Results (all on held-out, freshly generated instances)

Per-query ensemble completeness (same instances, every possible query — the latents serve ALL queries from one fixed computation, including threads never named in any text):

• Value after step 1: : 0.993
• Value after step 2: : 0.987
• Value after step 3: : 0.993
• Value after step 4: : 0.980
• Value after step 5: : 0.713
• Value after step 6: : 0.773
• Value after step 7: : 0.680
• Value after step 8: : 0.667

Training data is an infinite procedural stream (every instance seen at most once); "train accuracy" evaluates instances bit-exactly replayed from the training seed, "test" a disjoint seed — the match shows the organism runs the algorithm rather than memorizing instances.

Probe findings (ridge linear probes on dot residual streams)

run-ahead staircase (dot p decodes s_1..s_{p+1}); retention confirmed (s_1/s_2 still 0.95-0.99 at dot 8). Deep states s_5-s_8 cap at ~0.6-0.8 everywhere, matching the behavioral per-step gradient: errors are deep-chain COMPUTATION errors (serial cap ~6-8 hops), not forgetting.

Worked example

See examples.md for a full operating transcript (the model sees the scenario, emits only dots in <think>, then the query is revealed and it answers from the dot activations) and the surface-CoT version of the same instance that the curriculum started from.

Files

• LoRA adapter (PEFT, r=32, attn+MLP target modules) + tokenizer + lt_cfg.json (full config)
• training_code/ — complete training/eval/probe code snapshot (see its README)
• plots/ — load-bearing eval, probe grid + logit lens, training curve

Intended use

Activation-oracle / interpretability research target: the reasoning trace exists ONLY in the dot activations (logit lens is near-dark), with deterministic ground truth from the task generator. Wired into the cds-jb/AVBench recog eval as suite latent_threads (token-exact rows; row_metadata.selected=False rows query threads with no textual trace anywhere in the transcript). Part of the latent-threads collection together with its four siblings.

Trained 2026-06-12 (wandb group lt2, MATS10-CS-JB/cot-oracle). Built on the bottleneck-mask methodology of the pointer-chase filler organism (cds-jb/qwen3-8b-pointer-chase-filler-cot).