Production and the Effect Ledger

The effect ledger (brain/agency/effect_ledger.py) is the denominator the reward function was missing. Before it, reward was denominated in internal events — and internal events are free and infinite — so the rational policy was to churn cognition forever and never produce anything. Reading paid exactly what making paid. The ledger fixes that by making durable outward effects the thing that pays.

What the ledger is

An append-only, content-addressed log of real outward effects, written at the moment of the effect. Duplicates earn nothing (novelty dedup), so re-emitting the same artifact can't farm reward. It lives in brain/agency/ — the action side — not in cognition.

What counts as an effect

• Symbolic productions — synthesized principles, crystallized skills, resolved experiments, established causal edges → symbolic_artifact, recorded via brain/symbolic/symbolic_effects.py.
• Goal handler artifacts — research memos, housekeeping reports → file_write, registered at the goals runner's DONE-step chokepoint from Step.artifacts.
• Delivered notes and replies — things a person actually received (through the expression membrane).
• Verified sandbox checks — produce_and_check outputs that passed → tool_run_effect.
• Tracked-work sections — durable sections of long-form work (brain/data/tracked_work/).

Artifact files are captured under brain/data/effect_artifacts/ (EFFECT_ARTIFACTS_DIR in brain/paths.py); the ledger stores content hashes.

How reward keys on it

• A credited effect pays production reward at record time (finalize_cycle).
• Goal closure and milestone checks ground on the ledger (has_qualifying_effect) — a goal cannot claim completion without a qualifying recorded effect.
• Making actions pay per attempt, so the per-cycle gradient never favors pure intake over production, even before an attempt succeeds.
• Quality is enforced upstream by the quality standard: passing the bar is what makes an artifact creditable, so template-stamped or low-effort output doesn't pay.

Fail-able aspirations

The complement to paying for production is being allowed to fail: long-horizon aspirations carry guards instead of living forever, so a goal that stops producing is failed and replaced rather than absorbing attention indefinitely.

Code pointers

• brain/agency/effect_ledger.py — the ledger (design: ORRIN_PRODUCTION_REWARD_PLAN_2026-06-18)
• brain/symbolic/symbolic_effects.py — symbolic production recording
• goals/runner.py — the DONE-step artifact chokepoint
• brain/data/effect_artifacts/, brain/data/tracked_work/ — runtime artifact trees