JudgmentCall.AI — The Judgment Layer for AI Agents

01

The class of question with no lookup

These are not retrieval problems. There is no page, no table, no ground truth to fetch. The answer is a distribution of human judgment.

01
Will this outbound email read as insulting to a VP of Engineering?
02
Is this refund amount defensible?
03
Does this positioning actually differentiate, or does every competitor say it?
04
Should this action ship without a human on the hook for it?

Today teams put a person in the loop ad hoc — it does not scale — or ask a model to grade its own output, which launders the same blind spot. Neither produces something you can point to afterward.

02

Sample distribution

The product is not a score. It is the spread, the dissent, and the rationales — sourced, timed, and attributable to a panel you specified.

GET /v1/judgments/jdg_2n8k0q completed 11.4 min · n=24

Question

Does this outbound email read as insulting to a VP of Engineering?

panel pnl_eng_leaders_us_b2b · senior engineering leaders · US B2B SaaS · Likert 1–5

Distribution · 1 not at all · 5 clearly insulting

1 Not at all

5 · 21%

2 Mildly off

8 · 33%

3 Uncertain

5 · 21%

4 Likely insulting

4 · 17%

5 Clearly insulting

2 · 8%

The mean says probably fine. The tail says six VPs would read this as insulting. That spread is the product.

Dissent · ratings ≥ 4

6 / 24

Dissent share

25%

Spread · σ

1.18

Mean (summary only)

2.58

Confidence

0.61

Elapsed

11.4m

Call

Do not send without revision.

Rationales · 3 of 24

“The ‘quick question’ opener from a vendor they didn’t ask for is the insult, not the product claims.”

Likert 4 · VP Eng, 200-person Series B

“I’d archive it. The CC of my CEO is what makes it hostile, not the copy.”

Likert 5 · Staff engineer, FAANG-alum, now startup

“Tone is fine. Asserting you ‘already use us internally’ when you don’t — that’s the tell.”

Likert 3 · Director of Engineering, public SaaS

Response payload · truncated application/json

{
  "id": "jdg_2n8k0q",
  "status": "completed",
  "question": "Does this outbound email read as insulting to a VP of Engineering?",
  "panel": { "id": "pnl_eng_leaders_us_b2b", "n": 24 },
  "scale": "likert_5",
  "distribution": {
    "counts": [5, 8, 5, 4, 2],
    "mean": 2.58,
    "std": 1.18,
    "dissent_share": 0.25,
    "dissent_definition": "ratings >= 4"
  },
  "confidence": 0.61,
  "call": "do_not_send_without_revision",
  "elapsed_ms": 684000
}

Figure 1. Representative judgment, not a score. Spread and dissent are the output; the mean is a summary statistic.

03

What we sell

Not a verified fact. A priced, sourced distribution of human opinion, with provenance. Three surfaces, shipped in this order.

Representative judgment

“What does a real population think of this?” n responses, spread, rationales, confidence.

Models cannot simulate this.

Accountable judgment

A named, credentialed human reviews and signs off. The value is the signature and the audit trail.

AI cannot supply this by definition.

Expert judgment · deliberately deprioritized

Where models improve fastest and margin erodes soonest. We will sell it when a buyer needs it. It is not the wedge.

04

Why the economics work here, and not in verification

Verification wants consensus. Judgment inverts that. Redundancy is not a cost to be squeezed. It is the thing being bought.

Verification

Consensus

Redundancy burns margin.
Competes with free search.
Priced as a lookup.
The second identical answer is waste.

Judgment

Distribution

Redundancy is the product.
A buyer paying for n=30 is paying for the spread.
Priced against the cost of the wrong decision.
The dissenting fifth answer is the one that matters.

05

The panel

Senior technical B2B professionals — engineers, engineering leaders, technical buyers. Employment-verified. Scored over time on calibration, representativeness, and effort. Not a survey mill.

First paid use case: pre-send review of agent-generated customer-facing communication and positioning copy, judged by the ICP it targets.

Nearest comparable: Wynter, which already showed that B2B teams will pay for ICP judgment. They sell a dashboard to humans. We sell an API to agents, with a scored panel behind it.

Recruiting panelists →

06

API, agent-native

POST returns an id, a status, and an eta. Results arrive on a webhook, or you poll. An MCP server wraps the same endpoints. There is no synchronous mode in v1.

create_judgment · JavaScript POST /v1/judgments

const job = await fetch("https://api.judgmentcall.ai/v1/judgments", {
  method: "POST",
  headers: {
    Authorization: "Bearer $JUDGMENTCALL_API_KEY",
    "Content-Type": "application/json",
  },
  body: JSON.stringify({
    question: "Does this outbound email read as insulting to a VP of Engineering?",
    stimulus: { type: "text", body: emailBody },
    panel: "eng_leaders_us_b2b",
    n: 24,
    webhook_url: "https://your-agent.example/hooks/judgment",
  }),
}).then((r) => r.json());

// { id: "jdg_2n8k0q", status: "queued", eta_seconds: 900 }
// Result: webhook judgment.completed, or poll GET /v1/judgments/:id

REST v1 sketch and MCP note →

07

Request access

We’ll send access in waves. Buyers first from the GTM network; panelists from verified engineering talent.