method · 16 June 2026 5 min

Confidence calibration: what high / medium / low means on a score chip

Q: What makes a listing 'high' confidence?

All three pillars resolved against primary data: vision ran on a gallery meeting the photo-count floor, the AVM had at least the district-comparables floor with a known surface, and the listing has valid geocoded coordinates.

Q: Why is my listing's Window-View confidence low?

Either the gallery had fewer than 3 photos (below the vision scorer's image threshold), or the vision step was skipped — recorded as a 'vision-skipped-*' methodology tag on the score record.

Q: When does the M-Value AVM drop to low confidence?

When the district has fewer than 3 comparables (below the AVM's floor), or when listing surface (m²) is missing so per-m² normalisation can't be computed.

Q: What breaks the WC 2030 Catalyst pillar?

Missing geocoded coordinates — without lat/lng, proximity to host-city catalysts cannot be measured, so the pillar is not computable.

Q: Why don't you just show a chip anyway?

We render no chip rather than a fabricated one. If any pillar fails to resolve against primary data, the chip is suppressed — fail-closed presentation is the policy. A 'v1.deterministic' methodology tag also means the listing was not model-scored and should not be shown as high-confidence vision output.

A score chip is only as honest as the data behind it. marocain.investments rates overall confidence 'high' only when all three pillars (Window-View, M-Value AVM, WC 2030 Catalyst) resolve against primary data — and SUPPRESSES the chip rather than fabricate one if any pillar fails. Here is exactly what drops each pillar.

The rule

A score chip on a listing is only as honest as the data behind each pillar. marocain.investments rates overall confidence "high" only when every pillar — Window-View, M-Value AVM, WC 2030 Catalyst — resolves against primary data. If any pillar cannot be resolved, we suppress the chip rather than fabricate one — fail-closed presentation is the policy.

Window-View pillar

Two failure modes drop this to low:

Thin galleries: fewer than 3 photos falls below the vision scorer's minimum-image threshold.
Vision skipped: a score_methodology value matching vision-skipped-* means the vision step never ran on that listing.

A related case: score_methodology = 'v1.deterministic' means the listing was not model-scored at all — so it should not be presented as high-confidence vision-derived output. Because vision coverage across the catalogue is still partial, a non-trivial share of listings currently carry a vision-skipped-* tag and cannot earn high Window-View confidence until the drain backfills them.

M-Value AVM pillar

Two failure modes drop this to low:

Sparse comparables: fewer than 3 district comparables falls below the AVM's minimum-comparables floor (see the M-Value deep-dive).
Missing surface: surface_m2 IS NULL makes per-m² normalisation impossible.

WC 2030 Catalyst pillar

Confidence here depends on geocoded coordinates: without lat/lng, proximity to host-city catalysts cannot be measured, so the pillar is not computable.

What "high" actually means

In practice, high = all three pillars green simultaneously:

Vision ran (no vision-skipped-* tag) on a gallery meeting the photo-count floor.
AVM had at least the comparables floor in the listing's district, with a known surface.
The listing has valid geocoded coordinates for catalyst scoring.

Anything less, and we'd rather render nothing than fake it.

The honest ledger

What this proves:

The three-pillar gating rule and the suppress-rather-than-fake presentation policy.
The specific data conditions that downgrade each pillar (photo count <3 and vision-skipped-* for Window-View; <3 district comparables and NULL surface for M-Value; NULL lat/lng for WC 2030 Catalyst).
That v1.deterministic is not model-scored and should not be displayed as high-confidence vision output.

What it does NOT prove:

The exact share of catalogue listings currently in each confidence tier (vision coverage is partial; we don't quantify the distribution here).
The numeric weighting of pillars within the composite score, or how "medium" is bounded between "high" and "low".
Backfill ETA for vision-skipped listings to migrate into the high-confidence pool.

*Internal methodology document — this describes how marocain.investments' own confidence calibration is built. Reasoning is platform-canonical. Decision-support, not a certified appraisal.*

What makes a listing 'high' confidence?

All three pillars resolved against primary data: vision ran on a gallery meeting the photo-count floor, the AVM had at least the district-comparables floor with a known surface, and the listing has valid geocoded coordinates.

Why is my listing's Window-View confidence low?

Either the gallery had fewer than 3 photos (below the vision scorer's image threshold), or the vision step was skipped — recorded as a 'vision-skipped-*' methodology tag on the score record.

When does the M-Value AVM drop to low confidence?

When the district has fewer than 3 comparables (below the AVM's floor), or when listing surface (m²) is missing so per-m² normalisation can't be computed.

What breaks the WC 2030 Catalyst pillar?

Missing geocoded coordinates — without lat/lng, proximity to host-city catalysts cannot be measured, so the pillar is not computable.

Why don't you just show a chip anyway?

We render no chip rather than a fabricated one. If any pillar fails to resolve against primary data, the chip is suppressed — fail-closed presentation is the policy. A 'v1.deterministic' methodology tag also means the listing was not model-scored and should not be shown as high-confidence vision output.