Status: ACTIVE — Bounty eligible
Claim ID: THD-AI-001
Version: 2.0
Domain: Artificial Intelligence / Optimization
The Claim
For the frozen Random Hierarchy Model (RHM) system in this package, canonical THD predicts a lower normalized held-out cross-entropy trajectory area than an exposure-matched WDP duration-allocation comparator.
Define:
D = THD - AREA_MATCHED_PERM_WDP
Lower endpoint values are better, so negative D favors THD.
The registered population-level claim is:
- n=2: mean D <= -0.040
- n=3: mean D <= -0.060
- n=4: mean D <= -0.060
All three depth propositions are part of the claim.
What Is Being Compared
Primary THD arm:
THD_3n_6n2_9n3
Primary comparator:
AREA_MATCHED_PERM_WDP
The comparator is globally rescaled so its total integrated learning-rate exposure matches the THD arm at the same depth.
This is a matched whole-schedule comparison. It is not a claim of pure phase-order causality.
Primary Endpoint
NORMALIZED_HELDOUT_CE_AUC_0_100
This is the normalized area under the held-out cross-entropy trajectory from 0% through 100% of the registered schedule.
Registered Test Conditions
- RHM: v=6, m=6, classes=6, s=2
- depths: n={2,3,4}
- H=20
- width=144
- batch size=128
- optimizer=SGD
- base learning rate=0.1
- terminal learning-rate factor=0.1
- deterministic CPU execution
- N=24 paired worlds per block
- familywise one-sided alpha=0.01
- Bonferroni depth alpha=0.01/3
Update budgets:
- n=2: 2,040
- n=3: 6,120
- n=4: 13,680
THD phase durations:
- n=2: W=120, P=480, D=1,440
- n=3: W=180, P=1,080, D=4,860
- n=4: W=240, P=1,920, D=11,520
Falsification Rule
A 24-world PRIMARY counterclaim block creates a FALSIFICATION_CANDIDATE at any depth where the one-sided Student-t lower confidence bound for mean D is strictly above the registered negative margin:
- n=2: lower bound > -0.040
- n=3: lower bound > -0.060
- n=4: lower bound > -0.060
Exact equality is inconclusive.
Bounty-Final Replication
A primary falsification candidate is not bounty-final by itself.
A second fresh 24-world REPLICATION block is required.
Every depth that triggered in PRIMARY must independently satisfy the same falsification rule again in REPLICATION.
If all originally triggering depths replicate, the claim is:
FALSIFIED — BOUNTY QUALIFIED
Challenge Registration
Before running a bounty-eligible PRIMARY block, the challenger must register:
- Claim ID and version
- challenger name or team
- a unique challenge ID
- a challenger-selected nonce
with Creation Unified.
The included make_worlds.py deterministically derives the 24 worlds from the registered challenge ID, nonce, block label, and claim version. This prevents world substitution after results are known without requiring an institutional preregistration system.
A REPLICATION block uses the same challenge ID with a new registered nonce and the COUNTERCLAIM_REPLICATION block label.
Interpretation Limits
This claim does not assert that:
- THD is universally optimal across AI systems;
- THD beats raw unscaled WDP;
- the result is a pure isolated phase-order causal effect;
- the result transfers automatically to CIFAR, ImageNet, transformers, or language models;
- this test proves THD generally.
It concerns only the frozen RHM system and registered contrast in this package.
Download Registered Test Package
ZIP SHA-256: 74e61d55228bb620dea1866fd91fe9eb824a9e858499ac377e75d43cb292077a