Experiment 023: Physical AI Supervisor Commit-Ledger Stress

Date: 2026-07-09 Status: Research complete Classification: Experimental runtime path validation

Goal

Stress the Action-Outcome supervisor’s supervisor-specific commit-ledger sink contract under repeated projection failures, fresh runtime objects, bounded batches, and repair passes.

Hypothesis

If the atomic commit ledger is sufficient for the current supervisor sink contract, then projection failures after commit should be repairable without duplicating decision or evidence rows, even when each repair pass is performed by a fresh runtime object.

Procedure

1. Create the default proposal, history, decision, evidence, commit, and ownership tables.
2. Configure the runtime with batch_limit=2, max_sink_errors_per_pass=1, and a bounded proposal query limit.
3. Insert one positive and one negative historical outcome.
4. Run six passes; before each pass, insert two new proposals.
5. On every odd pass:
   • replace the evidence projection with a malformed table,
   • run a fresh runtime and require the pass to fail on evidence projection repair,
   • restore the projection schema,
   • run another fresh runtime to repair committed state and process remaining proposals.
6. After every pass, verify commit, decision, and evidence row counts and proposal-id uniqueness.

Acceptance Criteria

Criterion	Required
Six bounded passes complete	pass
Three projection faults are injected	pass
Faulted passes fail closed through the sink error budget	pass
Repair passes converge after schema restoration	pass
Final commit rows equal total proposals	pass
Final decision rows equal total proposals	pass
Final evidence rows equal total proposals	pass
Commit, decision, and evidence proposal ids remain unique	pass

Command

./build/tests/zepto_tests \
  --gtest_filter='ActionOutcomeSqlAdapterTest.CommitLedgerStressRepairsProjectionFailuresAcrossRestarts' \
  --gtest_brief=1

Result

See docs/research/results/physical_ai_supervisor_commit_ledger_stress_023.md.

Summary:

• 6 passes completed.
• 12 proposals converged.
• 3 malformed-projection faults were injected and repaired.
• Commit, decision, and evidence tables each ended with 12 unique proposal ids.

Interpretation

Experiment 023 supports keeping the supervisor-specific commit-ledger sink contract for the current shadow runtime. The ledger is an effectively-once boundary for this bounded proposal/decision/evidence path, but it remains deliberately narrower than a generic multi-table SQL transaction primitive.

Next Product Step

Do not block the Action-Outcome supervisor shadow pilot on generic transaction work. Track generic multi-table transactions separately if broader SQL surfaces need them.