Stop paying LLM prices for every decision
Sieve is a family of open-weight decision models. Give it text and a question with a fixed set of answers, and it returns a probability for every answer. No free text, nothing to parse, and a confidence you can audit. Run it in your own cloud, tune it on your own decisions, and own the result.
Sieve-31B-v3 scores 71.1 on the Jev Decision Index, tied for first among 237 models. Jev 1.14 scores 60.24.
Try a decision
Rear-end collision. Police report attached. Repair estimate $3,850. Policyholder for 9 years, no prior claims.
QuestionWhere should this claim go?
- Fast-track0.82
- Standard review0.15
- Investigate0.03
Confidence 0.73
Act: fast-track this claim
Illustrative example. Probabilities are not from a live model.
// 01 — how
One read, and a probability for every answer
An LLM writes an answer word by word, then your code parses it and retries when it drifts. Sieve reads the input once and scores every allowed answer in a single pass.
An LLM call for every decision
- Prompt and context
- Model writes text
- Parse, validate, retry
- Decision
Billed for tokens in and outWaits while the model writesNo calibrated confidence to audit
Sieve
- Input and question
- One forward pass
- Probability per answer
- Act, or send to a person
No text generatedAbout 50 ms on Sieve-9B-v2Every decision calibrated and logged
Where Sieve fits: the hybrid AI architecture
Keep your LLMs for the work that needs them. Let them reason and talk, let Sieve make the judgment calls, and let your software act. Assurance keeps the record.
Same contract as Jev
Sieve speaks the System One request format, so existing Jev code moves by changing the base URL. Rolling out November 2026.
# POST /v1/systemone { "model": "sieve-latest", "state": { "claim": "Rear-end collision, police report, estimate $3,850" }, "questions": { "route": { "type": "choice", "instructions": "Where should this claim go?", "criteria": { "fast_track": null, "standard": null, "investigate": null } } } } # response { "route": { "choice": "fast_track", "confidence": 0.73, "probabilities": { "fast_track": 0.82, "standard": 0.15, "investigate": 0.03 } } }
// 02 — results
Every number here is published
Every number here links to an evaluation run you can inspect on Hugging Face.
- 71.1Decision Index score for Sieve-31B-v3Tied for first of 237 models. Jev 1.14 scores 60.24. Entries within 0.9 points share a rank.View results
- 97.8%Agent tool selection on BFCL, Sieve-31BPicks the right tool for an agent nearly every time.Model card
- 50.5 msAnswer time for Sieve-9B-v22.1 times faster than Sieve-27B-v2, which answers in 106.1 ms, in about 20 GB of GPU memory.View results
- 0.5%Wrong answers among Sieve-4B's 95%-confidence callsA 3.1-point calibration error.View results
- 75.3Score on private tasks from domains the model has never seenSieve-31B-v3 scores 75.3 on new-domain tasks against 68.3 on private versions of skills it was trained for. It performs better on the unfamiliar half.Model card
Decision Index 0.4.1, board data of 28 October 2026. The Full score is 60% private same-skill tests and 40% private new-domain tasks; public benchmarks do not count. Leaderboard positions move as models are added.
// 03 — models
Twelve open decision models, from 0.8B to 31B
Pick for the job. Smaller models are faster and cheaper; larger ones handle harder reasoning. All of them answer in the same format.
| Model | Best at | Published result | Runs on |
|---|---|---|---|
| Sieve-31B-v3 | Hard reasoning and agent tool selection | 71.1 on the Decision Index, tied for first | One 96 GB GPU |
| Sieve-27B-v2 | Intent detection, claim verification, RAG hallucination checks | 59.89 on the Decision Index | One 80 GB GPU |
| Sieve-9B-v2 | The balance of accuracy, speed and cost | 47.00, answering in 50.5 ms | About 20 GB of GPU memory |
| Sieve-4B | Confidence scores you can trust | 3.1-point calibration error | 4.7B parameters |
| Sieve-2B-v3 | Lightweight intent detection and routing | 34.45, with 81.7 macro-F1 on CLINC150+OOS | 2.3B parameters |
See all Sieve models on Hugging Face. Eight of them are on the Decision Index board. Model cards credit every base model and open-source component.
// 04 — compare
Same kind of model, a different deal
Jev made decision models mainstream, and it is a strong hosted API. With Jev you rent one shared model over the internet. With Sieve you can run it in your own environment, tune it on your own decisions and own the result.
| Jev (TypeSafe) | Sieve | |
|---|---|---|
| Decision Index | 60.24 (Jev 1.14) | 71.1 (Sieve-31B-v3) |
| Where it runs | TypeSafe's hosted API and the gateways that resell it | Your VPC, on-premises or air-gapped; or hosted |
| Who owns the model | TypeSafe; the same weights serve every account | You: open weights and custom adapters you own |
| Tuned on your decisions | Not offered | Sieve Foundry trains on your decision logs |
| Your data | Sent to TypeSafe; zero retention on enterprise plans | Never leaves your network with Sieve Vault |
| Cloud catalogs | Not listed on AWS Bedrock, Microsoft Foundry or Google Vertex | AWS Marketplace listing in progress |
| Audit and model risk | A confidence score per answer | Confidence plus Assurance: drift, audit log, validation pack |
| Switching cost | Proprietary hosted service | Same request format: change the base URL |
Jev is a product of TypeSafe. Comparison based on TypeSafe documentation and public sources, October 2026.
// 05 — products
Five offerings, one decision stack you own
Each one stands alone. Most teams start with a Foundry model running on Vault, governed by Assurance, then add decisions from their industry canvas.
Sieve Open
Available on Hugging FaceOpen-weight decision models from 0.8B to 31B parameters, with a Jev-compatible API. Run them anywhere.
The pain it removes
- LLM bills that grow with every decision
- Waiting while a model writes an answer
- Free text that needs parsing and retries
Why teams want it
- A probability for every allowed answer, in one pass
- Weights you can inspect and run yourself
- Existing Jev code moves by changing the base URL
Outcomes you can measure
- Cost per 1,000 decisions against your current model
- p95 decision latency
- Retries and parse failures removed
Best for: Developers and data teams proving the cost and latency case before any contract.
Sieve Vault
Founding customers nowSieve as a licensed, versioned container for your VPC, on-premises or air-gapped. Calibration measured monthly, every decision logged.
The pain it removes
- InfoSec will not send customer data to a third-party AI API
- Data-residency and sovereignty rules
- Outages and price changes you don't control
Why teams want it
- No outbound calls; runs inside your boundary
- Signed images and an offline license
- A calibration commitment checked every month
Outcomes you can measure
- Customer data processed outside your network: zero
- Time from InfoSec review to approval
- Decisions served within your own SLOs
Best for: Banks, insurers, payers and public bodies whose data cannot leave their boundary.
Sieve Foundry
Taking founding customersA custom, calibrated decision model trained on your rubric and decision history, with retraining and drift monitoring. You own the adapter.
The pain it removes
- Generic models miss your policies and edge cases
- Prompt tuning stalls short of production accuracy
- You can't own a model you rent
Why teams want it
- Trained on your decision logs and outcomes
- Adapter weights you own and can move
- Retraining on a set cadence as policy changes
Outcomes you can measure
- Accuracy on your held-out cases against your current model
- Share of decisions automated at your confidence threshold
- Human review volume
Best for: Enterprises with their own rubrics and a history of decisions with known outcomes.
Sieve Assurance
Version 1, November 2026Calibrated confidence, drift monitoring and a per-decision audit log, with a validation report in the format model-risk teams review. Works with any decision model.
The pain it removes
- Model risk cannot sign off on free-text judgments
- No record of why a decision was made
- Drift goes unnoticed until an incident
Why teams want it
- A confidence on every decision; low confidence goes to a person
- A complete audit log you own
- An open conformance suite, plus a paid runtime monitor
Outcomes you can measure
- Calibration error, tracked monthly
- Decisions with a full audit record: 100%
- Time to answer an examiner's question
Best for: Any team running decision models in a regulated process, whichever model it uses.
Industry Decision Canvas
Financial services, Q1 2027Your industry's common decisions as prebuilt, typed APIs, each with a rubric, evaluation set and validation template. Financial services first.
The pain it removes
- Every team rebuilds the same decisions from scratch
- Months from idea to a validated model
- No shared standard across business lines
Why teams want it
- Start from decisions that already work
- Validation templates included
- Tune the ones that matter with Foundry
Outcomes you can measure
- Days to the first production decision
- Decisions live per quarter
- Validated contracts reused across business lines
Best for: Vault customers that want several decisions live in their first year.
// 06 — uses
Six decisions to start with
Pick one decision with history and a clear owner. These come up in almost every bank, insurer and health payer.
Claims triage
Insurancechoice
Fast-track, standard review or investigate, with a confidence for each.
AML alert escalation
Bankingyes or no
Escalate or close, with the probability behind the call.
Complaint routing
Banking and insurancechoice
The right team and the right regulatory tag the first time.
Prior-authorization review
Health payerschoice
Approve, pend or send to a clinician.
Collections next action
Lendingchoice
Call, text, hold or offer a plan, within policy.
Agent tool-call approval
Any AI agentyes or no
Let the agent act, or stop and ask a person.
// 07 — start
From one decision to production
Every step has an exit you control. The benchmark costs nothing, and the pilot fee is waived if we miss the criteria we agree on day one.
Benchmark
Weeks 0 to 2
Free
- One decision, up to 2,000 past cases
- Scorecard against your current model
- Exit: your call on a pilot
Pilot
Weeks 3 to 8
$40K, credited
- Up to five decisions
- In your cloud, or on redacted samples
- Exit: agreed criteria met, or no fee
Production
Weeks 9 to 16
License
- Vault in your environment
- A Foundry model on your logs
- Exit: model-risk sign-off, live
Scale
Quarter 2 onward
Expand
- Add decisions from your industry canvas
- Quarterly retraining
- Monthly drift and calibration reports
You bring
- A decision owner
- Up to 2,000 past cases with outcomes
- Your current model, as the baseline
- A cloud account, or redacted samples
- Model-risk and InfoSec contacts
You get
- A scorecard: accuracy, calibration, latency and cost per 1,000 decisions
- A calibrated model for your decision
- A per-decision audit log
- A draft validation report
- A go or no-go recommendation
Success criteria, agreed on day one
- Accuracy at or above your current model
- p95 latency under your target
- Calibration error under an agreed bound
- Cost per 1,000 decisions below today's
Bring us one decision
In ten business days we'll show you what it costs, how fast it runs and how sure it is on Sieve, side by side with what you use now.
- Decision benchmark
Free
One decision, scored on your past cases. Results in ten business days. - Paid pilot
$40K
Six weeks, up to five decisions. Credited in full toward Foundry or Vault, and waived if we miss the agreed criteria. - Founding customer terms
First five
For organizations that sign by 15 December 2026: the Assurance monitor free for a year, a two-year price lock and a seat on the product council.
// 08 — book
Book your benchmark
Tell us the decision you want to test. The first ten organizations get the benchmark free.
Prefer to write directly? hello@sthanika.ai