Learning & Adaptation

Learning in Orrin is not one algorithm — it's several mechanisms, each learning independently from the reward signal, some immediately and some with delay. The UI's Learning room surfaces the result as before→after→because diffs.

Immediate learning

• Function selector — the main contextual bandit that picks the next cognitive function; updates from immediate reward each cycle.
• depth_bandit (UCB1) — learns how many draft→critique→revise rounds the inner loop should run.
• thinking_depth — chooses shallow vs. deep reasoning chains for goal pursuit.

Delayed learning (credit that can't be scored at action time)

• Evaluator (brain/eval/) — rewards a past decision when a memory it tagged is retrieved within ~50 cycles, or its goal closes within ~200. This is how "storing the right memory" or "setting up a good goal" eventually pays.
• Demand-expectations — learns which actions actually relieve which demands and routes that prediction error back into the control signals.

flowchart LR
    A[Decision] --> O[Outcome later]
    O --> E[Evaluator: memory used? goal closed?]
    E --> CR[Credit back to the original decision]
    CR --> B[Bandit / signal update]

Value authority: learning steers what runs and what's committed to

Learned outcome value doesn't just tune the bandit's internals — it has explicit authority over both halves of behavior:

• Selection — brain/think/think_utils/selection/score_actions.py adds each action's learned reward EMA as an additive term in candidate scoring (with a cap on the exploration group), so an action whose realized reward has collapsed stops winning on priors alone.
• Commitment — brain/cognition/planning/commitment_value.py scores which goal holds the driver slot: base tier/priority, plus a credited-effect value EMA, minus staleness (driver cycles without a credited effect) and avoidance-streak penalties. A committed-but-unacted goal loses rank and rotates out; once released, its penalties decay so it can recover when it's actionable again. The survival-tier floor is preserved — the learned adjustment can cross priority ranks but never a tier boundary. Per-goal signals persist in brain/data/commitment_signals.json.

Grounded reward

Reward is denominated in durable outward effects, not internal churn, via the effect ledger — so learning pushes toward producing things, not just thinking about them. Quality is gated by the Quality Standard.

Optional self-shaping (OpenAI-only)

brain/cognition/finetuning/finetune_pipeline.py filters traces with outcome ≥ 0.65, submits a fine-tune job, and repoints model_config.json on completion so generation drifts toward what has worked. Symbolic-only mode never touches it.

Code pointers

• brain/think/think_utils/select_function.py, brain/think/depth_bandit.py
• brain/eval/ — delayed-learning daemons
• See also Action Selection and Bandit