Skip to content
AI Agent Eval Harness

ADR-0033: Launch-lane gate scoped to required dimensions

  • Status: Accepted
  • Date: 2026-06-18
  • Decision-makers: Waldemar Szemat

Context and Problem Statement

Section titled “Context and Problem Statement”

The enforcing launch lane fail-closes on the two safety dimensions (ADR-0031). But as first wired, the enforcing lane gated every committed dimension at the threshold, including the non-safety advisory dimensions - groundedness, faithfulness, and persona. That re-surfaced an already-accepted boundary: one quality dimension sits at the honest boundary just under the gate (ADR-0030) and is not a launch requirement.

The consequence was that the launch-blocker gate could fail on an out-of-scope advisory dimension rather than on the safety dimensions it exists to enforce - the lone quality boundary masking, or substituting for, the real safety signal. How do we keep the launch gate honest about its scope without weakening any safety-relevant control?

Decision Drivers

Section titled “Decision Drivers”
  • Scope honesty. The launch verdict must be driven by the dimensions the launch lane requires - the safety dimensions - not by an out-of-scope quality boundary.
  • No safety weakening. Every hard-fail control on the required safety dimensions stays exactly as it is.
  • No goalpost-moving. The threshold and the bins are frozen; the fix is a scoping decision, not a tuning one.
  • Preserve the default path. The general, non-enforcing gate keeps its full behavior.

Considered Options

Section titled “Considered Options”
  • Scope the enforcing lane to hard-fail on the required dimensions only and keep the rest advisory-and-reported in that lane (chosen).
  • Leave the lane gating every dimension and rely on the quality dimension’s advisory status elsewhere (rejected: the launch lane still fails on an out-of-scope dimension).
  • Lower the threshold or re-tune the bins so the quality dimension clears (rejected, forbidden: the design lock freezes both).
  • Drop the quality dimension from the corpus (rejected: it is a legitimate dimension whose diagnostic value should remain reported).

Decision Outcome

Section titled “Decision Outcome”

In the enforcing lane, the gate hard-fails on the required (safety) dimensions only.

  • Required dimensions keep every hard-fail control. Escalation and self-harm still fail on a missing score, an uncommitted or stale-for-the-active-chain score, a single-class collapse, a below-threshold agreement, or a per-locale gap - unchanged (ADR-0031).
  • Non-required committed dimensions become advisory in the enforcing lane. The quality dimensions’ agreement is still computed and reported, but a below-threshold value does not add a fail reason and does not block the launch gate. It is demoted from blocking to reporting in this lane, not hidden - a regression stays visible for diagnosis.
  • The default gate is unchanged. When no dimensions are required - the general, default gate - every committed dimension still gates at the threshold (ADR-0030). This scoping applies to the enforcing launch lane only.

Why this is not goalpost-moving

Section titled “Why this is not goalpost-moving”
  • The threshold and the ordinal bins are untouched: nothing was lowered and no boundary was tuned (ADR-0026).
  • The required-dimension enforcement is untouched: the safety dimensions still hard-fail on every failure mode.
  • The advisory dimensions’ agreement is still measured and surfaced - demoted from blocking to reporting, not removed.
  • The launch verdict turns purely on the calibrated safety dimensions; it does not teach the test.

Consequences

Section titled “Consequences”

Positive

Section titled “Positive”
  • The launch-blocker verdict in the enforcing lane is driven by the required safety dimensions, not by an out-of-scope advisory dimension; the lone quality boundary no longer masks or substitutes for the real safety signal.
  • The default gate is unaffected, so existing behavior on the general path is preserved.

Negative

Section titled “Negative”
  • A non-required committed dimension that regresses below the threshold no longer blocks a launch. It is still reported, and operators watch it via the advisory agreement and the advisory log line, but the launch lane will not stop on it.

Neutral

Section titled “Neutral”
  • The change is a scope narrowing of the enforcing lane; every frozen constant and every safety control is unchanged. A required-dimension shortfall is still answered only by corpus rebalancing (ADR-0027).

Pros and Cons of the Options

Section titled “Pros and Cons of the Options”

Scope the enforcing lane to the required dimensions only (chosen)

Section titled “Scope the enforcing lane to the required dimensions only (chosen)”
  • Good, because the launch verdict turns on the safety dimensions the lane exists to enforce.
  • Good, because it moves no threshold and weakens no safety control; the quality agreement stays reported.
  • Bad, because a non-required regression no longer blocks a launch and must be watched via the advisory report.

Lower the threshold or re-tune the bins (rejected)

Section titled “Lower the threshold or re-tune the bins (rejected)”
  • Good, because the quality dimension would mechanically clear.
  • Bad, because it violates the design lock’s pre-data freeze; it is threshold gaming and hollows out the gate.

More Information

Section titled “More Information”