Skip to content

Latest commit

 

History

History
93 lines (75 loc) · 4.42 KB

File metadata and controls

93 lines (75 loc) · 4.42 KB

LevelField validation report

Generated 2026-08-31 by scripts/validate.ts. Reads data/scores/index.json (produced by scripts/score-all.ts) and each curated contract's expected.category field (data/curated/*.json); does not itself run any classification.

Method and its honest limits

This report checks two things, both against the curated contract set only (the live DreamDEX track has no category label and is excluded from every table below):

  1. Category median ordering — do median scores rise in the order the ACDC taxonomy predicts (market_data ≤ statistical ≤ public < institutional < individual)?
  2. Spearman rank correlation — across all scored curated contracts, does a contract's score rank track its category's expected rank?

Neither check is proof that LevelField's scores predict real-world insider activity. The reference classifications behind the scored contracts are single-run by design — one classification per contract, not a majority vote across independent runs (that only happens for live/ad-hoc contracts via voteDimension, see packages/scoring/src/vote.ts) — so a single misjudged dimension moves a contract's score with no averaging to catch it. The category labels themselves are also single-run, hand-assigned judgment calls, not an independent ground truth. What this report actually supports is a narrower claim: LevelField's deterministic scoring engine reproduces, from category-blind per-dimension levels, an ordering consistent with the taxonomy ACDC validated against 435,000+ settled Polymarket markets (data/anchors/anchors.yaml). It is evidence of internal consistency with that taxonomy, not external validation of predictive power, and the sample size (currently 16 scored curated contracts) is far too small to support a statistical claim on its own.

Coverage

16/16 curated contracts are scored. Contracts below are missing a data/classifications/{marketId}.json reference classification and were skipped by score-all.ts; they are excluded from every table below except the last.

  • (none — every curated contract has a classification)

1. Category median ordering

category n scored median score
market_data 1 3
statistical 1 19
public 2 23.5
institutional 8 60
individual 4 95
adjacent pair median A median B result
market_data <= statistical 3 19 PASS
statistical <= public 19 23.5 PASS
public < institutional 23.5 60 PASS
institutional < individual 60 95 PASS

2. Spearman rank correlation

rho = 0.930, n = 16.

Computed as the Pearson correlation of tie-aware ranks (average rank within ties) between each contract's overallScore and its category's position in the ACDC order (market_data=1 … individual=5). Ties are expected and handled explicitly: multiple contracts can share a category (e.g. institutional).

3. Circuit breaker hits

marketId category score circuit breaker
curated-celebrity-breakup individual 95 CB-1
curated-ceo-resignation individual 80 CB-1
curated-injection-test individual 95 CB-1
curated-presidential-pardon individual 95 CB-1

4. All curated markets

marketId category score band circuit breaker status
curated-award-show institutional 45 moderate scored
curated-bill-passage institutional 73 elevated scored
curated-btc-120k market_data 3 low scored
curated-celebrity-breakup individual 95 high CB-1 scored
curated-ceo-resignation individual 80 high CB-1 scored
curated-company-layoffs institutional 65 elevated scored
curated-court-ruling institutional 54 elevated scored
curated-cpi-above-3 statistical 19 low scored
curated-earnings-beat institutional 55 elevated scored
curated-election-winner public 21 low scored
curated-fed-rate-cut institutional 49 moderate scored
curated-football-match public 26 moderate scored
curated-injection-test individual 95 high CB-1 scored
curated-military-strike institutional 78 high scored
curated-presidential-pardon individual 95 high CB-1 scored
curated-protocol-upgrade institutional 68 elevated scored