| # Phase 1: deterministic execution diagnostics |
|
|
| Frozen before first scientific run: `configs/phase1.json` (2026-09-16-v1). |
| CPU only, one numerical-library thread; no GPU and no package installation. |
| The existing project environment reuses base Conda site packages; each run |
| records actual package versions and source/config SHA256 hashes. |
|
|
| ## Questions and interpretation |
|
|
| 1. E1: Can a shrinking coherent model error coexist with increasing error of |
| the all-A/all-B execution? Evaluate actual KL and model joint TV, output TV, |
| and the *bound* on uniform head TV. No claim of optimal-policy performance. |
| 2. E2: How does an actual self-generated pilot change the posterior and output? |
| Compare a single alternative, the two-level manuscript ladder, and a single |
| alternative matched in initial drift. Same w/Delta does not match KL or |
| drift. Show actual model statistics and retain every grid point. No expected |
| direction of improvement is an acceptance test. |
| 3. E3: What does the best randomized adaptive policy attain at small n for a |
| fixed oracle? Compare to predefined schedules, keeping R fixed across K. |
| This does not optimize the outer KL neighborhood and is not an exact proof. |
|
|
| ## Numerical safeguards |
|
|
| - Pre-round conditional heads are frozen within each batch. No shared latent |
| mixture component is sampled during execution. |
| - E2 uses M ~ Bin(c, p0), not target prefixes. Its pilot set is selected uniformly |
| using private randomness and is not part of the delivered output; this |
| symmetry is essential for count TV to equal full TV. |
| - E3 must pass an independent literal-coordinate LP check and the n=1 private |
| randomization regression. Flow, objective, recovered-policy and dual residuals |
| are reported. Solver output remains a floating-point diagnostic. |
| - Main calculations are deterministic finite sums, not histogram TV estimates; |
| floating-point and normalization checks are reported, not mislabeled exact |
| arithmetic. There are no Monte Carlo confidence bands in this phase. |
| - No neural-language-model or practical-context-length claims follow. |
| - Resource caps: 120 s per LP solve, 250,000 state-actions per case; whole study |
| 45 min timeout, 8 GiB address-space cap. Stop/report failures rather than omit. |
| - This phase does not include the optional biased model or large-n stress tests. |
|
|
| ## Plotting |
|
|
| Visual style reference, accessed 2026-09-16: |
| https://github.com/ChenLiu-1996/figures4papers/ |
| Use blue #0F4D92, red #B64342, teal #42949E, violet #9A4D8E; |
| sans-serif labels, frameless legends, no upper/right spines, redundant line styles, |
| white background, vector PDF and 300-dpi PNG. Scientific calculations and |
| plotting implementations are independent of the reference repository. |
| Do not use decorative uncertainty bands for deterministic calculations. |
|
|
| ## Outputs and versioning |
|
|
| Each invocation creates a new results/run-* directory. Do not overwrite earlier |
| results. `status.json` gives actual completion/failure; source hashes identify |
| the implementation. Results are exploratory finite diagnostics, not a |
| preregistration or a claim that this is the final paper-ready experiment set. |
| The manuscript is not modified by this runner. |
|
|
