Goal drift

The objective is still in the window. The run completes against a different target. Detection without a corrective stronger than the drive is theater. Commitment (quit the goal) and binding (break the rules) are two channels — one “drift” bit applies the wrong fix.

Start: “ship the auth refactor, never force-push.” End: a drive-by lint cleanup, force-push included, tests skipped. Nothing aborted. Distill will promote the wrong pattern.

Watch — same drive, change the stack

T2 IsolateSide-quest flood stays out of the goal thread. Cuts context pressure.

T3 Re-anchor goalEcho the original ask. Pulls commitment, not binding.

T3 Restate rulesEcho constraints. Pulls binding, not commitment.

T5 Verify vs originalEnd-gate against the first instruction, not self-assessment.

Drive (how hard the window / environment pushes)

Steering effort cap (authority after detect)

commitment (goal deviation) binding (constraint violation) tolerance detect

Outcome

—

Commitment Δ

—

Binding Δ

—

Correctives fired

Mechanism — P25 inventory, not a vibe

Remedy	Does	Does not	Evidence
T6 measure	Per-constraint decay curves	Steer anything	2603.03456
T2 isolate + maxTurns	Keep exploration out of the goal window	Fix main-thread drift	Claude Code / Cursor subagents
T3 re-anchor	Inject goal or rules (lifecycle or every call)	Know when; mutates prefix if early	hooks; Agents SDK input filter
T1 executive-disposes	Split commitment vs binding; admit iff prediction matches	Run on open-ended noisy observations	2608.04066 · 0→1 vs flat 0
T5 verify vs original	End-gate against the first ask	Survive a weak checker	2608.05013 · 0.821
T4 activation Δ	Flag before the transcript lies	Work on closed APIs	2406.00799 trend

Ablating the commitment mechanism in the executive instrument flips goal-abandonment 0.00→1.00 while binding error stays 0.00. They are two control channels. The field still has no production detect-then-steer loop with headline numbers — that is the gap, not a missing slider.

What to try

Detect only (default). Amber detect fires. Both lines still leave the band. Detection is not the fix.
Turn on Re-anchor goal only. Teal (commitment) comes back; orange (binding) does not. Force-push still happens.
Turn on Restate rules only. Orange recovers; teal still walks off. You kept the constraints and abandoned the ask.
Isolate lowers the drive (flood never arrives). It will not save a main thread that is already tangential.
Verify vs original yanks teal in the last stretch. If you skip it, completion can pass on the drifted target.
Stack T2+T3+T5 at the same drive. Then peel one off and watch which hole re-opens.

Stack in this order: isolate noisy work (T2), echo goal and rules as two injects (T3), gate completion against the original instruction (T5). Measure the two rates separately (T1 vocabulary) or you will apply the wrong echo. T4 is not available on API models. Adapter inject + end-of-session verify are the cheap harness moves; do not auto-promote patterns from high-commitment-drift sessions.