dropout-decay / docs /plan.md
Mandeep Sidhu
Make research artifacts self contained
618af58

Cross-Regime Hypothesis Testing Plan

Date started: 2026-05-30

This is the standing protocol for testing the dropout-pressure hypothesis across regimes. Use this file before launching new experiments so formula changes are backtested against existing results first, instead of repeatedly rerunning expensive training.

For the detailed explanation of how coefficients are derived and how the formula is tested, see formula_coefficient_methodology.md.

Current operating decision: static coefficient backtests are internal gates; final evidence should be streaming multi-seed validation reports per regime. When a new regime or formula variant appears, fit it against existing saved results first, then decide whether new MPS experiments are actually needed.

Research Hypothesis

For a fixed training regime, the useful dropout rate is governed by pressure from model size, available unique data, and cumulative sampled training tokens. As a stream grows, this pressure changes, so a formula-derived dropout schedule can track the moving useful regularization region better than a hand-picked static dropout.

The current candidate formula family is:

p_t = clamp(p_min, p_max,
            A * log10(P / U_t)
          + B * log10(C_t / U_t)
          + D * log10(P / U_t) * log10(C_t / U_t)
          + C0)

The simpler first-order ablation is:

p_t = clamp(p_min, p_max,
            A * log10(P / U_t)
          + B * log10(C_t / U_t)
          + C0)

Where:

Symbol Meaning
P model parameter count
U_t unique tokens available at stream stage t
C_t cumulative sampled training tokens consumed by the optimizer by stage t
p_t active dropout rate at stage t
A model/data pressure coefficient
B sampled-token pressure coefficient
D interaction coefficient between model pressure and sampled-token pressure
C0 regime offset

Regime Definition

A regime is the full experimental environment in which coefficients are assumed to stay valid:

architecture family
+ tokenizer
+ corpus family
+ optimizer and learning-rate protocol
+ dropout placement and semantics
+ streaming protocol
+ evaluation distribution

Within a regime, P, U_t, and C_t are formula inputs. Changing those values should not require new coefficients. Refit coefficients when the corpus, tokenizer, architecture class, optimizer recipe, dropout semantics, streaming protocol, or validation distribution changes.

Non-Negotiable Rules

  1. Use MPS only for torch training experiments. If MPS is unavailable, stop and report it.
  2. Before launching any new MPS training, backtest the current formula family on all relevant existing saved results.
  3. Do not rerun a regime merely because the formula changed. Refit and backtest offline first.
  4. Treat coefficient fitting and streaming validation as different claims: coefficient fitting estimates useful static dropout; streaming validation tests whether those estimates form a good path-dependent schedule.
  5. Keep exploratory one-seed results separate from paper-grade multi-seed results.

Backtest-First Workflow

Run this workflow whenever a formula family changes or a new regime is added.

Step 1: Freeze Candidate Formula Families

Define the exact formula families being tested before looking at new training results:

Name Formula Purpose
base_abc A*x + B*y + C0 first-order pressure law ablation
interaction A*x + B*y + D*x*y + C0 current main candidate
optional higher-order quadratic or corpus terms only if simpler forms fail

Where:

x = log10(P / U_t)
y = log10(C_t / U_t)

Step 2: Inventory Existing Results

Before training anything new, enumerate saved runs by regime and decide which ones can be used for offline fitting or validation.

For each result source, record:

Field Required detail
regime name short stable label
run path directory containing summary.csv, metrics.jsonl, or equivalent
model specs model names and parameter counts
token prefixes unique-token limits used
sampled tokens steps * batch size * block size, or equivalent
dropout grid rates tested
seeds seed count
target extraction grid best, quadratic optimum, or boundary-marked optimum
quality flags bracketed, boundary optimum, flat curve, noisy curve

Step 3: Fit Within Each Regime

For every regime separately:

  1. Extract the observed static optimum for each (P, U, C) cell.
  2. Fit base_abc.
  3. Fit interaction.
  4. Optionally fit a higher-order variant only if the first two fail.
  5. Report coefficient values, RMSE, MAE, and residual direction.

Use boundary optima carefully. Keep them in the report, but downweight or flag them if the static dropout curve is not bracketed.

Step 4: Validate Without New Training

For every regime with enough cells, run:

Validation Meaning
leave-model-out fit on some model sizes, test held-out model size
leave-prefix-out fit on some unique-token prefixes, test held-out prefix
leave-source-out fit on one run source, test another run source
cross-regime transfer fit on one regime, test another regime

Expected result:

within-regime fit should be good;
cross-regime raw coefficient transfer may be weaker;
formula structure should still explain why coefficients differ.

Step 5: Backtest Streaming Runs Already on Disk

For existing streaming runs, do not refit on the streaming outcome first. Instead:

  1. Generate schedule values from the frozen coefficients.
  2. Compare them to the tested decay/static conditions already present.
  3. Report stage-wise and final-loss deltas versus the best static baseline.
  4. Mark whether the formula schedule wins, ties, or loses.

This separates two questions:

Can the formula estimate static useful dropout?
Can the static estimate be used directly as a streaming schedule?

Decision Gates

Coefficient Gate

Promote a formula family for a regime only if it satisfies:

Criterion Target
within-regime MAE preferably under 0.05 dropout
leave-model MAE preferably under 0.05 dropout
leave-prefix MAE preferably under 0.05 dropout
residual bias no systematic over/under prediction across P/U or C/U
interpretability coefficients have a defensible pressure-law explanation

If base_abc fails but interaction passes, present base_abc as the first-order law and interaction as the necessary second-order correction.

Streaming Gate

Run paper-grade streaming only after the coefficient gate passes.

A decay schedule passes strongly if:

mean final validation loss beats the best static baseline across seeds
and the win appears in most paired seed comparisons.

A decay schedule passes weakly if:

it ties the best hand-picked static dropout while avoiding bad static choices
across stream stages.

It fails if:

it loses to a simple static baseline in most seeds
or improves early loss only by sacrificing final loss.

Failure Handling

Use this decision tree before adding new experiments:

Formula changed or new regime added
        |
        v
Backtest on all existing saved results
        |
        v
Within-regime static fit passes?
        |
   no -> inspect target extraction, boundary cells, feature family
        |
   yes
        v
Held-out static validation passes?
        |
   no -> test interaction or limited higher-order correction offline
        |
   yes
        v
Existing streaming backtest passes or ties?
        |
   no -> consider streaming-specific transform before new training
        |
   yes
        v
Launch narrowed multi-seed streaming validation

If a pass condition fails, do not immediately launch a larger sweep. First decide whether the failure is due to:

Failure type Response
unbracketed static optimum add only the missing dropout-side points
flat/noisy curve increase seeds or eval batches before changing formula
bad held-out prefix add pressure interaction or revise C/U treatment
bad held-out model inspect parameter-count scaling and architecture invariance
streaming loses despite static fit fit a static-to-streaming transform, then backtest

Standard Experiment Order

Use this order for every regime.

  1. Backtest current formula family on all existing data.
  2. Fit coefficients within each existing regime.
  3. Produce cross-regime coefficient and error table.
  4. Decide whether the formula family is stable enough.
  5. If stable, run narrowed multi-seed streaming in the best current regime.
  6. If still stable, add a new regime with minimal calibration.
  7. Immediately backtest the new regime against all previous regimes.
  8. Only then run expensive streaming validation in the new regime.

New Regime Script Runbook

Use this exact command sequence for any new regime. Replace placeholders such as <regime>, <CORPUS_OR_PARQUET_PATH>, <MODEL_SPEC>, and <TIMESTAMP> with absolute choices before launching. Do not skip from calibration directly to streaming: the schedule must be frozen from the coefficient fit before the streaming run starts.

This section is intentionally verbose. Its purpose is to make future regimes auditable: an external reader should be able to tell what each script did, what file it produced, and which decision gate came next.

New Regime Step 0: MPS Preflight

Run this before any torch training command:

.venv/bin/python -c "import torch; print({'mps_built': torch.backends.mps.is_built(), 'mps_available': torch.backends.mps.is_available(), 'cuda_available': torch.cuda.is_available()}); raise SystemExit(0 if torch.backends.mps.is_available() else 1)"

What this does:

Check Meaning
mps_built PyTorch was compiled with Apple MPS support
mps_available this machine can actually run MPS now
cuda_available should not be used for this project

Decision rule:

if mps_available is false: stop and report
if cuda_available is true: still do not use CUDA

Also check for duplicate experiment processes before launching a long run. This is not part of the coefficient method, but it prevents corrupt timing/resource comparisons.

New Regime Step 1: Static Dropout Calibration Screen

Run:

.venv/bin/python scripts/run_experiments.py \
  --mode screen_static \
  --corpus <CORPUS_OR_PARQUET_PATH> \
  --text-column <TEXT_COLUMN_IF_PARQUET> \
  --cache-dir .cache/dropout_decay_<regime> \
  --output-dir runs/<regime>_static_screen \
  --models <M1=layersxheadsxdim> <M2=layersxheadsxdim> <M3=layersxheadsxdim> \
  --seeds 1 2 \
  --token-limits <U1> <U2> <U3> <U4> \
  --dropout-rates 0 0.02 0.04 0.06 0.08 0.10 0.14 0.18 0.20 0.26 0.30 \
  --steps <STATIC_STEPS> \
  --batch-size <BATCH> \
  --block-size <BLOCK> \
  --eval-batches <EVAL_BATCHES> \
  --train-eval-batches <TRAIN_EVAL_BATCHES> \
  --trace-eval-batches <TRACE_EVAL_BATCHES> \
  --vocab-size <VOCAB_SIZE> \
  --val-tokens <VAL_TOKENS> \
  --lr <LR> \
  --weight-decay <WEIGHT_DECAY> \
  --grad-clip 1.0 \
  --screen-early-stop

What this script run does:

scripts/run_experiments.py --mode screen_static trains a grid of static dropout models. It does not test the final decay hypothesis. It estimates the best static dropout rate for each calibration cell:

cell = (model parameter count P, prefix/unique tokens U, sampled tokens C)

For each cell, the script evaluates a fixed dropout grid and writes the validation curve. The curve is used later to extract the target dropout p*.

Expected outputs under runs/<regime>_static_screen/screen_static/<TIMESTAMP>/:

File Use
metrics.jsonl per-run raw metrics; includes token limit, model, seed, losses, and tokens seen
model_selection.csv per-cell static dropout curve and selected best dropout
summary.csv / summary.json compact aggregate summary
trace.jsonl lower-frequency trace for diagnostics
RESULT_SUMMARY.md human-readable first-pass summary

Why this is needed:

The coefficient formula is not fitted from streaming outcomes. It is fitted from static dropout optima. This separation is essential: calibration estimates where useful regularization sits; streaming validation tests whether following that moving estimate helps.

Recommended cheap calibration:

Dimension Default
models at least 3 model sizes if testing coefficient generality
token prefixes at least 4 prefixes
seeds 1-2 for calibration, 5 only for final streaming validation
dropout grid include low, middle, and high values so the optimum can be bracketed

Decision rule:

continue if most cells have a bracketed or near-bracketed optimum
refine if many best dropouts sit at the edge of the grid
stop and inspect if validation curves are flat/noisy enough that p* is unstable

New Regime Step 2: Fit First-Order Base Coefficients

Run:

.venv/bin/python scripts/fit_dropout_coefficients.py \
  --run-dirs runs/<regime>_static_screen/screen_static/<TIMESTAMP> \
  --output-dir runs/coefficient_calibration/<regime>_base \
  --target quad \
  --weighting heuristic \
  --feature-set base \
  --min-rate 0.0 \
  --max-rate 0.30

What this script run does:

scripts/fit_dropout_coefficients.py reads model_selection.csv and metrics.jsonl from the static screen. It converts each calibration cell into:

x = log10(P / U)
y = log10(C / U)
target = observed useful static dropout p*

With --feature-set base, it fits the first-order ablation:

p* ~= A*x + B*y + C0

With --target quad, the target p* is the local quadratic minimum around the best dropout grid point when the curve is bracketed. If the curve is not bracketed, the script falls back to the grid best and marks the cell as weaker evidence.

With --weighting heuristic, the fit downweights cells that are less reliable:

Cell condition Why it is weaker
boundary optimum true optimum may be outside the tested dropout grid
not bracketed local quadratic minimum is less trustworthy
very flat curve many dropout rates perform nearly the same
noisy best loss target dropout is less stable

Expected outputs under runs/coefficient_calibration/<regime>_base/:

File Use
coefficients.json fitted A, B, C0, metrics, and cross-validation scores
fit_diagnostics.md readable coefficient table, formula, fit metrics, and cell residuals
calibration_cells.csv one row per fitted cell with target, prediction, residual, and flags
next_dropout_suggestions.csv dropout rates to add if a cell needs refinement

Why this is needed:

The base model is the simplest pressure-law hypothesis. It is the ablation that tells reviewers whether the interaction term is actually necessary.

Decision rule:

if base MAE and held-out errors are already low: keep it as a strong ablation
if base has biased residuals or higher MAE: compare against interaction next

New Regime Step 3: Fit Interaction Coefficients

Run:

.venv/bin/python scripts/fit_dropout_coefficients.py \
  --run-dirs runs/<regime>_static_screen/screen_static/<TIMESTAMP> \
  --output-dir runs/coefficient_calibration/<regime>_interaction \
  --target quad \
  --weighting heuristic \
  --feature-set interaction \
  --min-rate 0.0 \
  --max-rate 0.30

What this script run does:

This repeats the same target extraction and weighted least-squares fitting, but uses the interaction pressure law:

p* ~= A*x + B*y + D*x*y + C0

The extra term D*x*y lets model/data pressure and sampled-token pressure interact. Empirically, this has mattered because dropout pressure is not always additive: the useful effect of seeing more cumulative sampled tokens can depend on how oversized the model is relative to the available unique data.

Expected outputs are the same as Step 2, but under:

runs/coefficient_calibration/<regime>_interaction/

Decision rule:

promote interaction if it lowers MAE/RMSE, improves leave-prefix/leave-model
validation, and does not create obvious residual bias

Do not promote the interaction form merely because it has more parameters. The paper needs the base-vs-interaction comparison to show that the extra term buys predictive accuracy, not just in-sample flexibility.

New Regime Step 4: Optional Static Refinement

Only run this if fit_diagnostics.md or next_dropout_suggestions.csv shows that important cells are weakly identified.

Run:

.venv/bin/python scripts/run_experiments.py \
  --mode screen_static \
  --resume-from runs/<regime>_static_screen/screen_static/<TIMESTAMP> \
  --use-cached-data \
  --cache-dir .cache/dropout_decay_<regime> \
  --output-dir runs/<regime>_static_refined \
  --models <ONLY_AFFECTED_MODELS> \
  --seeds 1 2 \
  --token-limits <ONLY_AFFECTED_PREFIXES> \
  --dropout-rates <SUGGESTED_RATES> \
  --steps <STATIC_STEPS> \
  --batch-size <BATCH> \
  --block-size <BLOCK> \
  --eval-batches <EVAL_BATCHES> \
  --train-eval-batches <TRAIN_EVAL_BATCHES> \
  --trace-eval-batches <TRACE_EVAL_BATCHES> \
  --vocab-size <VOCAB_SIZE> \
  --val-tokens <VAL_TOKENS> \
  --lr <LR> \
  --weight-decay <WEIGHT_DECAY> \
  --grad-clip 1.0

What this script run does:

This adds only missing static dropout points. It should not rerun the full grid. --resume-from lets the experiment skip rows already completed in the original static screen. --use-cached-data reuses the cached tokenizer and token arrays so refinement is measuring dropout/model behavior, not data preprocessing differences.

When to use it:

Trigger Refinement action
best dropout is at grid edge add rates beyond or near that edge if allowed
curve is too coarse near optimum add rates around the local best
static curve is flat add seeds or eval batches before changing the formula

After refinement, rerun Steps 2 and 3 with all relevant run dirs. At minimum, rerun the promoted feature family. If the paper will compare base versus interaction after refinement, rerun both.

.venv/bin/python scripts/fit_dropout_coefficients.py \
  --run-dirs \
    runs/<regime>_static_screen/screen_static/<TIMESTAMP> \
    runs/<regime>_static_refined/screen_static/<TIMESTAMP> \
  --output-dir runs/coefficient_calibration/<regime>_interaction_refined \
  --target quad \
  --weighting heuristic \
  --feature-set interaction \
  --min-rate 0.0 \
  --max-rate 0.30

Decision rule:

refinement is complete when the promoted coefficient fit has acceptable MAE,
held-out errors, and no obvious residual direction across P/U or C/U

New Regime Step 5: Generate Frozen Streaming Anchors

Run:

.venv/bin/python scripts/make_streaming_anchors.py \
  --coefficients-json <PROMOTED_COEFFICIENTS_JSON> \
  --name <regime>_interaction \
  --parameters <WINNER_MODEL_PARAM_COUNT> \
  --stream-token-caps <U1> <U2> <U3> <U4> <U5> \
  --stage-steps <STAGE_STEPS> \
  --batch-size <BATCH> \
  --block-size <BLOCK> \
  --min-rate 0.02 \
  --max-rate 0.65 \
  --precision 3

What this script run does:

scripts/make_streaming_anchors.py turns coefficients.json into the exact dropout schedule used by locked_stream. For each stream prefix, it computes:

P = chosen model parameter count
U_t = stream prefix tokens at stage t
C_t = cumulative sampled optimizer tokens through stage t
x_t = log10(P / U_t)
y_t = log10(C_t / U_t)
p_t = clamp(p_min, p_max, A*x_t + B*y_t + D*x_t*y_t + C0)

The script prints two things:

  1. a JSON diagnostic table with raw and clipped dropout values
  2. a final one-line anchor spec, for example:
<regime>_interaction:250000=0.300,500000=0.260,1000000=0.180,2000000=0.090,4000000=0.020

That final line is copied into the next command as --anchor-decays.

<PROMOTED_COEFFICIENTS_JSON> should point to the coefficient file selected by the coefficient gate. In a clean first pass, this is usually:

runs/coefficient_calibration/<regime>_interaction/coefficients.json

If optional refinement was needed and accepted, use the refined coefficient file instead:

runs/coefficient_calibration/<regime>_interaction_refined/coefficients.json

Decision rule:

freeze this anchor spec before streaming starts
do not edit the schedule after looking at streaming validation losses

If the anchor schedule looks pathological before training, such as all values clipping at p_min or p_max, inspect the coefficient fit and calibration cells before launching streaming.

New Regime Step 6: Five-Seed Locked Streaming Validation

Run:

.venv/bin/python scripts/run_experiments.py \
  --mode locked_stream \
  --use-cached-data \
  --cache-dir .cache/dropout_decay_<regime> \
  --output-dir runs/<regime>_<model>_streaming_validation_5seed \
  --models <WINNER_MODEL_NAME=layersxheadsxdim> \
  --seeds 1 2 3 4 5 \
  --stream-token-caps <U1> <U2> <U3> <U4> <U5> \
  --dropout-rates 0 0.02 0.04 0.06 0.08 0.10 0.14 0.18 0.20 0.26 0.30 \
  --anchor-decays <FROZEN_ANCHOR_SPEC_FROM_STEP_5> \
  --stage-steps <STAGE_STEPS> \
  --batch-size <BATCH> \
  --block-size <BLOCK> \
  --eval-batches <EVAL_BATCHES> \
  --train-eval-batches <TRAIN_EVAL_BATCHES> \
  --trace-eval-batches <TRACE_EVAL_BATCHES> \
  --log-every 250 \
  --vocab-size <VOCAB_SIZE> \
  --val-tokens <VAL_TOKENS> \
  --lr <LR> \
  --weight-decay <WEIGHT_DECAY> \
  --grad-clip 1.0

What this script run does:

locked_stream is the paper-grade test. It simulates a stream by increasing the available prefix tokens over stages. For each seed, it trains:

Condition type Meaning
static dropout baselines same dropout at every stream stage
anchor decay schedule frozen coefficient-derived dropout at each stream stage

The static baselines must be broad enough to make the comparison fair. The claim is not that decay beats weak static choices; the claim is that it can beat the best static dropout available in the tested grid.

Expected outputs under runs/<regime>_<model>_streaming_validation_5seed/locked_stream/<TIMESTAMP>/:

File Use
metrics.jsonl raw row-level results for each condition, seed, and prefix
summary.csv / summary.json aggregate condition and stage summaries
trace.jsonl progress traces for diagnostic plotting
config.json exact run configuration
RESULT_SUMMARY.md built-in readable summary

Primary evaluation metrics:

final validation loss at largest prefix
mean trajectory validation loss
stage-wise validation loss
paired seed delta versus the best static baseline
rank consistency across seeds

Decision rule:

strong pass: decay has best mean final loss and beats best static in most or all
paired seeds

weak pass: decay ties best static while avoiding bad early/late static choices

fail: decay loses to a simple static baseline in most paired seeds or wins early
only by sacrificing final loss

New Regime Step 7: Summarize Streaming Validation

Run:

.venv/bin/python scripts/summarize_streaming_multiseed.py \
  --metrics runs/<regime>_<model>_streaming_validation_5seed/locked_stream/<TIMESTAMP>/metrics.jsonl \
  --output-dir runs/<regime>_streaming_report/<model>_validation_5seed \
  --report docs/<regime>_streaming_report.md \
  --title "<Regime Name> Streaming Validation" \
  --date <YYYY-MM-DD> \
  --context "<regime/model/token/step description>" \
  --conditions <regime>_interaction static_dropout_0.1 static_dropout_0.08 static_dropout_0.06 static_dropout_0.14 static_dropout_0.18 static_dropout_0.2 static_dropout_0.04 static_dropout_0.02 static_dropout_0 static_dropout_0.26 static_dropout_0.3

What this script run does:

scripts/summarize_streaming_multiseed.py performs no training. It reads the saved metrics.jsonl file and writes standardized artifacts comparable across regimes.

Expected outputs:

File Use
docs/<regime>_streaming_report.md human-readable regime report for paper discussion
condition_summary.csv condition ranking by final validation loss
stage_summary.csv stage-wise trajectory table
paired_final_deltas.csv per-seed final-loss comparison against the best static baseline

The most important table is paired_final_deltas.csv. A mean win is useful, but paired seed wins are stronger because they reduce initialization-bias concerns.

Decision rule:

if the decay schedule wins 5/5 paired seeds: promote regime to strong evidence
if it wins 3-4/5: inspect effect size, variance, and trajectory tradeoff
if it wins 0-2/5: treat as a failed regime or schedule and do not bury it

New Regime Step 8: Smoke Check And Commit

Run:

.venv/bin/python -m py_compile \
  scripts/run_experiments.py \
  scripts/fit_dropout_coefficients.py \
  scripts/make_streaming_anchors.py \
  scripts/summarize_streaming_multiseed.py

What this script run does:

This is a code integrity check. It does not validate the scientific result, but it catches syntax or import errors in the scripts required to reproduce the regime.

After the smoke check, update this docs/plan.md ledger and commit:

docs/<regime>_streaming_report.md
runs/<regime>_streaming_report/<model>_validation_5seed/
runs/<regime>_<model>_streaming_validation_5seed/locked_stream/<TIMESTAMP>/
runs/coefficient_calibration/<regime>_interaction/

Do not commit temporary checkpoints or scratch corpora. Curated source corpora and token caches that are intentionally part of reproducibility should live under data/ and .cache/dropout_decay* with explicit provenance.

Current Regime Ledger

Regime Status Role
OpenWebText10K static/coefficient regime offline backtest complete retrospective support for interaction pressure law; do not rerun unless necessary
TinyStories static/coefficient regime active main coefficient evidence
TinyStories streaming regime 5-seed validation complete current main streaming evidence; interaction decay beats best static in 5/5 paired final-loss comparisons
OpenWebText10K streaming regime 5-seed clean validation complete OpenWebText10K interaction decay beats best static in 5/5 paired final-loss comparisons
WikiText-103 streaming regime 5-seed validation complete formula-derived L12 decay beats best static in 5/5 paired final-loss comparisons

Current Formula Status

The latest TinyStories analysis supports the interaction form more strongly than the first-order ABC form:

p_t = clamp(p_min, p_max,
            A*x_t + B*y_t + D*x_t*y_t + C0)

x_t = log10(P / U_t)
y_t = log10(C_t / U_t)

For the current TinyStories regime, the latest fitted coefficients are:

A  = -0.089261
B  = -0.129754
D  =  0.255069
C0 =  0.081525

Absolute TinyStories-regime formula:

p_t = clamp(p_min, p_max,
          -0.089261 * log10(P / U_t)
        - 0.129754 * log10(C_t / U_t)
        + 0.255069 * log10(P / U_t) * log10(C_t / U_t)
        + 0.081525)

Use these only as the current TinyStories-regime coefficients. They are not assumed to transfer numerically to the OpenWebText10K regime or any future corpus regime. The cross-regime claim we are testing is that the pressure-law structure transfers, while coefficients may be regime-specific.

Current Evidence Summary

Evidence item Current reading
OpenWebText10K static/coefficient regime backtest complete; interaction MAE 0.0148 on OpenWebText10K+5M versus base ABC MAE 0.0389
TinyStories static optima interaction form fits static dropout optima better than base ABC
TinyStories held-out prefix supports pressure dependence on unique tokens
TinyStories held-out model supports pressure dependence on model size
TinyStories streaming, 5 seeds interaction has best mean final loss; interaction beats best static in 5/5 paired final-loss comparisons
OpenWebText10K streaming, 5 seeds interaction decay has best mean final loss; top decay schedules beat best static in 5/5 paired comparisons
WikiText-103 streaming, 5 seeds formula-derived L12 decay has best mean final loss; beats best static in 5/5 paired comparisons
cross-regime raw coefficient transfer weaker than within-regime fit; supports regime-specific coefficients rather than universal numeric coefficients

Latest TinyStories 5-seed streaming final-loss table:

Condition Mean final 4M validation loss Std
interaction decay 2.5311 0.0213
smooth_low decay 2.5321 0.0203
baseabc decay 2.5357 0.0175
static 0.08 2.5444 0.0211
static 0.12 2.5477 0.0178
static 0.18 2.5644 0.0182

Paired final-loss result:

Decay schedule Paired wins vs best static
interaction 5/5
baseabc 5/5
smooth_low 4/5, with the one miss only +0.0003

The immediate risk is no longer seed count for TinyStories or OpenWebText10K. The main remaining risk is external validity beyond the three tested text regimes and robustness across controlled architecture or token-budget changes. The current defensible claim is:

Formula-derived dropout schedules track the moving useful dropout region and
avoid poor static dropout choices as stream scale changes.

The stronger claim:

Formula-derived dropout decay beats the best static dropout.

is supported at n=5 in TinyStories, OpenWebText10K, and WikiText-103. The strongest schedule in each of the three regimes beats the per-seed best static baseline in all five seeds.

Latest OpenWebText10K 5-seed streaming final-loss table:

Condition Mean final 4M validation loss Std
openwebtext10k_interaction decay 4.3981 0.0095
hold_30_then_decay 4.4052 0.0112
mild_30_to_08 4.4073 0.0085
fitted_l16_static_law 4.4124 0.0084
static 0.14 4.4455 0.0120
static 0.30 4.4668 0.0141
static 0.02 4.5358 0.0091
static 0.00 4.5943 0.0216

OpenWebText10K condition provenance:

Condition Provenance How to interpret it
openwebtext10k_interaction coefficient-derived interaction schedule main OpenWebText10K formula hypothesis test
hold_30_then_decay heuristic schedule-search ablation manually specified after exploratory single-seed OpenWebText10K schedule search; not generated from coefficients
mild_30_to_08 heuristic schedule-search ablation manually specified after exploratory single-seed OpenWebText10K schedule search; not generated from coefficients
fitted_l16_static_law older fitted/static-law schedule retained as a comparison to the earlier aggressive fitted path
static conditions fixed dropout baselines same dropout at every stream prefix

The heuristic OpenWebText10K schedules were chosen from failure analysis, not from the final coefficient formula. The older fitted_l16_static_law path started too high (0.60 -> 0.40 -> 0.30 -> 0.14 -> 0.02), while static dropout 0.30 looked useful early but worse at the final 4M-token stage and static dropout 0.14 was the strongest static endpoint. This motivated two manual ablations:

hold_30_then_decay = 0.30 -> 0.30 -> 0.20 -> 0.10 -> 0.02
mild_30_to_08      = 0.30 -> 0.24 -> 0.18 -> 0.12 -> 0.08

These ablations support the broader mechanism that stream-dependent dropout can matter, but they should not be used as evidence that the coefficient formula generated those exact schedules. The formula claim for OpenWebText10K should be based on openwebtext10k_interaction.

Paired final-loss result:

Decay schedule Paired wins vs best static
openwebtext10k_interaction 5/5
hold_30_then_decay 5/5
mild_30_to_08 5/5
fitted_l16_static_law 5/5

The best static baseline in the clean OpenWebText10K run is static dropout 0.14. The interaction schedule improves mean final validation loss by about 0.0473 and wins every paired seed comparison. This promotes OpenWebText10K from exploratory support to a second multi-seed streaming validation regime.

Latest WikiText-103 5-seed streaming final-loss table:

Condition Mean final 4M validation loss Std
wikitext103_formula_l12 decay 4.0808 0.0195
wikitext103_probe_blend decay 4.0961 0.0145
wikitext103_low_decay decay 4.1020 0.0166
static 0.10 4.1105 0.0188
static 0.08 4.1116 0.0186
static 0.06 4.1197 0.0082
static 0.14 4.1221 0.0155
static 0.18 4.1304 0.0130
static 0.04 4.1331 0.0227
static 0.20 4.1394 0.0167
static 0.02 4.1459 0.0165
static 0.26 4.1784 0.0145
static 0.00 4.1835 0.0165
static 0.30 4.1946 0.0141

Paired final-loss result:

Decay schedule Paired wins vs best static
wikitext103_formula_l12 5/5
wikitext103_probe_blend 4/5
wikitext103_low_decay 4/5

The best static baseline in the clean WikiText-103 run is static dropout 0.10 by mean final loss. The formula-derived L12 decay improves mean final validation loss by about 0.0297 and wins every paired seed comparison. This promotes WikiText-103 to a third multi-seed streaming validation regime.

Completed Static Backtest Gate

The first offline coefficient backtest is complete. It is retained as supporting artifact, not as final proof:

runs/coefficient_calibration/cross_regime_backtest/

Main reading:

the interaction pressure-law structure is supported in both the OpenWebText10K
regime and the current TinyStories regime, but coefficient values are
regime-specific.

Do not claim universal numeric coefficients. For final paper evidence, use streaming multi-seed reports for each regime.

Immediate Next Action

Reconcile the TinyStories, OpenWebText10K, and WikiText-103 five-seed streaming reports into the paper outline. The strongest current claim is now supported in three regimes: formula-derived or regime-fitted decay schedules beat the best static dropout baseline in paired five-seed final-loss comparisons.

The next empirical weakness is no longer a missing third text regime. The next useful strengthening step is to test robustness across a controlled architecture or token-budget change inside one established corpus regime, while preserving the same MPS-only, five-seed validation standard.

Next Training After Current Gate

No MPS training should launch until the three completed five-seed streaming reports are read together. Since a third held-out text regime is no longer the limiting issue, use the next run only for a narrowed robustness test:

completed: TinyStories 5-seed streaming report
completed: OpenWebText10K 5-seed clean streaming report
completed: WikiText-103 5-seed clean streaming report
next: reconcile three-regime evidence into the paper, then choose one narrowed
      robustness test
avoid: broad new sweep before three-regime report reconciliation

Evaluate with paired seed comparisons:

final 4M validation loss
mean trajectory validation loss
stage-wise validation loss
decay minus best-static delta per seed
rank consistency across seeds

Because TinyStories, OpenWebText10K, and WikiText-103 decays win across paired seeds, promote the cross-regime streaming claim to "supported in three regimes." Do not yet claim universal numeric coefficients. The current defensible paper-level claim is that the pressure-law structure and regime-specific fitting procedure can produce dropout schedules that beat the best static dropout baseline across multiple text regimes.

Latest streaming report:

docs/tinystories_streaming_report.md
docs/openwebtext10k_streaming_report.md
docs/wikitext103_streaming_report.md
runs/streaming_tinystories_multiseed_validation_l12/combined_5seed_summary/
runs/openwebtext10k_streaming_report/l16_updated_formula_clean_5seed/
runs/wikitext103_streaming_report/l12_validation_5seed/