$42 per billion input tokens. Output tokens are free for Jev 1.13.0. Text input requires preprocessing for images and scans.
Model and pricing documentationPre-alpha strategy / Cathodic protection review
Strong reasoning.
Precise questions.
Affordable scale.
Turn cathodic protection records and approved engineering knowledge into guidance a specialist can trace and challenge. Use a capable orchestrator to prepare the work, Jev for repeated judgments, and deterministic code for the calculations.
01 / Design around the evidence
Give each part a clear job.
The orchestrator earns its place by turning a complicated policy into a small, explicit decision. A domain expert approves that policy before it is reused across records.
- 01
Approved knowledge
Client procedures, field records, applicable references, and specialist-approved criteria. Preserve the original source and its version.
- 02
Capable orchestrator
Find the relevant passage, identify missing evidence, and propose precise questions with explicit options and boundary cases.
- 03
Jev decisions
Evaluate bounded text questions: yes/no probability, a choice, or a score level. Retain probabilities and confidence for review.
- 04
Code + review
Apply approved arithmetic and thresholds, link the reason to its source, and route uncertain or consequential findings to a specialist.
Example: code detects that a required measurement is missing. The orchestrator selects the applicable procedure. Jev checks whether a field note describes an unresolved isolation issue. The record enters the review queue with the missing field, cited procedure, and advisory judgment visible.
This is a proposed workflow. The company's original Pass/Fail, the calculated condition grade, and AI review urgency remain separate. An AI-generated interpretation does not become an approved engineering rule.
02 / Why Jev fits this role
Good instructions make the cheap model useful.
Jev's bounded answers keep a repeated judgment in a format the software can use. Precise criteria and relevant evidence reduce ambiguity; a strong orchestrator helps prepare both.
The proposed advantage is consistency and cost across many narrow decisions. The orchestrator handles complex interpretation and explanations; Jev handles the prepared judgment. Expert approval and tests determine whether that division works on this application.
TypeSafe's typed decision interface provides the primitives. The performance of this proposed infrastructure workflow still needs to be measured.
Public TypeSafe figures
Useful evidence, with its conditions attached.
TypeSafe also reports 193.6× faster for its System One workflows. These headline factors describe the vendor's tests.
TypeSafe homepageJev$0.0000810.114 seconds
LLMs$0.0138808.566 seconds
One displayed comparison, separate from the headline. Those costs imply approximately 171× cheaper and 75× faster for that example.
View the exampleTypeSafe's workflow evaluations average four example domains equally. Their reference labels come from GPT-6 Astra and Claude Fable 5.1 at high thinking; other models use provider defaults. Agreement with that model consensus is not an independently verified engineering outcome. Our pilot needs expert-labeled records.
03 / Public API pricing
Compare the model bill plainly.
Standard USD rates per million text tokens, verified October 11, 2026. The example column assumes 10,000 separate records, each with 2,000 input tokens and 500 billable output tokens.
| Model | Input / 1M | Output / 1M | Example / 10k | Proposed role |
|---|---|---|---|---|
| TypeSafe Jev 1.13.0 | $0.042 | Free | $0.84 | Narrow typed judgments |
| OpenAI GPT-6.1 Sol | $2.00 | $10.00 | $90.00 | Balanced orchestrator pilot |
| OpenAI GPT-6 Astra | $10.00 | $50.00 | $450.00 | Difficult reasoning and exception review |
| Claude Opus 5.5 | $4.00 | $20.00 | $180.00 | Primary Claude orchestrator candidate |
| Claude Fable 5 / 5.1 | $10.00 | $50.00 | $450.00 | Demanding, long-running orchestration |
| Claude Sonnet 5.5 | $2.00 | $10.00 | $90.00 | Lower-cost Claude candidate |
| DeepSeek V4 Pro 0813 | $1.32 | $3.96 | $46.20 | Cost-focused candidate; text inputs |
| DeepSeek V4.1 Flash | $0.30 | $1.20 | $12.00 | Lower-cost candidate to evaluate |
| Kimi K3 | $3.00 | $15.00 | $135.00 | Long-context orchestrator candidate |
| Kimi K2.6 | $0.95 | $4.00 | $39.00 | Lower-cost Kimi candidate |
| OpenAI GPT-5.3 Codex | $1.75 | $14.00 | $105.00 | Coding-specialized API option |
| OpenAI GPT-6 Luna | $0.10 | $0.50 | $4.50 | Cheap routine-task comparison |
| Claude Haiku 5.5 | $0.10 | $0.50 | $4.50 | Cheap routine-task comparison |
These are uncached standard rates. DeepSeek uses peak pricing here; off-peak input/output rates are half these values. Haiku 5.5's listed rate applies to inputs up to 100k tokens. GPT-6 models listed here have higher pricing above 272k input tokens. Caching, batch discounts, priority service, extra reasoning tokens, tools, OCR, hosting, review labor, and taxes can change the bill. Equal token counts are a budgeting convention; tokenization differs by provider. Every model name links to its official price source.
Start with GPT-6.1 Sol and Claude Opus 5.5.
Test both on the same approved policies and expert-labeled records. Escalate difficult interpretations to GPT-6 Astra or Claude Fable 5.1 where tests justify the added cost. Sonnet, DeepSeek, and Kimi are additional candidates; price alone does not establish fitness.
This shortlist is a design recommendation, not a measured winner. GPT-6.1 Sol · Claude selection guidance.
Codex is an agent environment; pick its API model.
A Codex-style orchestrator still needs an underlying model and a controlled tool workflow. A ChatGPT or Claude subscription is separate from an application's API token bill.
Jev also needs to compete against inexpensive general models such as Luna and Haiku on the actual task. Its low price is useful only if the complete workflow meets the agreed accuracy and review requirements.
04 / Change the assumptions
A workload budget you can inspect.
Prepare a reusable decision workflow, run routine records through Jev, and use the frontier model for exceptions. This calculator estimates token costs; it calls no AI APIs.
Illustrative token budget
Jev + selected frontier model
- Routine Jev input
- $0.84
- Frontier workflow preparation
- $1.60
- Frontier exceptions (500 records)
- $4.50
Alternative: frontier model on every record$90.00
These architectures perform different steps. The comparison assumes their results meet the same approved acceptance criteria; it is not a benchmark or promised saving.
Token costs only. Jev output is free. Ingestion, retrieval, scans, retries, infrastructure, security controls, engineering, and specialist review need separate budgets.
05 / Data control is part of the design
Decide the processing boundary first.
The infrastructure concept calls for local or dedicated private processing of sensitive customer data. A hosted AI API sends its input outside that private environment, even when the payload is reduced.
Keep restricted records inside the agreed boundary.
Use local ingestion, retrieval, deterministic rules, and an approved private inference deployment. Hosted Jev and frontier APIs are excluded for restricted content unless the customer authorizes that external path.
Private or dedicated vendor deployment availability, pricing, and contractual terms need confirmation. An on-premises Jev installation requires vendor confirmation.
Send only the permitted evidence.
Remove identifiers, GPS coordinates, and unrelated context where appropriate. Approve the vendor, data region, retention terms, and access controls before routing any customer information to an API.
TypeSafe says it does not train on customer inputs and offers enterprise zero data retention by arrangement. Those terms are not the same as local processing. Published enterprise terms.
06 / Turn the concept into a tested offer
A focused pilot before broader claims.
Choose one inspection workflow, approve its evidence and rules, and compare the complete system against specialist judgments.
- Define the task. Agree on what a finding means, its review route, and the permitted data boundary.
- Build an expert-labeled set. Include missing data, contradictions, boundary cases, and misleading document content.
- Compare complete workflows. Measure missed important findings, false alerts, unsupported explanations, review volume, end-to-end time, and actual cost.
- Set acceptance gates. Calibrate confidence thresholds on held-out records, document specialist approval, and recheck after model or policy changes.
The current 155-record CP first pass is advisory. It does not establish a validated commercial accuracy rate or autonomous engineering decision system.
07 / Deployment choice
Pre-alpha · Under evaluationChoose where the orchestrator runs.
A hosted frontier model offers access to strong reasoning without buying inference hardware. A self-hosted model puts the compute and data boundary under company control, with a larger upfront investment and ongoing operating responsibility.
Decide the direction during pre-alpha. Agree on the data boundary, response-time target, budget, and support owner before production integration or hardware procurement. The pilot will test that choice; changing it later may require migration and revalidation.
| Decision | Hosted frontier API | Self-hosted model |
|---|---|---|
| Upfront cost | Usually lower infrastructure spending. Application development, integration, and validation still need funding. | GPU compute, storage, networking, deployment, and validation. Hardware sizing follows the workload test. |
| Ongoing cost | Metered model usage plus application hosting, support, and any contracted capacity. | Power, cooling, administration, security updates, backups, and hardware replacement; licensing where applicable. |
| Data & control | Approved inputs go to the provider under its agreed terms. The provider operates the model infrastructure. | Greater control over data location, access, model versions, and retention. The company or its operator owns the security and maintenance work. |
| Speed & capacity | Depends on the chosen model, network, service limits, and workload. Rate limits or higher-capacity service may need planning. | Depends on model memory needs, available hardware, input size, and simultaneous users. Local hosting does not guarantee an instant answer. |
| Quality & fit | A strong candidate for complex interpretation and exceptions; CP performance still needs evaluation. | An eligible model must pass the same CP evaluation. Adequate hardware and a permissive license do not establish reasoning quality. |
Size the workload before the rack.
One suitable GPU workstation may be enough for an initial evaluation. Production could require more memory, multiple GPUs, or multiple servers. We have not selected a configuration.
A rack holds server chassis; it has no universal GPU capacity. The chassis, electrical supply, cooling, and network determine what it can support. Two towers also need an appropriate design to share work or provide failover.
Model memory requirements and the number of simultaneous field users matter more than GPU count alone. A large model fitting in memory does not prove it can respond quickly enough.
Capacity context: Dell documents an eight-GPU, 6U server. This illustrates how configurations differ; it is not the proposed $10,000 system.
Define what “fast” means.
The pre-alpha should distinguish confirmation that a field entry was saved from completion of its AI review. Staff should be able to see when analysis is still pending.
Set an agreed response-time target, then measure the complete path from field submission to a usable result at the expected number of simultaneous users. Check typical and slower-case results, connectivity interruptions, and recovery.
Vendor benchmarks show that concurrency and input size affect response time. They are useful sizing context, not a CP performance promise. NVIDIA performance context.
More control, with responsibility.
Self-hosting can keep inference and records inside an agreed environment and reduce external data transfers. That benefit depends on access controls, secure updates, monitoring, backups, and recovery being operated properly.
A local orchestrator calling hosted Jev is a hybrid deployment. Jev still receives the evidence sent to it. A fully private path needs approved private processing for every component that handles restricted data; Jev’s on-premises availability requires vendor confirmation.
Quick explanation: model versus harness
The model provides the reasoning. The harness coordinates the work around it: context, tools, permissions, and task progress.
Hermes Agent can fill that harness role, broadly like Claude Code or Codex. Running the harness on company hardware does not make its connected model local. Hermes can connect to either a hosted provider or a local model.
These are examples of the software layer, not a finalized implementation choice. Hermes documentation · Claude Code overview · Codex harness overview.
08 / Build on reviewed experience
Pre-alpha · Implementation planA “wiki” built from human-approved results.
If the platform moves into company use, a private database will preserve every analysis run and build a searchable library of cases reviewed by a company-authorized CP3 or CP4 specialist. That approved “wiki” gives the orchestrator relevant experience to consult on future work.
- 01
Record each run
Save the submitted evidence, generated results and grades, model and rule versions, and processing status. New results enter a review queue; incomplete or failed runs remain traceable.
- 02
Specialist review
A designated CP3 or CP4 reviews the evidence and records approval, corrections, rejection, or a request for more information, with their identity, reason, and review date.
- 03
Preserve the judgment
Approved cases enter the reference library. Corrections retain the original output and review history. Superseded or withdrawn decisions leave the current reference set.
- 04
Consult approved cases
The orchestrator retrieves relevant, current cases within the permitted customer scope and cites their review records. Each new result still follows the agreed human review process.
Recorded does not mean approved. The run log contains all attempts; the approved “wiki” contains only accepted reference cases. Retrieval supplies reviewed evidence to the model—it does not automatically retrain it or turn a past judgment into a universal rule.
The database can run locally or in an approved private hosting environment, with controlled access, backups, and retention. Client records stay outside this public website. Company reviewer permissions and the production retrieval connection will be established during implementation.
The strategy
Keep the evidence visible.
Make the decision reviewable.
A capable orchestrator and a low-cost decision model can support this goal when approved rules, sources, privacy boundaries, and human responsibility remain explicit.