454–718 ms at the local provider boundary.
Recorded trial, not a benchmark ↗
OPENJEV + POLLARD-JEV / FIELD ENGINEERING NOTES
Small decisions.
Visible boundaries.
Local observations, bounded model choices, and an explicit budget. Here is what we ran, what we learned, and what remains simulated.
RECORDED TRIALSSeptember 26–27, 2026
Actual local inference
on synthetic observations
This is a static proof page. It does not call OpenJev, Gemini, or a worker. These records do not establish that a model service is running now.
Two unchanged frames reused local state.
One recorded dry sequence ↗
During these OpenJev trials.
Gemini roles remained fixtures.
THE SUPERVISORY PATH
Model proposals pass through local controls.
The local controller owns alarms, holding, and admissible actions. A model response is a proposal; it does not clear a hazard or command a motor.
- 01
Observe
Structured observations enter the controller. The recorded trials used synthetic inputs.
LOCAL FACTS - 02
Gate locally
Check change, freshness, and deadlines. Reuse unchanged valid state when the policy permits.
AVOID REPEATED WORK - 03
Ask OpenJev
A typed
ACTUAL LOCAL TRIALSpollard-jevrequest constrains the local NLI model to permitted choices. - 04
Validate again
Recheck the returned choice, evidence age, current allowlist, and deadline after inference.
POST-VALIDATION - 05
Supervise
Accept only a bounded supervisory choice. Uncertain or invalid output leaves local alarms intact.
NO ACTUATOR COMMAND - 06
Escalate if allowed
Optional Gemini interpretation passes separate request, token, and dollar gates. Its roles were fixtures here.
FIXTURE IN THESE RUNS
Supervisory choices
holdrequest_evidenceescalate_geminicontinue_monitoringChoices have no model-supplied motion parameters. A latched alarm or uncertain evidence removes passive continuation from the permitted choices.
01 / RECORDED MODEL EXECUTION
A real local model, a narrow trial.
A pinned OpenJev checkpoint ran through pollard-jev on an RTX 4090 Laptop GPU. Observations and supervisory actions were simulated. [E1]
| Case | Attempts | Recorded result |
|---|---|---|
| Warm-up [E2] | 1 | Dry facts accepted passive monitoring; 921 ms. |
| Dry monitoring [E3] | 1 | 718 ms. Three frames used one inference and two local state reuses. |
| Wet episode [E4] | 2 | 531 ms dry / 454 ms wet. The wet decision abstained; alarm stayed latched. |
| Conflicting observations [E5] | 1 | 485 ms. The model abstained; alarm stayed latched and review remained required. |
| Authenticated tunnel smoke check [E6] | 1 | A separate recorded request reached the local model through the tunnel. |
| Total recorded attempts | 6 | 4 mission decisions + 1 warm-up + 1 tunnel smoke check. |
What this establishes: local model execution, state reuse in the dry sequence, and conservative handling of abstentions. It does not establish field accuracy, autonomous physical operation, or a latency distribution.
TWO SEPARATE ACCOUNTING RESPONSIBILITIES
A token allowance is not a dollar ledger.
Workflow economy requires both visible model work and cumulative spending control.
Pollard recording and allocation
Record model work and allocate request/token allowances around the workflow. pollard-jev supplies typed decision and freshness contracts.
Starting a new loop or mission does not create a new shared spending allowance.
Durable dollar reservations
A separate SpendLedger reserves cumulative USD for each outbound billable attempt and retains unresolved reservations.
Pollard's model records do not automatically include every cloud or provider bill. A refused budget keeps local holding and alarms active.
02 / SEPARATE OFFLINE COMPARISON
Fewer Gemini calls did not mean less total work.
Eight missions without injected provider faults, using illustrative fixture accounting. Observation-quality faults and the deliberate blind spot remained. [E7]
| Modeled strategy | Gemini only | Hybrid |
|---|---|---|
| Gemini attempts | 21 | 18 |
| Jev attempts | 0 | 5 |
| Total attempts | 21 | 23 |
| Simulated tokens | 1,696 | 2,290 |
Three fewer Gemini attempts, two more total attempts, and 594 more simulated tokens. The different fixture price schedules do not demonstrate real cost savings. Detection improvement was not demonstrated.
03 / THE CURRENT BROWSER DEMO
80 → 3 is a modeled request comparison.
80Analyze every observation
→3Evidence-gated requests
For one completed default patrol with no injected faults, local gates filter 57 routine and 20 duplicate observations. Actual model calls are zero.
The token and dollar estimates use disclosed illustrative assumptions. This simulator runs deterministic Python in your browser; it does not run OpenJev, pollard-jev, Pollard, or Gemini inference.
This demonstration is separate from the recorded local model trials and the fixture strategy comparison above.
EVIDENCE IDENTIFIERS / WHOLE-FILE SHA-256
Traceable claims without publishing private endpoints.
These are the names and content hashes of the reviewed local records. Raw artifacts are not hosted here; private endpoints, service details, and credentials are omitted. Hashes identify the reviewed bytes, not an independent certification of their contents.
- E1jev-openjev-runtime.jsonRuntime and four mission decisions
9fc1f9d41b3bb0cb37b379d2a7f4f00109865cfab89f9a6aa1fb18701b068f60 - E2jev-openjev-warmup.jsonOne local warm-up attempt
e26f7ad5d865c9c9e01d6270f411b9804346b005daec5cd50f47e7735f568a11 - E3jev-openjev-dry_monitoring.jsonDry sequence and state reuse
85fd40a43e08e64dd458033ee46f50a81845405b4d883fe199861637ba4c5766 - E4jev-openjev-wet_episode.jsonDry acceptance, wet abstention, latched alarm
2a6e9541e104f6dea00708cc68104879f7f89fd2b1d96b2f2e4aca2eaba8322c - E5jev-openjev-conflicting.jsonConflicting evidence and required review
43bd4cfa1cd3e0064c41ed26c715f73f31a387ab8719d6f9de45d68973b75b1e - E6jev-openjev-ngrok-proof.jsonOne authenticated tunnel smoke check
ead87e1c56f3795a87e02831e03205d5f01e8f8a7d880e27fedfcae3dd81105f - E7jev-comparison.jsonOffline strategy fixture comparison
7218beaa403231fd3343cb09e361cc3a3e2f1baeda7aaddbc4659f9a4ac1f2e0
Browser simulation source hashes are available in the public build manifest. No energy or carbon savings have been measured.