Approval gate
Stop before you act
Manual autonomy routes every side-effecting call through a gate. It shows the payload verbatim rather than a paraphrase, and a rejection carries a reason back into the run so the next attempt is better instead of identical.
Merge the experiment branch into main and tag the promoted policy.
bash scripts/promote.sh --experiment exp-clip-vs-potential --into main --tag policy/ppo-clip@v5
workspace/human/approval-gate.tsxWrite the finding from this run into the workspace notes.
## Reward shaping Clipping at 1.0 with a -0.01 step cost raised the goal-reach rate from 11% to 94%.
risk: low — same gate, quieter chromeAsk user
Stop to ask
The other half of human-in-the-loop: the run pauses on a question set rather than guessing. Answered exchanges collapse into a one-line record so the reply that follows still has a visible cause.
Which shaping approach should the promoted variant use?
single → multi → text, steppedworkspace/chat/panel/question-answered.tsxPreference labelling
Human signal, with the numbers attached
Two candidates side by side, each carrying its own measured facts. A preference given without seeing return, steps and cost is a preference about prose — so the facts sit inside the card being judged.
Reach the goal tile from a random start within 60 steps, without farming the proximity term.
07 / 40workspace/human/preference-compare.tsxTakeover & capture
Taking the controls
The seam between the policy and the person. Who was driving for which steps is recorded, because that is what separates usable demonstration data from a contaminated rollout — and the recorder arms before it records so a fumbled first move never becomes the lesson.
Click through both stateslatency > 120ms turns the readout amberState: idle
idle → armed → recording → captured