worth-ledger

name	worth-ledger
description	Measure whether shipped changes were actually liked and what they were worth — elicit valuations from the user, record worth against tokens spent, and feed the results back into stopping thresholds and direction weights. Use at the end of a cycle, after a batch of changes lands, or when deciding where effort pays.

Worth Ledger

Unmeasured improvement is indistinguishable from motion. The worth ledger closes the loop: every cycle's output gets appraised — was it liked? what was it worth? — and the appraisal recalibrates everything upstream: the hunt's stopping threshold V, the interview's direction weights, and the honest answer to "is this project still paying for attention?"

Lives at .forward/worth.jsonl, one JSON object per appraised change or batch:

{"ts":"2026-07-09","change":"idempotent ingest retries","commits":["abc123"],"area":"ingest/retry","finds_addressed":2,"probes_spent":9,"verdict":"loved","worth_probes":40,"note":"user: 'this was the thing silently corrupting batches'"}

worth_probes denominates value in the same unit as spend (probes ≈ one focused attention-unit), so value-per-token is a dimensionless ratio you can compare across areas, cycles, and projects. area and finds_addressed link each appraisal back to the hunt ledger — without them the calibration below is not computable, only narratable. When only a verdict exists (no elicited magnitude), use the default anchors — disliked 0, shrug 1, fine 3, loved 9 — and record that anchors were used; an elicited worth overrides anchors and records its rationale in note.

Eliciting from the user

Appraisal is one light moment per cycle, batched — never a survey after every diff. Show the shipped changes compactly (one line each, with the observable effect, not the implementation), then ask two things:

• Verdict per change or batch: loved / fine / shrug / disliked. "Shrug" is a real answer and important to make cheap to give — a user who must praise everything gives you no signal.
• Worth, anchored concretely: "Which of these would you have paid a dedicated hour for? Which shouldn't I have bothered with?" Relative ranking beats absolute numbers; people rank reliably and estimate poorly.

If the user engages beyond the batch question, mine the conversation — a spontaneous "oh that one was great" is the highest-fidelity datum the ledger ever gets. If no human is available, self-appraise conservatively, mark "appraiser":"self", and let a later human verdict overwrite it.

Feeding back

• Stopping threshold: per area, V = Σ worth_probes / Σ finds_addressed over that area's records — probes of value per find, computable by replay with no interpretation. High-worth areas earn longer hunts (silence threshold 3·V grows); persistently shrug-rated areas get released sooner.
• Direction weights: sum worth by dimension (depth/breadth/reach/...). A dimension whose changes keep landing "loved" is where the next autonomous cycle defaults; a dimension the user keeps shrugging at is evidence the interview's excitement answer has gone stale — re-ask it.
• The project-level curve: value-per-probe across cycles is the health metric. Declining curve + statistical silence everywhere = the well is dry; say so and recommend stopping. A rising curve after a find-rich cycle is the hazard principle working — report it as license to keep spending.

Benchmarking honesty

The ledger is also the anti-slop instrument: it makes "the agent did a lot" and "the agent did a lot of good" distinguishable. Report both numbers, always — probes spent and worth returned — and never quote one without the denominator.