Can't go off-label
A pointer head scores only the options you list. There is no text to generate, so there is no way to answer with something else.
Decisions. Not text.
TL-Decision (Trillion Labs Decision) reads a message, a document or an agent log and answers typed questions: yes/no, pick one, or a score. Every allowed answer gets a calibrated probability. It never writes text, so it can't answer outside the options.
Get TL-Decision-4B
Open weights · kev / System One request format
One request.
Every question answered.
Send the state once with as many questions as you need. They share the text but can't read each other's answers.
{
"state": "Subject: Site down? ... none of our team can
log in, the dashboard shows a 502 ...",
"questions": {
"urgent": { "type": "noul",
"instructions": "Does the customer need a reply within the hour?" },
"team": { "type": "choice",
"instructions": "Which team should handle it?",
"criteria": { "outage": "service down",
"billing": "charges, refunds",
"feature": "requests, how-to" } },
"tone": { "type": "score",
"instructions": "How upset is the customer?",
"criteria": ["calm", "annoyed", "furious"] }
}
}urgentnoul
yes0.90
no0.10
teamchoice
outage0.86
billing0.03
feature0.11
tonescore
calm0.08
annoyed0.61
furious0.31
The request is the example from the TL-Decision-4B model card. The probabilities shown illustrate the response shape only; they are not a recorded model output.
Built to be
acted on.
Software can branch on these answers directly. Nothing to parse, nothing to retry.
Calibrated by default
Each checkpoint ships with a fitted temperature, so 0.9 means roughly 90%. Expected calibration error is 0.018 on transfer-v4.
A drop-in endpoint
Same request and response format as TypeSafe's Jev, through the open kev project. Point an existing client at a TL-Decision server.
From generalist
to decider.
Fine-tuning the Trida2.0-4B backbone for typed decisions, scored with kev's own harness.
hard-v1 · 1,083 questions
Long policies, trade-offs, multi-hop
0.867accuracy
The same backbone answering through a prompt scores 0.476.
transfer-v4 · 656 questions
Sources it never trained on
0.822accuracy
On par with Kev-4B (0.817), with lower calibration error (0.018 vs 0.042).
English suites
| Suite | TL-Decision-4B | Kev-4B | Trida2.0-4B, prompted |
|---|---|---|---|
| transfer-v4never-trained sources, 656 q | 0.822 | 0.817 | 0.692 |
| transfer-v9MMLU-Pro, evidence removed, 1,046 q | 0.784 | — | 0.653 |
| decision-v7 †kev's decision distribution, 1,184 q | 0.861 | 0.871 | 0.751 |
| hard-v1long policies, multi-hop, 1,083 q | 0.867 | — | 0.476 |
| devtools-v1code review, commit, safety, 1,074 q | 0.745 | — | 0.585 |
| documents-v1consumer complaints, 920 q | 0.891 | — | 0.775 |
| SemIfhand-authored, 144 q | 0.903 | — | 0.757 |
Accuracy on kev's development splits at TL-Decision's shipped temperature (T = 1.823). † Without Banking77 questions with more than 26 options. Kev-4B was scored with the same harness on the same items where shown.
Korean suites, test splits
| Suite | Trida2.1-4Bcoming soon | TL-Decision-4B | Kev-4B |
|---|---|---|---|
| KMMLU-Pro2,822 q | 0.466 | 0.418 | 0.417 |
| CLIcK1,995 q | 0.617 | 0.584 | 0.572 |
| KorNAT common knowledge6,007 q | 0.671 | 0.666 | 0.597 |
| KorNAT social values4,000 q | 0.447 | 0.485 | 0.537 |
| typed-decisions test2,000 q | 0.762 | 0.654 | 0.666 |
Trida2.1-4B (coming soon) has its own typed-decision head on a model that also writes text, scored at its fitted calibration temperature. Test items were screened for overlap with training data.
Read these numbers with care.
- English results are development splits; the locked test splits have not been read for this checkpoint.
- decision-v7 and hard-v1 training data overlap their development splits (26% and 36% of records under kev's rule). Kev trained on the same data, so the comparison holds, but both scores are flattered.
- SemIf has 144 questions, so differences under about 3 points are noise.
Serve it on
your machine.
TL-Decision-4B runs on the open kev server and benchmark harness. A single CUDA GPU is enough.
# kev is not on PyPI; install from source (Python 3.12/3.13, uv)
git clone https://github.com/jaredpalmer/kev.git && cd kev
uv sync --extra serve
uv run --extra serve python -m kev.serve \
--run trillionlabs/TL-Decision-4B --port 8009
# → POST /v1/systemone
uv run python -m kev.benchmark --run trillionlabs/TL-Decision-4B \
--suite evals/v4/transfer-v4curl -s localhost:8009/v1/systemone \
-H 'Content-Type: application/json' \
-d '{"state": "Refund still missing after 10 days.",
"questions": {
"team": {"type": "choice",
"instructions": "Which team should handle it?",
"criteria": {"billing": "charges, refunds",
"outage": "service down"}}}}'The calibration temperature is stored in head.pt and applied by default.
Set KEV_TEMPERATURE=1.0 for raw probabilities. TypeSafe's Python SDK works against the same endpoint.
Need text generation as well? Trida2.1-4B (coming soon) is a Trida model with typed decisions built in: one set of weights writes text and makes typed decisions.
A few details
What is a typed-decision model?
A model that returns a probability for each allowed answer instead of generating text. It is built for the decisions software makes all day: routing, triage, risk scoring, guardrails.
Is TL-Decision the same as Jev?
No. Jev is TypeSafe's hosted model. TL-Decision uses the same request and response format, as implemented by the open kev project, and is not affiliated with TypeSafe.
What is it built on?
Trillion Labs' Trida2.0-4B backbone (Qwen3.5-4B architecture: 32 layers, 24 Gated DeltaNet and 8 full-attention), fully fine-tuned, with a pointer head trained from scratch.
How does it read a request?
Each question becomes its own causal row: the state, the instructions, each option, then a decide token. The head scores each option's hidden state against the decide token, and a softmax over the options gives the probabilities. This is kev's encoding, byte for byte.