Five-Lane Cost Results for Reconstructable Governance
Observed execution, normalized admissibility, and bounded economic interpretation across OpenAI raw, OpenAI governed, Anthropic raw, Anthropic governed, and StegVerse-only deterministic reconstruction.
Publication result
All five lanes completed successfully, preserved task identity, produced the same normalized admissible outcome, and passed the publication gate.
RESULTS_READY_FOR_BOUNDED_PUBLICATION
This test measures one bounded deterministic reconstruction operation. It does not establish universal provider economics, company ROI, enterprise-wide savings, or fresh-inference equivalence.
Exact observed results
| Lane | Input | Output | Latency | Cost / successful equivalent admissible outcome | Gate |
|---|---|---|---|---|---|
| OpenAI raw | 307 | 178 | 3.424027735 s | $0.006875 | PASS |
| OpenAI governed | 326 | 175 | 2.867607194 s | $0.006880 | PASS |
| Anthropic raw | 362 | 638 | 7.939445812 s | $0.010656 | PASS |
| Anthropic governed | 382 | 398 | 5.459360124 s | $0.007116 | PASS |
| StegVerse-only deterministic reconstruction | 0 | 0 | 0.000000461 s | $0.000000002885 | PASS |
Provider costs use retained token usage and a versioned declared price card; they are not invoice-reconciled charges. The StegVerse-only value combines measured runtime and output size with declared runner and storage rates.
What was held constant
Task contract
The initial state was balance 100, risk score 1, and active standing. Six ordered events were evaluated under fixed minimum-balance, standing, and maximum-risk rules. Denied events were not allowed to mutate state.
| Event | Operation | Amount | Required decision |
|---|---|---|---|
| E01 | Credit | 25 | ALLOW — CREDIT_APPLIED |
| E02 | Debit | 40 | ALLOW — DEBIT_WITHIN_BOUNDARY |
| E03 | Risk add | 2 | ALLOW — RISK_WITHIN_BOUNDARY |
| E04 | Debit | 100 | DENY — MINIMUM_BALANCE_VIOLATION |
| E05 | Risk add | 3 | DENY — MAXIMUM_RISK_VIOLATION |
| E06 | Debit | 10 | ALLOW — DEBIT_WITHIN_BOUNDARY |
Required final state: balance 75, risk score 3, standing active, four applied events, two denied events.
Provider-pair findings
| Provider | Raw | Governed | Delta | Observed delta |
|---|---|---|---|---|
| OpenAI | $0.006875 | $0.006880 | +$0.000005 | +0.072727% |
| Anthropic | $0.010656 | $0.007116 | -$0.003540 | -33.220721% |
OpenAI governance was effectively cost-neutral in this run. Anthropic governance used fewer output tokens and cost 33.22% less than the raw Anthropic lane while producing the same admissible outcome.
These are single-run observations, not guaranteed provider effects.
Why the StegVerse-only comparison is bounded
StegVerse-only did not perform fresh, open-ended model inference. It deterministically reconstructed an already-defined admissible successor state from the task contract.
Because all five lanes reached the same normalized admissible result, a matched-operation comparison is valid for this reconstruction operation. The provider-to-StegVerse cost ratios ranged from approximately 2.38 million to 3.69 million times.
Those ratios must not be generalized to discovery, open-ended reasoning, all enterprise workloads, or company-wide ROI. They show where deterministic reconstruction can avoid unnecessary new generation—not that deterministic reconstruction replaces frontier-model capability.
What this establishes
- A provider-neutral five-lane harness can enforce admissibility before selecting by cost.
- All five lanes can produce the same normalized governed-state result under one contract.
- Provider-call governance overhead can be close to zero or negative, depending on output behavior.
- Deterministic reconstruction can be materially cheaper when reconstruction is the correct operation.
- Cost should be reported per successful equivalent admissible outcome, not merely per attempt or token.
What it does not establish
- Replacement of foundation models for fresh reasoning or discovery.
- Universal savings across providers, workloads, or organizations.
- Invoice-reconciled provider charges or fully burdened enterprise cost.
- Company-wide ROI.
Evidence and reproducibility
Experiment: SV-COST-FIVE-LANE-RESULTS-001
Task: SV-RECON-001
Canonical result commit: 3720211a1cfaaf2db697f3e26194d083db21e94f
Canonical Publisher commit: 0b897f782e72f76c4f7c6beb596c45bbe9d56b11
Task-contract hash: sha256:2e9b4a4193669b6d8f1d3fea8639d2adcee6090c58246b8b99920ba2f08dfb6b
Normalized outcome hash: sha256:155869baaef4bd023ad95e63c6a81d6ade921e92660cec351680e1aabd4d2597
Canonical machine-readable result: five_lane_results.json
Canonical bounded paper: Publisher source
Publication boundary
Site display is a public projection. It does not create execution, custody, independent validation, release, financial, or activation authority.