Reference (partial) — Question wording and thresholds are brand-configurable. Behavior below is verified from the Jev match-check service and Admin display rules (November 2026).

Overview

Match checks are automated yes/no questions about a member and a candidate (their About me / Looking for text, profile answers, decline history, and matchmaker Admin insights). When a question scores high enough, Admin shows a flag — often in red — so staff review the pair before suggesting it.

Questions are data, not hard-coded releases: ops add rows on Settings → General Codes with type matchCheck (filter the list by that type). Each active row is one question the system asks Jev about the pair.

Read flags on View Potential Matches, on Check Match, and (after create) during automated PENDING APPROVAL release — see compare list rank & match checks for how lines look on screen.

Where checks run

WorkflowWhat staff sees
Rank potentials on potential matches Checks run with the list-ranking pass. Results feed the LLM rank column warning icon and the Checks before suggesting block on compare (before an introduction exists).
New automated introduction in PENDING APPROVAL Checks run again when Match Advisor finishes. A stop (for example earlier decline) or a hold from flagged questions can block automatic release — staff may still Approve manually. See pending approval workflow and stuck in pending approval.
Code-only checks (no Jev call) verified — Pair was introduced before and someone declined; empty About me on a side. These always apply when relevant — see common Jev topics.

Add or edit a question

1
Open General Codes
Go to Settings → General Codes, set Type to matchCheck, and click Add Code (or open an existing row).
2
Write the question in plain English
The name field is the full question text. Write about the member and the candidate (the two people in the pair). Example style: “Does the member’s About me ask for a home type the candidate’s profile contradicts?”
3
Tune the Extra JSON
Set how strongly a “yes” must score before the pair is flagged, whether to ask the question both ways round, optional short label for Admin, and pair vs member scope — see Extra JSON fields.
4
Save and dry-run
Save the code, run Rank potentials on a test member with known pairs, and read flags on compare before changing production thresholds. Prefer inactive over delete when retiring a question — old assessment rows may still reference the key.

More context on General Codes: General Codes.

Extra JSON fields

When you add a matchCheck row, the Extra editor starts with sensible defaults. Only keys present in Extra are editable in Admin.

Extra key (in JSON)Plain meaningTypical value
threshold Jev “yes” score from 0 to 1. At or above this value, the question holds the pair (red on compare; may block auto-release from pending). 0.7 (70%) — same default Admin uses for new rows.
bothWays When true, the system also asks the question with member and candidate swapped. true for most pair questions; set false for member-only patterns that should not mirror.
humanTopic Short label on the flag line in Admin (for example “School”, “Parnassa plan”). If empty, Admin falls back to built-in topic labels — see common Jev topics. One or two words ops will recognize in the table tooltip.
scope pair (default) — question is about this member ↔ this candidate; a yes can hold the pair. member — question is about the member alone; a yes is a note on every pair for that member, not a hold by itself. Leave default unless product asked for a member-wide note question.
key Optional stable id stored on assessment rows. If omitted, the system uses an auto key tied to the General Code id. Only set when engineering or ops need a fixed id in exports.

Pair vs member scope

Pair questions are the default: each flag applies to one member–candidate combination. Use these when the concern is whether these two people fit together.

Member questions run once per member (not per candidate). A “yes” shows as context on every potential row for that member but does not by itself stop automatic release the way a pair hold does. Use this for patterns like “this member repeatedly declines for the same quality” where no single candidate is at fault.

Thresholds and colors

verified — On compare and in the potential-matches tooltip:

  • Red (hold) — Jev “yes” score at or above the question’s threshold (default 0.7 / 70%), plus built-in stops such as Declined before and About me is empty.
  • Orange (soft) — Score from 60% up to (but below) that question’s threshold — for example 61%–69% when threshold is 70%. Shown on compare only; the potential-matches warning icon ignores soft-only flags. If you set threshold below 60%, there is no soft band for that question — scores at or above the threshold become holds directly.
  • Muted sub-line — List ranker disagreed with Jev; compare still shows the line for audit, but the potential-matches icon is suppressed even when the score would otherwise be a hold.

Lowering threshold on a General Code row makes that question fire more often; raising it makes flags rarer. Change one question at a time and re-run ranking on test pairs. Step-by-step scenarios: Worked threshold examples.

Worked threshold examples

Use these patterns when tuning a matchCheck row. Numbers assume the default threshold 0.7 unless noted.

ScenarioJev scoreWhat ops seesAutomated pending release
Borderline soft 68% on a pair question Orange line on compare; no warning icon on the potential-matches rank column. Does not block by itself — other gates (advisor, soft deal breakers, pending mode) may still hold the row.
Hold at default threshold 72% Red hold line; rank-column icon when the ranker did not dispute the flag. After Match Advisor completes, a pair hold can set pending due to match checks until staff review — see pending approval.
Make the question stricter Same pair now scores 66% after you raise threshold to 0.85 Previously a hold at 72%; now only orange (soft) — fewer false alarms, but staff may miss borderline pairs unless they open compare. Weaker gate on auto-release for that question until scores climb above the new threshold.
Make the question more sensitive Score 66% after you lower threshold to 0.65 Soft band becomes a red hold (66% ≥ 65%). More introductions stay in PENDING APPROVAL until someone approves — expect higher pending volume.
Ranker disagrees 90% but compare shows a muted disputed sub-line Flag text visible on compare; potential-matches icon hidden for that flag. Treat as “review both readings”; do not assume the icon alone means a hard hold.
Earlier-decline stop (code check) Not a percentage — pair introduced before and someone declined Declined before in red; always a stop. Auto-release blocked regardless of advisor success; staff decline or follow override policy.

When Match Advisor fails

verified — If Match Advisor itself errors, a hold that exists only because the Jev call failed does not block automated promotion (the row may still advance under your brand’s existing failure rules). If Match Advisor succeeded but Jev returned an error, the system treats missing checks as a hold and keeps PENDING APPROVAL. Stops such as Declined before still apply even when Jev errors. Troubleshooting: stuck in pending approval.

Labels on screen

Each flag line shows:

  • A topic label — from humanTopic on your General Code when set, otherwise built-in labels such as Location.
  • Which side — member's side or potential's side when the question is directional.
  • A confidence percentage when Jev returned a score.

Full catalog of built-in topic keys: compare list rank → Common Jev topics.

When no brand rows exist

verified — If your brand has no active matchCheck General Codes, the system uses a small built-in question set (earlier-decline wording, text-against-profile, and a member-scope repeated-decline pattern). Adding brand rows replaces that default set for your brand — plan migrations with engineering before deleting the last row.

Brand 0 rows apply when a brand has none of its own active questions (shared fallback catalog).