Methodology

Jev supplies judgment. Code keeps control.

Nothing on this site asks a model to decide a priority. It asks six narrow questions about evidence, gets constrained probability outputs back, and then does ordinary arithmetic. This page is the whole method, including the parts that would be embarrassing if they were wrong.

1,000Accounts

6,000Jev judgments6 questions × 1,000 requests

9.37sWall clockconcurrency 100

$0.0775Cost1,845,032 input tokens

From evidence to action

The model reads the account. Your code makes the decision.

01 · Input
One account record
Usage, support, renewal dates and notes.
Removed before sending: id + planted archetype
02 · Jev
6 parallel questions
One request. The same evidence for every question.
- Needs attentionnoul
- Churn evidencenoul
- Expansion signalnoul
- Urgencyscore
- Primary issuechoice
- Suggested ownerchoice
03 · Typed outputs
Evidence, not a verdict
Noul
Probability from 0 to 1
Score
Urgency across an ordered rubric
Choice
Distribution over named options
04 · Your code
Rules → a shortlist
triage.ts applies thresholds and weights. No more inference.
Act now: urgency ≥ 3 and attention > 0.8
Act now / Review / No action, plus retention and sales lenses.

Ranking weights: attention 35 · churn 25 · expansion 15 · urgency 15 · renewal 10.

Human review: on flagged accounts, Choice confidence below 0.55 marks routing as uncertain.

Step one

One request per account, six questions

TypeSafe evaluates every question in a request against the same state in parallel, so six questions cost one round trip rather than six. The account record goes over with its planted archetype and its id stripped out — the model never sees which scenario the generator drew it from.

POST https://api.typesafe.ai/v1/systemone

{
  "model": "jev-latest",
  "state":  { …the account record… },
  "questions": { …the six below… }
}

{
  "type": "noul",
  "instructions": {
    "question": "Does this account contain evidence that a person at the vendor should act on it within the next two weeks?",
    "guidance": "Weigh `notes`, `usage30dDeltaPct`, `activeUsers30dDeltaPct`, `support`, `engagement` and `commercial` together. A single moved number is not evidence; a coherent story across several fields is."
  },
  "criteria": {
    "true": "There is concrete evidence of something worth a person's time: a deteriorating trend with a plausible cause, an unresolved failure, a change in the relationship, or a live commercial opening.",
    "false": "Nothing here is materially different from a normal, well-running account. Routine variation, seasonal dips, ordinary how-to tickets and a distant renewal are not evidence."
  }
}

{
  "type": "noul",
  "instructions": {
    "question": "Is there meaningful evidence that this customer's retention is at risk?",
    "guidance": "Judge the evidence for the customer reducing, not renewing, or replacing the product — not how large or important the account is."
  },
  "criteria": {
    "true": "Something in the record points at leaving: sustained disengagement, an unaddressed failure, a lost champion or changed sponsor, a competitor or consolidation review, payment or budget obstacles, or discount demands tied to a near renewal.",
    "false": "No retention risk beyond ordinary account noise. Falling usage on its own, with a stated benign cause and an engaged champion, is not churn evidence."
  }
}

{
  "type": "noul",
  "instructions": {
    "question": "Is there meaningful evidence of a near-term opportunity to sell this customer more?",
    "guidance": "'Near-term' means a conversation the vendor could open this quarter. Judge `commercial`, `notes`, and the direction of usage together."
  },
  "criteria": {
    "true": "The record shows actionable demand: additional seats requested, pricing or packaging asked for, a new team or region adopting, usage growing against a stated plan, or a sponsor referencing new budget.",
    "false": "No near-term buying signal. A healthy, stable account with no stated demand is not an expansion opportunity."
  }
}

{
  "type": "score",
  "instructions": {
    "question": "How soon does someone need to act on this account, given everything in the record?",
    "guidance": "Rate the timing only. A large opportunity and a large problem can both be urgent; a large account that is running fine is not."
  },
  "criteria": [
    "Nothing to do. The account is running normally and the next scheduled touchpoint is soon enough.",
    "Worth watching. One soft indicator has moved; note it and look again at the next regular review.",
    "Review within the next two weeks. Several indicators point the same way, or a renewal is approaching without a plan.",
    "Investigate now. There is a live problem or a live opportunity that decays if it is left until the next cycle.",
    "Immediate intervention. The account is failing, leaving, or ready to buy right now, and a delay of days changes the outcome."
  ]
}

{
  "type": "choice",
  "instructions": {
    "question": "If someone picks this account up, what is the single dominant theme they will be dealing with?",
    "guidance": "Choose the one theme that best characterises the account right now. Pick `none` when the account is simply running normally."
  },
  "criteria": {
    "adoption": "The dominant issue is that the customer is not using what they bought: low or falling usage, unconfigured workspaces, or stalled onboarding.",
    "support": "The dominant issue is unresolved product or service failure: live incidents, a backlog of tickets, or slow resolution.",
    "relationship": "The dominant issue is the human relationship: a lost champion, a changed executive sponsor, cancelled meetings, or a vendor review.",
    "commercial": "The dominant issue is money or contract mechanics other than the renewal itself: pricing pressure, discount demands, billing or payment problems.",
    "renewal": "The dominant issue is the renewal event itself. It is close and the outcome is not secured.",
    "expansion": "The dominant theme is a growth opportunity rather than a problem to fix.",
    "none": "There is no dominant issue. Nothing about this account needs characterising."
  }
}

{
  "type": "choice",
  "instructions": {
    "question": "Which role at the vendor is best placed to take the next action on this account?",
    "guidance": "Choose by the kind of work the next action requires, not by account size alone."
  },
  "criteria": {
    "CSM": "Customer Success Manager. Adoption, value realisation, relationship health, and the routine renewal motion.",
    "TAM": "Technical Account Manager. Implementation, configuration or integration work that needs hands-on technical ownership.",
    "AE": "Account Executive. A commercial conversation: additional seats, pricing, packaging, or negotiating terms.",
    "Support": "Support. An open technical failure that should be resolved through the support process rather than by a relationship owner.",
    "Leadership": "Vendor leadership. The account is valuable or severe enough that an executive needs to be personally involved.",
    "No action": "Nobody needs to pick this up. Leave it in the normal review cycle."
  }
}

Step two

The final decision never leaves the codebase

Bands, rankings and both lenses are arithmetic over the six answers. Changing a weight re-ranks 1,000 accounts instantly and costs nothing, because the evidence and the question meanings have not changed.

band =
  urgency >= 3 && needsAttention > 0.8
    ? "act-now"
  : needsAttention > 0.55
    || churnSignal    > 0.55
    || expansionSignal > 0.65
    ? "review"
  : "healthy"

priorityScore =
    needsAttention   * 35
  + churnSignal      * 25
  + expansionSignal  * 15
  + urgency / 4      * 15
  + renewalProximity * 10

// A Choice distribution flatter than 0.55 is not routed
// automatically — it is flagged for a person.

Step three

Did it actually find the right ones?

The generator knows which archetype produced each account, and Jev never sees it. That makes this table a real check rather than a demo. The generator planted 181 accounts that should need attention and 819 that should not. The triage surfaced 181, missing 1 planted problem and flagging 1 account that was meant to be quiet.

Same questions. Different signals.

Each bar shows the share of an archetype sent to each action band. Dots show mean model outputs from this synthetic run, not measured churn rates or evidence of calibration.

Healthy
n = 618
Routing outcomes
Act now
0
Review
0
No action
618
Mean signals
0%50%100%
Attention
10%
Churn
8%
Expansion
12%
Healthy noise
n = 201
Routing outcomes
Act now
0
Review
1
No action
200
Mean signals
0%50%100%
Attention
22%
Churn
15%
Expansion
14%
Adoption concern
n = 45
Routing outcomes
Act now
2
Review
43
No action
0
Mean signals
0%50%100%
Attention
78%
Churn
71%
Expansion
8%
Support escalation
n = 25
Routing outcomes
Act now
21
Review
4
No action
0
Mean signals
0%50%100%
Attention
87%
Churn
71%
Expansion
10%
Relationship risk
n = 21
Routing outcomes
Act now
12
Review
9
No action
0
Mean signals
0%50%100%
Attention
84%
Churn
92%
Expansion
7%
Renewal risk
n = 25
Routing outcomes
Act now
9
Review
16
No action
0
Mean signals
0%50%100%
Attention
77%
Churn
82%
Expansion
15%
Expansion opportunity
n = 34
Routing outcomes
Act now
1
Review
33
No action
0
Mean signals
0%50%100%
Attention
73%
Churn
6%
Expansion
96%
Ambiguous
n = 31
Routing outcomes
Act now
0
Review
30
No action
1
Mean signals
0%50%100%
Attention
69%
Churn
38%
Expansion
25%

Bars are normalized within each archetype; group sizes differ. Exact counts, signal means and urgency scores are in the table below.

Full audit data

Scroll the table horizontally to see every signal and urgency score.

Synthetic archetypes: planted accounts, routing outcomes and mean model outputs
Archetype planted	n	Act now	Review	No action	Attention	Churn	Expansion	Urgency
Healthy	618	—	—	618	10%	8%	12%	0.63
Healthy noise	201	—	1	200	22%	15%	14%	1.23
Adoption concern	45	2	43	—	78%	71%	8%	2.85
Support escalation	25	21	4	—	87%	71%	10%	3.17
Relationship risk	21	12	9	—	84%	92%	7%	3.12
Renewal risk	25	9	16	—	77%	82%	15%	3.02
Expansion opportunity	34	1	33	—	73%	6%	96%	2.99
Ambiguous	31	—	30	1	69%	38%	25%	2.49

Attention, churn and expansion are mean model outputs across each archetype; urgency is a mean rubric score, not a probability. The separation between a planted problem and planted noise is produced by the model reading the record, not by a rule keyed to the field the generator moved.

Caveats

What this does not show

The accounts are synthetic. The archetypes were written to be separable, which makes them easier than real CRM data, where the notes are shorter, staler and more contradictory.
The planted scenarios are generator intent, not observed outcomes. This is a synthetic evaluation set: nothing on this page shows whether a flagged account would actually have churned.
Typed output guarantees the interface, not the truth. Thresholds worth trusting have to be evaluated on your own data and your own consequences — these were chosen by hand, on this portfolio.
Cost is computed from the input tokens the API reported, at Jev's published $0.042 per million input tokens with output tokens free.

Jev supplies judgment. Code keeps control.

From evidence to action

One account record

6 parallel questions

Evidence, not a verdict

Rules → a shortlist

One request per account, six questions

The final decision never leaves the codebase

Did it actually find the right ones?

Same questions. Different signals.

Healthy

Healthy noise

Adoption concern

Support escalation

Relationship risk

Renewal risk

Expansion opportunity

Ambiguous

What this does not show