipeline

Coach: GenerateCurriculumProblems / ScenarioIdeas
        │
        ▼
Safety: LegalShield.AssessText → EthicsGood.EvaluateText
        │        (+ GamingEthicsGate if gaming)
        ▼
Accept / reject each example
        │
        ▼
MathNN: DistillTopicToAgent (RLAgentNet SupervisedQStep)
        PreferTrain + Align(ethics)
        │
        ▼
Coach debrief: ExplainMathRun / CoachEpisodeSummary

Commands

nn coach doctor
nn coach status
nn coach train math_algebra 48
nn coach train math_algebra 64 quadratic equations
nn coach gaming 32 health=0.8 aim=0.5
nn coach align math_algebra
nn coach safety on
nn coach hard on
nn coach score -0.15
nn coach coach on

Aliases (same handler):

llm coach train math_algebra 48
nn learn coach train math_algebra 48
domain llm coach …

Also still available:

llm math curriculum math_algebra 6
llm math distill math_algebra basic skills 32
llm math explain math_algebra
llm gaming coach <metrics>
nn learn align ethics
ethics good status
gaming ethics status

Gaming + film archetype

film seed anakin
film apply gaming
nn coach gaming 32
gaming rl train 1000

Coach summaries stay safety-gated; Anakin callouts are narrative only.

Lua

print(nn_coach_status())
ok, steps, detail = nn_coach_train("math_algebra", 48, "linear equations")
ok, detail = nn_coach_gaming(32, "health=0.9 deaths=0")

Config knobs

Setting Default Meaning
coach on LLM curriculum + debrief
safety on Ethics + LegalShield
gaming_ethics on Extra gate for gaming batches
hard_block on Reject when ethics not ok
min_ethics_score -0.15 Drop weaker scores
Generated from the project markdown docs on 2026-07-24. This is a static, self-contained site.