CRIA Combat Intelligence OS v1

Mission

Build a combat system where the player is doing grappling rather than selecting canned grappling animations.

The system must make Brazilian Jiu-Jitsu, submission grappling, wrestling and NO-GI/ADCC exchanges readable, responsive, tactically deep and rules-correct while remaining compatible with CRIA's 2D premium pixel-art renderer.

The simulation must be rich enough that a future 3D or hybrid renderer could visualize the same fight state without rewriting gameplay rules.

Authority order

1. data/combat/bjj_rulesets_verified_v1.json — legality/scoring authority.
2. data/combat/bjj_completion_gate_v1.json — full BJJ completion authority.
3. authoritative BJJ graph/reducer sources.
4. data/combat/combat_intelligence_contract_v1.json — enriched authoring/calibration semantics.
5. fighter/canon data.
6. data/research/grappling_video_corpus_contract_v1.json — evidence intake policy.
7. accepted expert-reviewed observations.
8. ML/video output.
9. this skill.

Never let a model, video classifier, animation or LLM override items 1–5.

Core separation

INPUT
  ↓
INTENT / CONTROL MAPPING
  ↓
DETERMINISTIC COMBAT SIMULATION  ← sole state authority
  ↓
RULES / SCORE / CLOCK
  ↓
TACTICAL AI + FIGHTER IDENTITY   ← proposes actions only
  ↓
AUTHORITATIVE RESULTING STATE
  ↓
ANIMATION BINDING / AUDIO / CAMERA / HUD

Render state is not combat state.

Animation playback cannot grant a sweep, pass, takedown, score or submission. It visualizes a transition already accepted by the simulation.

Why the existing reducer stays

BJJGraphReducerV2 already gives CRIA valuable properties:

• seeded determinism;
• position/role legality;
• gas costs;
• rules filtering;
• counter timing;
• stabilization before scoring;
• replayable logs.

Do not replace it with a neural network or physics sandbox.

Combat Intelligence v1 enriches the state and planning around the reducer and creates a migration path for a future reducer v3 only after tests prove parity.

Grappling state model

Contest state

Track independently from athlete pose:

• ruleset;
• regulation/overtime/finished;
• phase clock;
• positive-scoring phase versus no-positive-points phase;
• mat zone and out-of-bounds restart context;
• score;
• penalties/negative points;
• pending stabilization events.

This is mandatory for ADCC-style strategy because the same technical exchange has different strategic value before and after positive scoring activates.

Athlete state

Each athlete state should support:

• posture;
• support points and base quality;
• center-of-mass projection estimate;
• hip/shoulder/head orientation;
• inside position;
• frames and wedges;
• grip/tie graph and grip integrity;
• underhook/overhook/head-control state;
• leg-entanglement topology;
• knee-line and heel exposure;
• off-balance/kuzushi;
• pressure direction;
• mobility freedom;
• pinned segments;
• submission-danger components;
• gas;
• isometric load;
• local fatigue;
• decision confidence.

Unknown is a valid state. Never hallucinate an occluded limb or grip into certainty.

Interaction state

Model the two athletes as a contact graph, not as two independent pose skeletons.

Important edges include:

• grip/tie;
• frame;
• hook;
• post;
• wedge;
• pin;
• head control;
• chest/hip contact;
• leg entanglement;
• mat support.

This graph is the bridge between biomechanics, gameplay and paired animation.

Technique model

A CRIA technique is a structured causal transition.

Minimum dimensions:

TECHNIQUE
├── from / to
├── attacker/defender roles
├── rules + GI/NO-GI legality
├── preconditions
├── entry cues
├── required grips/ties/connections
├── body alignment
├── weight shift
├── force direction
├── attacker objective
├── defender responses
├── counter branches
├── success topology
├── failure topology
├── energy/isometric cost
├── risk profile
├── score event + stabilization
├── six paired animation phases
└── recovery state

A generic prompt such as perform armbar attack is never a technique definition.

Six-phase paired motion contract

Every combat animation remains aligned with the canonical BJJ phases:

1. anticipation — setup/intention; cancel or counter opportunity still exists;
2. entry — committed level/angle/grip/tie change;
3. establish — required contact topology is obtained;
4. stabilize — control is made score-valid where rules require it;
5. response — defender frame/counter/escape branch resolves;
6. recovery — both bodies settle into the authoritative resulting state.

The attacker and defender are one animation system with two visual bodies. Their timelines are not authored independently.

Real-time grappling contact

Adapt the useful principle of real-time contact systems without cloning proprietary implementation.

For grappling, contact windows mean:

• a grip/tie can become valid over a short timing range rather than one magic frame;
• feet/knees/elbows/posts must coincide with the intended support topology;
• a takedown entry can be redirected if the defender wins head/hip/inside position during the response window;
• animation blending may correct small visual spacing errors but cannot override the state graph;
• major contact mismatch is a QA failure, not something to hide with interpolation.

Prioritize responsive input-to-state timing over longer cinematic clips.

Fighter identity — CRIA Rhythm

Do not express fighter identity only through ratings.

A fighter identity profile should influence preferred:

• range;
• ties/grips;
• takedown families;
• guard families;
• passing families;
• top-control patterns;
• scramble frequency;
• submission families;
• pace;
• counter tendency;
• risk tolerance;
• late-match strategy.

Cria Rhythm rewards coherent execution of the fighter's game plan. It may affect confidence, transition efficiency or planning priority, but never legality or guaranteed outcomes.

Examples:

• Ruan: pressure, grip/control, top-game, progressive positional consolidation;
• Davi: timing, counter, scramble, redirecting overcommitment.

Two fighters with similar numeric technique values should still make different decisions.

Tactical AI stack

Use a hybrid architecture:

Layer 0 — hard rules

Filter impossible/illegal actions first.

Layer 1 — statechart

Manage match phase and tactical macro-state:

• neutral standing;
• grip/tie battle;
• takedown exchange;
• guard engagement;
• passing/control;
• scramble;
• submission offense/defense;
• boundary restart;
• late-score strategy.

Layer 2 — utility planner

Score legal candidates using:

• expected position value;
• scoring value under current clock phase;
• finish opportunity;
• counter exposure;
• energy and isometric cost;
• boundary risk;
• opponent tendencies;
• fighter-style fit;
• chain continuation value;
• escape urgency.

Layer 3 — short-horizon chain search

Evaluate likely action→counter→continuation branches rather than one move at a time.

Layer 4 — optional offline learning

RL/self-play or behavior cloning may help calibrate policies offline. It never becomes rules authority and must pass deterministic evaluation before use.

LLM boundary

LLMs are acceptable for:

• annotation assistance;
• research-query generation;
• coach/commentary drafts;
• living-world dialogue.

Never put an LLM in the frame-by-frame inner combat loop.

Evidence-first video laboratory

Do not claim that CRIA has learned from a video until that video exists in the source ledger and its derived observations have passed the review contract.

Pipeline:

RIGHTS LEDGER
  ↓
FFmpeg normalization
  ↓
scene/round/exchange segmentation
  ↓
optional Whisper transcript
  ↓
SAM2 athlete masks
  ↓
MediaPipe / MMPose-RTMPose pose candidates
  ↓
Pose2Sim for controlled multiview captures
  ↓
MMAction2 / V-JEPA / InternVideo temporal evidence
  ↓
structured grappling observation
  ↓
expert BJJ + rules review
  ↓
DuckDB/Parquet + vector index
  ↓
calibration + key-pose reference

What competition video is good for

• technique frequency;
• preferred chains;
• tactical score/time decisions;
• entry timing;
• common counters;
• match rhythm;
• broad phase duration;
• visual/camera reference.

What single-view broadcast video is bad for

• exact hidden grips;
• exact occluded limb orientation;
• exact joint torque;
• submission safety thresholds;
• production-quality 3D reconstruction during heavy entanglement.

Never convert low-confidence inference into a fact.

Own motion-capture studio

For final movement fidelity, prefer CRIA-owned recordings.

Recommended session:

• 4–8 synchronized camera/phone views;
• calibrated mat scale and camera intrinsics/extrinsics;
• two consenting trained athletes;
• unbranded high-contrast rashguards or approved GI;
• technique script derived from the authoritative graph;
• normal-speed repetition;
• controlled-speed repetition;
• common counter branch;
• common failure branch;
• explicit tap/safety protocol;
• athlete and footage releases.

Record clean key states for setup, entry, contact, establishment, defense and recovery.

For multi-view reconstruction use Pose2Sim/RTMPose or another audited stack as research tooling. Keep original video and calibration files immutable and hashed.

Dataset policy

Research datasets can be very useful without being shippable.

Examples:

• ViCoS BJJ positions: benchmark/research only under its current noncommercial CC BY-NC-SA terms;
• Open FSW wrestling dataset: valuable close-contact action-recognition benchmark, but underlying media rights must be reviewed independently from repository code license;
• tiny/incomplete BJJ datasets: schema inspiration only, never evidence of model readiness.

Training rights, code license, checkpoint license and source-media rights are four separate fields.

Visual production bridge

Combat Intelligence produces movement/reference metadata; Sprite Forge V2 produces the visual candidate.

For pixel art:

1. lock both character identities;
2. select accepted starting paired pose;
3. author/generate the whole phase strip or coherent blocks, not isolated frames;
4. normalize shared scale/pivot/contact space;
5. validate content bounds and contact continuity;
6. compare phase endpoints to combat state;
7. run paired preview;
8. integrate in Godot only after rights/human review.

The primary shipping renderer remains premium 2D pixel art. The simulation must not encode pixel-art assumptions so a future 3D/hybrid representation remains possible.

Rules profiles

Support multiple rules profiles through adapters, not conditionals scattered through animation code.

At minimum:

• IBJJF GI;
• IBJJF NO-GI where relevant;
• ADCC World-style professional rules;
• ADCC Trials/Open variants;
• fictional CRIA clandestine rules only where canon explicitly defines them.

ADCC needs first-class concepts for:

• no-positive-points phase;
• positive/negative scoring phase;
• overtime;
• negative points;
• out-of-bounds restart while preserving established grips/position;
• tactical behavior changing with clock phase.

Validation metrics

Combat quality cannot be measured only by win rate.

Track:

• illegal action rate = 0;
• impossible source-position transitions = 0;
• deterministic replay mismatch = 0;
• score/rules mismatch = 0;
• paired contact mismatch rate;
• animation endpoint/state mismatch rate;
• input-to-state latency;
• policy diversity by fighter;
• action entropy without random nonsense;
• calibration error for empirical probabilities;
• expert agreement/disagreement;
• video model confusion by position/technique;
• performance on high-occlusion holdout;
• CPU/frame budget on target Android.

Fail-closed gates

Do not claim complete combat until:

• full 40/120/10 authoritative BJJ source is present;
• rule adapters are tested;
• paired animations exist for required transitions;
• F3 empirical calibration is complete;
• F4 balance is complete;
• expert biomechanical review is recorded;
• deterministic replays pass;
• visual runtime QA passes;
• physical Android evidence passes.

Engine migration

Current project/CI versions are below the current supported Godot line. Do not upgrade main as part of unrelated combat work.

Follow data/production/godot_migration_plan_v1.json on a dedicated migration branch, maintaining a dual-version evidence window until runtime/import/export tests are green.

Commands

python tools/combat/validate_combat_intelligence_v1.py
python -m unittest discover -s tests -p 'test_combat_intelligence_v1.py'
npm run quality

Definition of success

The system is successful when a knowledgeable grappler can explain why a transition was available, what contact/base/angle made it work, what counter window existed, why the rules scored or did not score it, and the paired animation visibly tells the same story.