grindability

name

grindability

description

The green-light triage for any long or autonomous work — ask "what's the oracle?" AND "is it grindable?" before committing to the shape. Use at the moment of green-lighting any overnight or unattended run, fan-out, "keep marching" arc, soak, migration walk, bulk verification, or agent-loop dispatch. Also triggers when sizing why an autonomous run underperformed. Companion to fan-out (the parallel shape).

grindability — the oracle's twin

Distilled from a digest of Brandon Sanderson's conversation with Dwarkesh Patel on AI and the future of math. Promoted to a skill so it fires at green-light time instead of waiting to be remembered.

Verifiable isn't the bar — grindable is. An oracle tells you whether an attempt succeeded; grindability is whether you can afford many attempts: cheap, deterministic, side-effect-free retries — a sandbox you can reset, a test you can rerun, a build you can throw away. Autonomy compounds exactly where both hold. Where each attempt touches the world (rate limits, live sites, sends, spends, one-shot state), no oracle quality makes long unattended grinding safe. The canonical contrast: a test suite in a resettable sandbox is grindable; a rate-limited live site is not — no matter how good the check.

The two questions, asked together

1. What's the oracle? The deterministic check that says pass/fail without you (build/test/lint/diff/exit-code).
2. Is it grindable? Can attempt N+1 run immediately after attempt N failed, at ~zero cost, with no residue? (Reset-able sandbox · replayable input · idempotent target · no external quota/party in the loop.)

The four quadrants → what each green-lights

	oracled	no oracle
grindable	★ Let it run — autonomous goal loops, overnight runs, fan-out, retry-until-green. This is where autonomy compounds.	Build the oracle first — the detector is the actual first task; grinding without one just launders guesses.
ungrindable	Human-cadence checkpoints: act in small verified steps, surface between them. The oracle helps judge each step, but retries are expensive/dirty, so no unattended loops. (Deploys, live-DB migrations, anything where a probe against the live system is the only proof.)	Don't automate. Do it by hand with eyes on, or reshape the task until it moves quadrants.

Raising grindability is usually the real first task

Most "ungrindable" work is one artifact away from grindable — and building that artifact beats grinding carefully:

• Live system → scratch container / disposable replica you can rebuild and replay against
• One-shot state → replay file / fixture / seeded world (e.g., a headless game driver replaying a recorded run)
• External API with quotas → recorded fixtures for the loop; the real call only at the final verify (the verify-on-the-real-path rule still applies to the last run)
• Prod-only reproduction → capture the failing input once, grind against the capture
• Costly model calls → grind the harness with a local/cheap model; the premium model only for the final pass

If no such artifact is buildable, that's the signal the work belongs on the user's cadence — which is a scheduling fact, not a failure.

The companion mechanism: entropy collapse

Long solo grinding degrades: an autoregressive model marinating in its own prior outputs loses diversity (each attempt more like the last — the same wrong idea, retried politely). The fix lives at the prompt level, not the sampler: fresh eyes per attempt (fan-out), or opposed mandates (one attempt told to attack what the previous defended). This is why council beats solo and why the fan-out skill's fresh-context-per-unit shape out-performs one tiring serial operator on grindable work.

Applying it (30 seconds at green-light)

1. Name the oracle. None? → building it is step 1 (or the work isn't autonomous yet).
2. Name the reset. "How do I cheaply undo/rerun attempt N?" No answer? → name the artifact that would create one, or drop to human-cadence checkpoints.
3. Pick the shape: grindable+oracled → autonomous loop / fan-out / overnight; add opposed mandates if attempts will iterate on each other.
4. Say the classification out loud in the green-light message — "grindable+oracled (sandbox X, oracle Y), letting it run" — so the shape is auditable.