DynamisLM is a compact specialized reasoning-LM research project for elite men's senior first-team top-division professional association-football performance science, backed by deterministic Python scientific authority.
The project is not an AI coach and does not currently ship a trained DynamisLM checkpoint. Its scientific contract is deliberately split:
- The language-model layer may resolve terminology, protocols, measurement identity, comparability intent, analysis class, evidence scope, interpretation, and refusal.
- Registered deterministic Python software is the authority for equations, arithmetic, units, signal processing, event/phase detection, metric derivation, statistics, uncertainty, thresholds, and registered comparability adjudication.
- If a required numerical operation is not registered, the correct result is
COMPUTATION_NOT_REGISTERED; the language model must not become an implicit calculator.
The V2 scientific world has three coupled domains:
- Performance testing;
- external load / exposure;
- longitudinal football context.
The Performance Testing domain preserves twelve test families:
- Countermovement Jump (CMJ)
- Drop Jump (DJ)
- Isometric Mid-Thigh Pull (IMTP)
- Squat / Squat Velocity-Based Testing
- Bench Press / Bench Press Velocity-Based Testing
- Bench Press Throw
- Medicine-Ball Throw Testing
- Short Linear Sprint / Acceleration Testing
- Maximum Sprint Velocity / High-Speed Sprint Testing
- 505 Change-of-Direction Testing
- 30–15 Intermittent Fitness Test
- Repeated-Sprint Testing / RSA
These twelve families are the Performance Testing scope, not the whole DynamisLM scientific world. V2's canonical empirical target population is:
SEX = MALE
AGE_CLASS = SENIOR
SPORT = ASSOCIATION_FOOTBALL
PROFESSIONAL_STATUS = PROFESSIONAL
SQUAD_LEVEL = FIRST_TEAM
COMPETITION_LEVEL = TOP_DOMESTIC_DIVISION
Women, academy/youth/U23, university, amateur/semi-professional, lower-division, futsal, rugby, other-sport, generic trained-adult and unseparable mixed cohorts are noncanonical empirical target populations. They may still provide explicitly classified indirect measurement evidence for population-independent mechanics, metrology, signal processing, statistics, device behavior and related methods; they do not establish canonical football priors, norms or thresholds.
This domain covers GNSS/GPS, optical tracking, inertial and vendor-derived measures, match/training exposure, total/relative distance, speed-zone metrics, HSR, sprint, acceleration/deceleration and provider/method/threshold identity.
RES-64 implements the external-load scientific identity extension and the registered deterministic boundary for unit conversion, duration-normalized distance, and threshold summaries from explicit timestamped velocity samples. Modality, provider/system, threshold basis and boundary, event definition, session aggregation, processing, and normalization are identity-bearing:
SAME_LABEL != SAME_MEASUREMENT
SOURCE_REPORTED_VALUE != DYNAMISLM_DERIVATION
Source/provider outputs such as PlayerLoad remain first-class
PROVIDER_DERIVED observations with exact Source A variable and provenance
identity. Their proprietary derivations are not reconstructed or silently
treated as DynamisLM computations. External-load comparisons fail closed when
thresholds, modalities, providers, segmentation, dwell rules, filtering, or
other material identity dimensions differ or are unknown.
This domain covers athlete, first-team squad, club/team, competition, season, training, match, testing session, match exposure, training exposure, microcycle, match-day-relative context and multi-device/multi-source longitudinal history.
The fundamental scientific object is:
ScientificMeasurementObservation
= ObservationContext
+ MeasurementIdentity
+ MeasurementResult
+ Provenance
The language-model/Python authority split and the distinction between Knowledge Scope, Computational Authority Scope and Claim Authority Scope remain binding. ScientificMeasurementObservation is the unit of typed scientific meaning, not a naked label/value pair.
CMJ is no longer merely the next authorized unit. The repository contains substantial CMJ authority through RES-50, including:
- acquisition identity;
- signal validation;
- weighing/system mass;
- events;
- net force / impulse;
- COM mechanics;
- jump-height estimators;
- phase system;
- trial selection;
- session aggregation;
- comparability;
- refusal;
- provenance;
- serialization and ranking authority.
Additional CMJ force, power, RFD, RSI-mod and asymmetry completion work belongs downstream under RES-65. No metric is implied here beyond what the existing registered implementation and tests establish.
RES-64 adds external-load identity, Source A interpretation mapping, unit and relative-distance operations, and a raw velocity threshold-summary operation. Acceleration/deceleration event algorithms, repeated-high-intensity effort algorithms, and proprietary PlayerLoad/IMA/RHIE derivations remain outside DynamisLM computational authority unless a later registered unit supplies their exact definitions and inputs.
RES-75 hosted CI and repository governance is sealed. RES-63 implements the
canonical empirical ingestion boundary; real source bytes and full canonical
tables remain outside Git under ~/data/dynamislm. Empirical data use for
model training remains blocked by MODEL_TRAINING_USE=NOT_AUTHORIZED_BY_RES63.
Question + measurement context
↓
semantic reasoning
↓
typed analysis/comparison request
↓
deterministic Python authority
↓
structured result + provenance
↓
bounded interpretation or refusal
The current V2 and historical architecture documents are curated in docs/architecture. Project-specific operational guardrails are versioned in skills.
Python 3.12 and uv are required. From a Linux-native checkout:
uv sync --frozen
uv run ruff check .
uv run ruff format --check .
uv run mypy .
uv run pytestThe runtime kernel uses only the Python standard library. Development tools are locked in uv.lock.
The source code is AGPL-3.0-only. Model weights and adapters are license-deferred until an upstream base-model license is selected and reviewed. Datasets and corpora remain licensed by their individual sources and provenance. Project name, logo, and branding rights are separate from software copyright. See NOTICE.md, CONTRIBUTING.md, and SECURITY.md.
The current executable program is defined by the canonical Linear sequence:
RES-59 → RES-60 → RES-61 → RES-62 → RES-63 → RES-64...RES-70 → RES-71
RES-59 re-seals the V2 scientific constitution; RES-60 through RES-63 establish typed population, football-world, longitudinal-record and canonical-dataset authority; RES-64...RES-70 expand deterministic scientific engine coverage; RES-71 is the Scientific Engine Qualification Gate. Project control remains in Linear; this repository records executable code and curated authority documents.
No operational model, inference, baseline-inference benchmark, paid GPU qualification, CPT/DAPT, SFT/PEFT, RLVR/GRPO or scaling work may begin before RES-71 records SCIENTIFIC_ENGINE_GATE = PASS. Candidate-model research notes are downstream planning only; no model is shipped here.
- No trained checkpoint, model weights, corpus, database, API, frontend, GPU runtime, or deployment is included.
- The other performance-test families and scientific-engine qualification remain downstream work in the RES-64→RES-71 sequence; RES-63 dataset qualification remains limited to its current ingestion boundary, while RES-64's external-load operations remain limited to their registered identities and methods.
- The current CMJ implementation is not a complete CMJ science program; additional force, power, RFD, RSI-mod and asymmetry methods remain downstream under RES-65.
- The generic kernel and CMJ vertical slice represent registered contracts and provenance; they are not a complete persistence layer or clinical/return-to-play authority.
- Comparability without a registered deterministic rule remains explicitly unresolved/insufficient.
Copyright 2026 Julio Rodriguez and contributors. Licensed under the GNU Affero General Public License, version 3 only.