evidence / Engineering depth
Possession Monte Carlo -- correlation that EMERGES, and a sim that grades itself
The engine simulates the mechanism, not the summary. Teammate correlation is not a parameter I fit -- it falls out of shared-possession mechanics, and the measured value matches reality where a prior simulator's hand-tuned matrix was badly wrong. The sim then grades its own forecast quality by game-state bucket and names the buckets it handles worst. The single truth-source for any figure below is docs/JOB_EVIDENCE_PACKET.md (Section F); edge_claimed=false throughout, and no dollar, ROI, or betting-edge figure appears anywhere on this page.
strongest single receipt
the claim
Most projection tools regress to a summary statistic -- a mean, maybe a variance -- and then staple correlations on afterward with a hand-tuned matrix. I built the opposite: a player-level possession Monte Carlo where the joint structure between players is not an input at all. Two teammates compete for the same finite pile of possessions, so the correct slightly-negative teammate scoring correlation emerges from the mechanics rather than being imposed. The measured emergent value matches realized boxscores, where a prior simulator that imposed a correlation matrix got the sign and magnitude wrong. The sim then grades its own calibration by game-state cell and publishes a ranked backlog of its worst buckets.
receipts
| claim | committed artifact |
|---|---|
| Emergent teammate pts-pts rho ~ -0.10 vs realized (no imposed matrix); fixes prior +0.65 | src/sim/basketball_sim.py |
| Joint pricing straight off the sample block; measured rho -0.104; `validate_joint_calibration` grades sim-joint vs realized (no ROI claimed) | src/sim/sgp_from_sim.py |
| Cross-sport state-cell heatmap: per-bucket PIT/CRPS + ranked worst cells, named ceilings, self-check test | scripts/platformkit/benchmarks/sim_heatmap/build_heatmap.py, .../test_build_heatmap.py |
| Per-sport heatmap outputs (LOCAL: `data/` gitignored; regenerated by `--refresh`) | data/frontend/ops/sim_heatmap_{nba,mlb,soccer,tennis}.json |
| Full engine design (mechanics, `_finalize`, joint pricing) | [docs/architecture/possession-simulator.md](../architecture/possession-simulator.md) |
| Truth-source row | [docs/JOB_EVIDENCE_PACKET.md](../JOB_EVIDENCE_PACKET.md) (Section F) |
why this matters
The hire signal for quant research is not "I built a simulator." It is that the simulator's joint structure is generative -- correlation emerges from the mechanics and lands on the realized value by construction, exactly the property a fitted matrix lacks and exactly where a prior version got the sign wrong at +0.65. Then, instead of one aggregate calibration number, the engine decomposes its forecast quality by game state, names the buckets it handles worst, states its coverage ceilings, and refuses to convert a validated joint structure into a betting-edge claim it did not earn. Building a generative model whose emergent joint behavior matches reality -- and grading it honestly, cell by cell -- is the work, not the headline.
edge_claimed = false everywhere. Every figure is a calibration / correlation / CRPS quantity against realized outcomes, never a dollar figure. Retracted measurement artifacts appear only in [JOB_EVIDENCE_PACKET.md](../JOB_EVIDENCE_PACKET.md), never on this page.
This site reports calibration and sharpness only, never a dollar edge, ROI, or bankroll result. An honest null is a success.