An autonomous AI agent (LLM with tool access) was subjected to seven pairs of executions of the same operational task: one in conventional mode and another under D'Artagnan Method axioms. Resource consumption was measured in real time by the platform's billing system, with 21 screenshots as documentary evidence.
Main result: Savings of 57% to 78% in the 5 tests with favorable axiomatic results, averaging 62%. Test 5 produced an inverted result serving as a negative control — demonstrating that violating an axiom nullifies the advantage.
Additional discovery: More axioms don't burden — they save. Cost dropped from 346 (3 axioms, with error) to 160 (5 axioms, no error) on the same task.
| Test | Task | Axiomatic | Conventional | Savings | Axioms |
|---|---|---|---|---|---|
| 1 | Server Status | 248 | 661 | 62% | A, B, C |
| 2 | Guardian (Security) | 121 | 282 | 57% | A, B, C |
| 3 | MCA Engine | 80 | 363 | 78% | A, B, C |
| 4b | Subdomains | 88 | 203 | 57% | A, B, C |
| 4a | Subdomains (discarded) | 236 | 272 | 13% | A, B, C |
| 5 | Logs (axiom violation) | 346 | 204 | -70% | A, B, C |
| 6 | Logs (5 axioms) | 160 | 375 | 57% | A, B, C, D, E |
| Metric | Value |
|---|---|
| Average savings | 62% |
| Minimum savings | 57% |
| Maximum savings | 78% |
| Average axiomatic cost | 139 créditos |
| Average conventional cost | 377 créditos |
| Reduction factor | 2,7x |
1. Axiom without protocol doesn't change behavior. The agent knew the axioms for weeks and kept wasting. Change only occurred when axioms were converted into explicit operational constraints. Knowing is not applying. Applying is programming.
2. More axioms save instead of burdening. The cost of prior validation (~20 credits) is a fraction of the retry cost per error (~200 credits). Axiomatic safety is investment, not overhead.
3. Violation proves causality. If all tests showed savings, the alternative explanation would be "the axiomatic mode is simply shorter." Test 5 refutes this: same form, axiom violated, inverted result. Savings come from discipline, not form.
Bare-metal server, 89 days online, 4 vCPU. Platform: Manus Pro (110,000 credits/month). Measurement: billing panel screenshots (21 captures). Experiment conducted on Aug 06-07, 2026 (Days 86-87 of 90).