STRATHMARK

STRATHMARK is an installable Python engine for AAA-compatible woodchopping handicaps. Version 2.0.0 uses a reproducible, prior-only statistical model and a deterministic joint mark optimizer. It works offline and exposes both Python and REST interfaces.

Install

python -m pip install "strathmark @ git+https://github.com/SquirmyWormy275/STRATHMARK.git@v2.0.0"
python -m pip install "strathmark[api] @ git+https://github.com/SquirmyWormy275/STRATHMARK.git@v2.0.0"

STRATHMARK 2.0.0 is published as the immutable v2.0.0 Git tag and GitHub release; there is no PyPI distribution for this version. Production consumers may pin the exact release commit for an even stricter lock. Other optional extras are ml, db, llm, and dev.

The NumPy/Pandas V2 core and its validated JSON artifact ship in the base package. No network, database, LLM, or native ML library is required to calculate marks.

Quick start

from datetime import date

from strathmark import HandicapCalculator
from strathmark.predictor import CompetitorRecord, HistoricalResult, PredictionContext, WoodProfile

competitors = [
    CompetitorRecord(
        name="Alice",
        competitor_id="competitor-alice",
        gender="F",
        history=[
            HistoricalResult("SB", 28.4, "Pine", 300, 5, date(2025, 3, 1)),
            HistoricalResult("SB", 27.9, "Pine", 300, 5, date(2024, 11, 15)),
        ],
    ),
    CompetitorRecord(
        name="Bob",
        competitor_id="competitor-bob",
        gender="M",
        history=[HistoricalResult("SB", 35.2, "Pine", 300, 5, date(2025, 3, 1))],
    ),
]

wood = WoodProfile(species="Pine", diameter_mm=300, quality=5)
results = HandicapCalculator().calculate(
    competitors,
    wood,
    event_code="SB",
    context=PredictionContext(prediction_as_of=date(2026, 1, 1)),
)

for result in results:
    print(result.name, result.predicted_time, result.mark, result.interval)

calculate() returns MarkResult objects ordered slowest to fastest. Important fields include predicted_time, mark, method_used, interval, performance std_dev, engine/model/calibration versions, optimizer metadata, warnings, and optional trusted ledger state.

quality=5 remains required by the legacy WoodProfile constructor, but wood quality is deliberately a numeric no-op in V2 until provenance-backed quality or moisture data exists.

Prediction Engine V2

The core models log finish time with robust hierarchical partial pooling. Verified inputs are competitor identity and strictly prior history, event, result date and exclusive UTC cutoff, diameter, species physical properties, and gender including missingness. Zero- and sparse-history competitors pool toward supported population priors. Chronological conformal intervals stay separate from race-performance std_dev.

Division, round/heat, venue, lane/stand, run order, exact material identity, wood quality/moisture, weather, equipment, fatigue, penalty/DNF state, same-tournament weighting, and field strength are accepted where compatibility requires but cannot change a V2 number. Numeric LLM prediction is retired; LLMs remain narrative-only.

The five legacy keys remain:

• manual: uncalibrated operator override;
• llm: always None numerically;
• ml: optional promoted residual learner, inactive in 2.0.0;
• baseline: authoritative V2 core;
• panel: labeled broad-prior fallback.

A field uses one immutable model bundle. Marks are chosen from 2,048 deterministic joint posterior samples, subject to the 3-second floor, effective ceiling, monotonic ordering, and a rounded-gap fallback.

Read Prediction Engine V2 for the complete evidence, calibration, optimizer, compatibility, and ledger contract.

Locked benchmark

On the frozen 128-row temporal test set, the V2 core recorded MAE 16.1301 seconds versus 20.5172 for the strict prior-only incumbent (21.38% lower), and RMSE 33.6904 versus 44.4791 (24.26% lower). The central 90% interval covered 94.53% of rows. These figures describe that fixed workbook and split; they are not a universal accuracy or fairness claim. Cohort samples are smaller, and the optional residual learner was inactive.

Verify the checked-in report and packaged artifact without reopening the locked data:

python train_model.py

REST API

pip install "strathmark[api]"
uvicorn strathmark.api:app --host 127.0.0.1 --port 8000

• GET /health — core availability and cutoff compatibility, calibration, residual, artifact, store, and narrative-LLM health; accepts optional prediction_as_of
• POST /predict — one stateless prediction and all five compatibility keys
• POST /calculate — stateless field prediction and mark assignment
• POST /simulate — independent fairness simulation, capped at 250,000 races
• POST /ledger/calculate — authenticated, idempotent trusted calculation
• POST /ledger/predictions/{prediction_id}/settle — authenticated immutable settlement
• POST /results and GET /results/{competitor_name} — authenticated local history
• POST /v1/shadow/calculate — trusted field calculation or exact receipt recovery
• POST /v1/shadow/receipts/lookup — immutable receipt lookup with live status
• POST /v1/shadow/status — trust, freshness, mirror, outcome, and drift status
• POST /v1/shadow/outcomes/apply — field-atomic settlement or void revision
• POST /v1/shadow/mirror/replay — bounded administrator mirror replay
• POST /v1/shadow/drift — bounded advisory-only drift report

Public calculation routes do not write trusted training evidence. Ledger and result routes require STRATHMARK_API_TOKEN.

All six /v1/shadow/* routes require both a consumer-specific bearer credential and a short-lived v2 actor attestation. The attestation binds the action, run revision, and canonical validated request digest. Trusted calculation also requires an attested offline-single-writer-durable topology and a current, integrity-verified local evidence snapshot. Receipt-bound settlement/void remains available when the current snapshot is stale or missing so evidence can be corrected safely. See the canonical Shadow Consumer Contract for roles, request limits, readiness gates, frozen OpenAPI checksum, and the complete offline lifecycle.

Stable design rules

• Mark floor: 3 seconds.
• System ceiling: 183 seconds; event configuration may lower it.
• Joint optimizer: deterministic, fixed seed/sample/pass budget, never wall-clock based.
• Fallback mark: 3 + round(slowest - predicted), bounded to floor and ceiling.
• Evidence: strictly earlier than one exclusive UTC request cutoff.
• Forecast interval and simulation std_dev: different quantities.
• Numeric LLM: prohibited.
• Output: plain text, no terminal-control formatting.

Documentation

• Prediction Engine V2
• Architecture
• Deployment runbook
• Shadow Consumer Contract
• Wiki source
• Contributing
• Changelog

Development

pip install -e ".[dev,api]"
pytest tests -q
ruff check .
ruff format --check .
python train_model.py

Tests use isolated temporary databases. Never point tests at a production database.

License

Apache License 2.0. See LICENSE.

Author

Alex Kaper — SquirmyWormy275