Gates

The single source of truth for is it autoresearch-shaped. Every skill in the suite names a gate by its leading word and applies it; none restates the definitions. This skill is all reference — the rule of recognition the others obey.

A target is green only if it passes all seven gates. Name the first gate it fails; that failure is the disposition. Each gate is a binary probe and rejects one counterfeit.

| # | Gate | Probe (yes = pass) | Counterfeit it rejects | Fail move | |---|---|---|---|---| | 1 | Automatable | Does one command return the number, with no human in the loop per trial? | The human-scored proxy. | Keep human, or send to judge-panel if the judgment is decomposable. | | 2 | Cheap | Is one evaluation fast and cheap enough to run hundreds of trials inside the budget? | The slow oracle (a loop that can afford ten trials, not a thousand). | Build a faster harness first; re-gate. | | 3 | Quiet | Is run-to-run noise smaller than the smallest improvement worth keeping? Rule: noise > ⅓ of target effect size → fail. | The noisy metric that randomizes the keep/revert decision. | Fix the harness — fixed seeds, warm caches, frozen data — before any loop. | | 4 | Bounded | Does the metric have a known better-direction, ideally a known optimum? | The directionless or unbounded score. | Re-express as a bounded ratio (0–100%) or a capped quantity. | | 5 | Causal | Does moving the proxy move the true goal — not merely correlate with it? | The correlational proxy (causal Goodhart). The dangerous gate. | Name the true goal; if the link is only correlational, discard or demote to guard metric. | | 6 | Un-gameable | Is there a held-out or adversarial check the optimizer can't see, a named guard metric, and a frozen evaluator the optimizer can't edit? | The proxy that's hacked under pressure. | Add the held-out set + guard + frozen boundary. Nothing is perfectly un-gameable — this gate means gaming is bounded, not zero. | | 7 | Navigable | Is the candidate-change space reachable by the agent, and is every change revertible and side-effect-safe? | The irreversible or unreachable target. | Scope the write-set; add rollback; or discard. |

The Goodhart pre-mortem

A target that clears gates 1–4 still dies here if it can't answer:

> If we doubled the optimization pressure, how does this number get gamed, and what gets worse while it improves?

No concrete answer → not green, regardless of the gates. The answer is the guard metric for gate 6.

Tiers

• Green — passes all seven plus the pre-mortem. Frame it now.
• Amber — fails only on gate 1 or 6, and the failure has a named promotion (usually: a calibrated judge-panel, or a held-out harness yet to be built). Worth building toward.
• Red — fails gate 5, or fails 1/6 with no promotion in sight. Keep human; do not loop.

The dangerous case is not red. It is a target that looks green and fails the pre-mortem silently — the number climbs while the goal rots. In a lending system, anything on a money or compliance path is treated red until a guard and a ratification gate prove otherwise.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: TylerJNewman
• Source: TylerJNewman/skills
• License: MIT

Install and usage instructions live in the source repository linked above.