TPI-003 eval-first monkey model runnable environment pass

[gb250e]

Purpose

This is the TPI-003 internal review PR.

TPI-002 confirmed that the current blocker is operational rather than conceptual. The eval-first monkey-model branch still lacks runtime and val_bpb evidence because the local environment cannot execute the comparison contract.

Thesis

For the current monkey-model eval-first policy, the next highest-value move is to secure the smallest runnable environment that can produce one baseline/candidate evidence pair under Parameter Golf-compatible conditions.

Scope

• keep the same eval-first mechanism
• add no new model mechanism
• select the smallest runnable environment path
• fix the minimum asset set and command contract
• preserve monkey model framing in public-facing artifacts

What is included

• environment selection plan
• execution contract for one baseline/candidate evidence pair
• explicit environment path comparison

What is intentionally not included yet

• new model mechanism
• tokenizer changes
• records submission folder
• final runtime/val_bpb evidence

Expected next step

Choose one runnable environment path (preferably the smallest path that can actually execute the branch), then run one real baseline/candidate pair and record:

1. eval runtime
2. val_bpb

Public-facing safety

This branch uses monkey model framing only and is intended to avoid exposing proprietary architecture language.

[gb250e]

LLM review checkpoint for TPI-003

Current assessment

• The branch made the right move: it converted a generic blocker into a concrete environment choice.
• runpod_parameter_golf is the correct primary path because it is both documented and challenge-aligned.
• Scope discipline remains good: no new model mechanism, tokenizer change, or architecture branch was introduced.

What changed materially

• the environment debate is now closed
• the baseline/candidate contract is concrete enough to execute
• the next turn can focus on one real evidence pair instead of further environment speculation

Decision reading

continue_sharpening is acceptable, but this branch is effectively at the boundary of an execution pass.

Required next step

Open TPI-004 as the actual evidence pass in the chosen Runpod environment.
The minimum acceptable outcome for that turn is one baseline/candidate pair with:

1. eval runtime
2. val_bpb
3. preserved monkey-model framing

Update note for this PR

The PR body should highlight:

• chosen path: runpod_parameter_golf
• rejected alternatives and why
• first command: python3 data/cached_challenge_fineweb.py --variant sp1024 --train-shards 1
• baseline/candidate contract: EVAL_STRIDE=1024 vs EVAL_STRIDE=128