Raw 300k Checkpoint and Interaction Result

Status: immutable pretraining milestone and descriptive transcript result. This is not an SFT checkpoint, benchmark promotion, reasoning claim, or evidence that the model has a reliable hidden-thought mechanism.

Checkpoint custody

Two-H100 flagship job 686732 completed cleanly on evc34 at exactly 300,000 steps after 153,869 seconds. The final log line before completion was:

step 299990 loss 1.6554 gnorm 0.11 lr 0.0005 281,959 tok/s
[done] 300000 steps in 153869s

The log contains 134 printed skip lines. The final four consecutive gnorm skips at steps 299546--299549 recovered, and the run subsequently completed every remaining step. The log is 498,688 bytes with MD5 28b9a13d596f48c39ee1139c004a17e8 and SHA-256 f359671e256fea784c063747a9d76641384dad8762e4bfae5bf6177fa308669e.

The trainer's terminal artifact is intentionally model-only. It contains no optimizer state and must use a fresh-optimizer rewarmup if resumed. Newton preserves it as both ckpt_0300000.pt and best_step300000.model.pt; the numbered and best names are hard links to the same read-only preserved inode, not links to the writable ckpt_final.pt.

The local Mac copy at train/flagship_out/ckpt_0300000.pt is mode 0444 and matches Newton by byte count, MD5, and SHA-256.

Nominal update tokens are not unique corpus tokens. The run necessarily replayed the mounted corpus; the two quantities must not be conflated.

Fixed direct interaction

Raw 300k was run through the existing deterministic seven-case, five-turn manual_capability_probe_v1 protocol on Apple MPS with max_new=128. The protocol asks for an initial answer, independent review, use of a verified intermediate fact, compact-state emission, and compact-state reuse.

Evidence bindings:

Strict scores

Raw 300k therefore shows no strict aggregate improvement on this fixed protocol. It still cannot reliably follow answer-only formatting, correct a prior response, emit a valid compact state, or reuse one.

Descriptive transcript change

One fixed case improved qualitatively despite failing the strict parser. For the sequential state update, raw 300k generated:

14 + 9 = 23
23 * 3 = 69
69 - 20 = 49
49 is the integer.

The strict score is false because the answer protocol requires the final integer at the start of the response and the model continued generating after 49. This differs from raw 200k, which answered 41, and raw 260k, which omitted the multiply. However, raw 190k had already generated the same correct 23 -> 69 -> 49 trajectory before entering a loop. The mode therefore disappeared and reappeared rather than improving monotonically. It is evidence of one fragile visible multistep behavior on a fixed probe, not evidence of broad or latent reasoning or a 260k-to-300k capability transition.

The remaining cases stay structurally weak:

• multiplication still asserts 29 times 16 = 496;
• base conversion does not apply positional weights correctly;
• sort/deduplicate repeats the input and drifts into code-like boilerplate;
• string insertion degenerates into repeated formatting;
• the logic answer is correct but review/state reuse fail;
• Python generation does not produce a valid answer-only function.

No <think>-style hidden reasoning was recovered by this protocol. The raw pretraining checkpoint mostly emits continuation-style explanations and templates rather than controlled deliberation.

Decision

1. Preserve 300k as a durable raw-pretraining anchor.
2. Do not claim that 300k is broadly smarter than 260k from this probe.
3. Preserve the sequential-state transcript as a qualitative lead for future fresh confirmation, not as a tuned score.
4. Do not launch the planned 600k language-balanced continuation until the 25B FineWeb replacement and both language-source approval records exist and pass their hash-bound gates.
5. Continue the R12 mechanism program independently; raw next-token pretraining alone has not produced reliable state transport or self-correction.