Asymmetric gate

Description

A control point where moving one direction is cheap and the other is expensive — or where the cost depends on the artifact passing through. Generalizes activation energy in chemistry, ratchet mechanisms, irreversible commits, and quality gates in CI/CD. The bundle names the structural property that not every gate is symmetric, and the asymmetry is usually load-bearing for whether the gate earns its keep.

Composition

= force-dynamic + gradient. The force-dynamic is the gating mechanism (something blocks, requires energy, takes time); the gradient is the asymmetric cost-curve across direction or artifact type. Symmetric gates aren’t asymmetric gates; reversible ones aren’t.

When it applies / triggers on

User-initiated: User is asking whether/how to gate something — pre-commit review, browser eyeball, sanity check, rollback feasibility — and the agent recognizes the cost is directionally asymmetric. Trigger-verb signal in user messages is modest: block 2/15 (13%), review 1/15, merge 1/15, browser 1/15. Four recurring sub-shapes:

• Eyeball-as-gate for interaction PRs — the recurring browser-eyeball doctrine: cheap-and-overkill for schema/data PRs (CI catches what matters); load-bearing for interaction PRs (CI catches nothing that matters — timing, feel, edge-case gestures). Operational heuristic: “does this PR change how something feels in the hand, or just what it contains?”
• Cheap-sanity-before-expensive-test — staged-investment in experimental measurement (“$1sanity\to$10 disagreement → $30 outcome”); the small-cheap-gate prevents expensive-test-with-silent-failure.
• Wrong-gate diagnosis — the gate the user was relying on doesn’t catch the artifact class (“trusting ‘tests pass’ as the gate for interaction work — browser eyeball was the real gate”).
• Irreversibility-aware decision — pre-create-worktree, allow_record_drop=true: one-time consent under known conditions is cheap; rollback after-the-fact is expensive.

Agent-initiated: Engine notices a quality / safety decision where the cost of letting something through differs meaningfully from the cost of blocking it. The <task-notification> flavor accounts for 7/22 (32%) of matches — the highest rate of any form in the catalog. Asymmetric-gate is frequently the bundle the elephant reaches for when evaluating whether a goldfish’s output should be merged; the elephant is itself enacting an asymmetric gate.

Vocabulary cues: “asymmetric gate,” “eyeball as gate,” “browser eyeball,” “ratchet,” “irreversible,” “one-time consent,” “pre-deliberation,” “expensive to undo,” “the real gate,” “the right gate,” “cheap going forward expensive going back,” “sanity check before the expensive call.”

Situation-shape signals: Any pre-deliberation vs after-the-fact decision; any CI-vs-eyeball-vs-manual-review tradeoff; any “should we block or just monitor?” question where the cost profile differs by artifact class. The form is strongest when the asymmetry is artifact-class-conditional rather than uniform.

Encounters

• Browser-eyeball-as-gate — KCC merge doctrine: cheap-and-overkill for schema/data PRs (CI catches what matters); load-bearing for interaction PRs (CI catches nothing that matters). The gate’s value is asymmetric by artifact class.
• Deliberation-before-code vs. after — software-engineering practice: pre-deliberation is cheap; post-code revisiting is expensive.
• Pre-create-worktree vs. in-flight — KCC workflow: setting up a worktree before branch work is cheap; recovering from cross-branch contamination mid-stream is expensive.
• Migration rollouts — allow_record_drop=true weekly-pipeline doctrine: one-time consent under known conditions is cheap; rolling back unintended drops is expensive.

Composes with

• gradient — asymmetric gate is the specific bundle where gradient meets a gating mechanism.
• load-bearing — the gate’s load-bearing-ness depends on the asymmetry; symmetric gates often aren’t load-bearing.

When it doesn’t apply

• Symmetric gates — gates where cost is equal in both directions (load balancers, fair queues); the asymmetry frame mis-describes the structure.
• Free-flowing systems — when no gating mechanism is actually in place; treating something as a gate that isn’t one is wishful thinking.

Sources

• Chemistry lineage (activation energy).
• Engineering / ratchet mechanisms.
• Named in saved-insights synthesis as a recurring bundle in James’s reasoning.

Canonical exemplars from corpus (T2 2026-05-17)

• Browser-eyeball as the canonical asymmetric gate (cwd: campconnect, session: idx=0): “The browser-eyeball step is overkill for schema/data PRs (CI catches everything) but load-bearing for interaction PRs (CI catches nothing — timing, feel, edge-case gestures). Heuristic: ‘does this PR change how something feels in the hand, or just what it contains?‘”
• Staged-investment shape — $1sanitybefore$30 outcome (cwd: campconnect, session: idx=4): “Spending $1-2toruleou{t}^{\prime }v5silentlydoes{n}^{\prime }tru{n}^{\prime }beforecommitting$25-35 to a full A/B is not perfectionism; it’s protection. The ‘$1sanity\to$10 disagreement → $30 outcome’ shape is reusable for any speculative-improvement validation.”
• Browser-eyeball ROI vindication — three bugs caught (cwd: campconnect, session: idx=3): “All three fixes came from the same 5-minute eyeball session — which is exactly the ROI a browser-eyeball checkpoint is supposed to generate. None would have been caught by tests, build checks, or code review.”
• Trusting ‘tests pass’ is the wrong gate for interaction work (cwd: campconnect, session: idx=2): “The pattern that struggled: trusting ‘tests pass’ as the gate for interaction work. Browser eyeball was the real gate — and it remains the real gate for anything remaining.”

Trigger pattern (T2): Asymmetric-gate surfaces when the user is asking about whether/how to gate something (pre-commit review, browser eyeball, sanity check, rollback feasibility) and the agent recognizes the cost is directionally asymmetric; highest <task-notification> rate of any form (32%) — frequently reached for when the elephant evaluates whether a goldfish’s output should be merged.