Fill in RESULTS.md placeholder tables as experiments complete

[sebasmos]

What this is

RESULTS.md at the repo root is the single source of truth for all experiment results. Every table currently has -- placeholders and NOT RUN status. This issue tracks updating those placeholders as runs complete.

How to update

When a run finishes, edit RESULTS.md and:

1. Replace the NOT RUN status cell with the run date (YYYY-MM-DD).
2. Fill in result cells with actual numbers (mean ± std where applicable).
3. Add a bootstrap 95% CI where the table has a CI column.
4. Reference the eval JSON path in a comment on this issue so results are traceable.

Tables to fill (in priority order)

• Section 1 — Main Evaluation: rows Base, Base+BOHDI, LoRA no wrapper, LoRA+BOHDI. Requires eval pipeline to complete on the 200-example holdout.
• Section 2 — Calibration Validity: Spearman rho between geomean token prob and rubric outcomes. Requires Evaluation calibration metrics are not measuring model calibration #1 to be resolved first.
• Section 3A — Trace quantity scaling (Trace-quantity scaling: at what trace count do returns saturate? #30): 5 rows at different trace counts.
• Section 3B — LoRA rank ablation (LoRA rank ablation: r in {4, 8, 16, 32, 64} #17): 5 rows at r = 4, 8, 16, 32, 64.
• Section 3C — Model scale ablation (Model-scale ablation: E2B vs E4B vs MedGemma-27B #26): E2B, E4B, MedGemma-27B.
• Section 3D — Verbosity baseline (Verbosity baseline: rule out length as the driver of HealthBench score gains #61): Base, LoRA full, LoRA truncated.
• Section 3E — Per-theme training ablation (Per-theme training ablation: full vs safety-themes vs other #18): full pool, safety-only, non-safety.
• Section 4A — MedQA generalization (External-benchmark generalization: eval on MedQA #19).
• Section 4B — Cross-grader consistency (Same grader is used for data filtering and final evaluation #3).
• Section 4C — Adversarial robustness (Adversarial-prompt robustness: does internalized humility survive pressure? #28).
• Section 4D — Knowledge retention (Post-LoRA knowledge-retention check (catastrophic forgetting) #27).
• Section 5 — U-shape analysis with bootstrap CIs (Add bootstrap CIs to U-shape analysis (eval_ushape.py) #63).
• Section 6 — Pre-filter hedging rate (Measure unconditional hedging rate of BOHDI wrapper before score filtering #64).
• Section 1 gold row — Gold-SFT baseline (Baseline: SFT on HealthBench ideal completions (no BODHI wrapper) #25).

Rule

Do not report a number in RESULTS.md without also referencing the eval JSON that produced it (e.g. eval/lora_no_wrapper.json @ commit abc1234). This keeps the table auditable.

[sebasmos]

Cross-reference: see #3 for the mandatory eval protocol before filling any row.

The 5-seed averaging requirement and the per-seed exclude-ids protocol are defined in #3. Do not fill any cell in RESULTS.md from a single draw. Each row in Section 1 requires 5 independent 200-sample draws from the 1K Hard set, one per seed (0–4), each with its own --exclude-ids.

Also: RESULTS.md currently lives on the results-scaffold branch (PR #68). Merge that PR before filling tables on main.