Composite scoring
Composite scoring is TypeSafe’s name for: several atomic Scores (or mixed primitives), then a formula you own. Example from the primitives page: ticket priority from (1) bug severity, (2) customer frustration, (3) how much the report gives an engineer to work with. Normalize, weight, combine in Python.
This page is for builders crossing from curious to production-minded Jev use. jev.pro is an independent unofficial field guide — not affiliated with TypeSafe AI. Cover intents; do not clone rival IA. Confirm facts on docs.typesafe.ai.
When the blend disagrees with the on-call team, change weights and replay — do not hope a monolithic “rate this ticket 1–10” prompt rebalances.
Why not one Score?
A single rubric that secretly mixes three dimensions produces low confidence and arguments about what a “3” meant. Atomic levels stay inspectable. You can log each dimension.
Keep arithmetic out of the model
Do not ask Jev to multiply weights. jev-1.13 is not a calculator. Sum in code. Do not interpret a fractional Score as a dollar amount or a count.
See the official composite scoring page for their worked request.
Worked sketch: vendor RFP pack
Procurement scores a vendor packet on three independent dimensions. Official pattern page uses resume screening — cite it; this sketch is RFP review, not hiring.
{
"packet": "SOC2 Type II letter (draft). 99.5% uptime 'target'. REST + webhook. No SSO mentioned."
}| Id | Type | Rubric (4 levels each — Score 2–10) |
|---|---|---|
security_clarity |
Score | none → letter-only → attested controls |
sla_clarity |
Score | marketing target → contractual SLO |
integration_fit |
Score | CSV only → REST → SSO + webhook |
# normalize by (n_levels - 1); weights live in code
s = answers["security_clarity"].score / 3
a = answers["sla_clarity"].score / 3
i = answers["integration_fit"].score / 3
if min(answers[k].confidence for k in
("security_clarity", "sla_clarity", "integration_fit")) < FLOOR:
return review("low_dimension_confidence")
blend = 0.50 * s + 0.25 * a + 0.25 * i # change without a new promptNouls can feed the formula (e.g. mentions_sso) but they are not weights. Confidence exists on each Score; abort the blend if any dimension is below floor — confidence-gated routing.
Pin jev-1.13.0. Vendor $0.042 / M input (claim). Jaggedness: no math inside Jev; 99.5% is text. Do not generate the award letter.
When not to use
- One dimension is the only thing reviewers care about — ask that Score alone.
- You need Jev to multiply or rank-normalize a table of 200 vendors in one thought. Loop in code; filter state.
- You hide weights in
instructionsso nobody can change them without a prompt review.
See atomic questions and Score. When security and SLA scores disagree with the procurement committee, change the 0.50 / 0.25 / 0.25 constants and replay the same packets. That is the point of composite scoring: inspectable dimensions, movable weights, no new rubric paragraph. Schema-safe blends are still not award decisions.
See also
FAQ
Can weights be Nouls?
Weights are yours. Nouls can be inputs to a formula, not the formula.
Should I hide weights in instructions?
No. Reviewers cannot tune a paragraph the way they tune a constant.
Official worked request?
On TypeSafe’s pattern page (resume screening) — we will not clone their full example.
Score level cap?
2–10 levels per Score. Choice is a different primitive (≤255 options).
Disclaimer
This is an independent unofficial site and is not affiliated with TypeSafe AI; official documentation is available at https://docs.typesafe.ai. Never treat jev.pro as TypeSafe official documentation. We do not sell, issue, or proxy API keys.