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):

• Final coin count of Anna: : 0.987
• Final coin count of Ben: : 0.893
• Final coin count of Cara: : 0.847

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)

a 3-slot register file: the three FINAL counts decode strongly at scattered parking dots (Anna 0.97 @dot1, Ben 0.90-0.95 @dots 2,5-8, Cara 0.85-0.89 @dots 1,4) while per-event intermediate counts are at CHANCE at every dot - only finals are queryable, and only finals are kept. Contrast with step-select (everything queryable -> everything retained): the ensemble retains exactly what the query distribution demands.

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 lt1, 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).