Weighted average of the three sub-scores — the signal composite — then mapped onto the displayed 0–10 scale through a published calibration curve (v3). No confidence multiplier, no hidden penalty. The popover on the score chip shows both the signal composite and the calibrated score, so the result stays reproducible.

signalComposite = round(
  innovationIndex × 0.35
  + functionalityScore × 0.45
  + activityScore × 0.20
, 1)
mindberScore = calibrate_v3(signalComposite)

The calibration curve is monotonic (a higher composite always yields a higher or equal score) and exists to spread scores across the full 0–10 range instead of clustering every product near the middle. A live, working tool with modest reach reads ~7; a category leader ~8.5–9; a weak or inactive tool stays low.

When one sub-score is missing, the missing input’s weight redistributes proportionally across the remaining two. Two of three inputs labels confidence as medium; all three as high. With fewer than two scored inputs, no Mindber Score is shown (insufficient) and the page renders “Score pending” instead.

Fallback weights when one input is missing

• No Activity → Innovation 44% · Functionality 56%
• No Innovation → Functionality 69% · Activity 31%
• No Functionality → Innovation 64% · Activity 36%

Confidence states

• High — all three sub-scores present. Default 35/45/20 weights apply.
• Medium — two of three sub-scores present. Fallback weights as above.
• Insufficient — fewer than two present. No Mindber Score is shown.

Score band interpretation

Bands are absolute — they do not shift with the category. The category percentile label (below) tells you how the product ranks against its peers separately.

Average of three sub-scores — Novelty, Differentiation, Concept clarity — each 0–10.

innovationIndex = round((novelty + differentiation + conceptClarity) / 3, 1)

• Novelty — how distinct the core concept is from the category baseline. LLM-inferred, grounded in the category’s top 3 tools and common features (see below).
• Differentiation — count of features present in this product but missing from >=50% of category peers. Scaled to 0–10: min(10, unique × 1.5). LLM-inferred, verified against the category common-features list.
• Concept clarity — how unambiguous the value proposition reads from the tagline and description alone. LLM-inferred, grounded in the source copy.

Confidence-weighted average of four sub-scores. Feature breadth and Performance reliability are deterministic — computed from source data with no LLM involvement. The other two are LLM-inferred.

functionalityScore = Σ(score × baseWeight × confidence) / Σ(baseWeight × confidence)

Base weights:breadth 0.25 · depth 0.30 · integrations 0.20 · performance 0.25. Confidence weighting means a sub-score with thin evidence pulls less weight than one backed by real data.

• Feature breadth (deterministic). Count of features from collected source records. Score = min(10, round(log2(count + 1) × 2)). 5 features → 5; 15 → 8; 30+ → 10.
• Feature depth (LLM). Quality of the top features vs the category leader. Cites specific named features.
• Integration surface (LLM with caveat). Public API + webhook + native integration evidence. Where no integration evidence exists, the score defaults to neutral (5) with confidence ≤ 0.4.
• Performance reliability (deterministic). Computed from the product’s website engagement score and homepage response time: 0.7 × (livenessScore / 10) + 0.3 × responseTimeScore. Response-time score is 10 for ≤500 ms, linearly degrading to 1 at ≥5000 ms.

Mindber Activity Score converts live public engagement checks into the same 0–10 scale used by the composite. It reflects update frequency and availability signals from public sources, not product quality or company health.

The underlying 0-100 engagement score is documented in the engagement methodology.

A score in isolation isn’t meaningful — a 7.3 in a crowded category is different from a 7.3 in a deep one. We compute per-category distributions (median, p25, p50, p75, p90) and bucket each final score into one of five labels:

• ≥ p90top 10%
• ≥ p75top 25%
• ≥ p50above median
• ≥ p25below median
• < p25bottom 25%

A baseline is only published once a category has at least 10 scored products; below that threshold, percentile is shown as unranked.

• Full baselines (median/p25/p75/p90) refresh weekly. A drift alert fires when a category median moves by more than 1.0 point.
• Percentile labels refresh daily. The underlying raw scores stay fixed until the product is regenerated — only the bucket label updates.
• Product scores are regenerated whenever the source inputs change (name, tagline, description, categories, candidate alternatives) or when the scoring methodology is upgraded.

• Every LLM-written sentence must cite a specific feature, use case, or baseline data point. Marketing adjectives are banned.
• A banned-word safety filter runs on every rationale; offending text is blanked out rather than retried.
• Any sub-score below 4/10, composite confidence below 0.6, or a safety-filter replacement flags the record for admin review before publication.
• Scores compare within category, never absolute. “Below median” is factual; “bad” is not a word we use.

Weighted average of the three sub-scores — the signal composite — then mapped onto the displayed 0–10 scale through a published calibration curve (v3). No confidence multiplier, no hidden penalty. The popover on the score chip shows both the signal composite and the calibrated score, so the result stays reproducible.

signalComposite = round(
  innovationIndex × 0.35
  + functionalityScore × 0.45
  + activityScore × 0.20
, 1)
mindberScore = calibrate_v3(signalComposite)

The calibration curve is monotonic (a higher composite always yields a higher or equal score) and exists to spread scores across the full 0–10 range instead of clustering every product near the middle. A live, working tool with modest reach reads ~7; a category leader ~8.5–9; a weak or inactive tool stays low.

When one sub-score is missing, the missing input’s weight redistributes proportionally across the remaining two. Two of three inputs labels confidence as medium; all three as high. With fewer than two scored inputs, no Mindber Score is shown (insufficient) and the page renders “Score pending” instead.

Fallback weights when one input is missing

• No Activity → Innovation 44% · Functionality 56%
• No Innovation → Functionality 69% · Activity 31%
• No Functionality → Innovation 64% · Activity 36%

Confidence states

• High — all three sub-scores present. Default 35/45/20 weights apply.
• Medium — two of three sub-scores present. Fallback weights as above.
• Insufficient — fewer than two present. No Mindber Score is shown.

Score band interpretation

Bands are absolute — they do not shift with the category. The category percentile label (below) tells you how the product ranks against its peers separately.

Average of three sub-scores — Novelty, Differentiation, Concept clarity — each 0–10.

innovationIndex = round((novelty + differentiation + conceptClarity) / 3, 1)

• Novelty — how distinct the core concept is from the category baseline. LLM-inferred, grounded in the category’s top 3 tools and common features (see below).
• Differentiation — count of features present in this product but missing from >=50% of category peers. Scaled to 0–10: min(10, unique × 1.5). LLM-inferred, verified against the category common-features list.
• Concept clarity — how unambiguous the value proposition reads from the tagline and description alone. LLM-inferred, grounded in the source copy.

Confidence-weighted average of four sub-scores. Feature breadth and Performance reliability are deterministic — computed from source data with no LLM involvement. The other two are LLM-inferred.

functionalityScore = Σ(score × baseWeight × confidence) / Σ(baseWeight × confidence)

Base weights:breadth 0.25 · depth 0.30 · integrations 0.20 · performance 0.25. Confidence weighting means a sub-score with thin evidence pulls less weight than one backed by real data.

• Feature breadth (deterministic). Count of features from collected source records. Score = min(10, round(log2(count + 1) × 2)). 5 features → 5; 15 → 8; 30+ → 10.
• Feature depth (LLM). Quality of the top features vs the category leader. Cites specific named features.
• Integration surface (LLM with caveat). Public API + webhook + native integration evidence. Where no integration evidence exists, the score defaults to neutral (5) with confidence ≤ 0.4.
• Performance reliability (deterministic). Computed from the product’s website engagement score and homepage response time: 0.7 × (livenessScore / 10) + 0.3 × responseTimeScore. Response-time score is 10 for ≤500 ms, linearly degrading to 1 at ≥5000 ms.

Mindber Activity Score converts live public engagement checks into the same 0–10 scale used by the composite. It reflects update frequency and availability signals from public sources, not product quality or company health.

The underlying 0-100 engagement score is documented in the engagement methodology.

A score in isolation isn’t meaningful — a 7.3 in a crowded category is different from a 7.3 in a deep one. We compute per-category distributions (median, p25, p50, p75, p90) and bucket each final score into one of five labels:

• ≥ p90top 10%
• ≥ p75top 25%
• ≥ p50above median
• ≥ p25below median
• < p25bottom 25%

A baseline is only published once a category has at least 10 scored products; below that threshold, percentile is shown as unranked.

• Full baselines (median/p25/p75/p90) refresh weekly. A drift alert fires when a category median moves by more than 1.0 point.
• Percentile labels refresh daily. The underlying raw scores stay fixed until the product is regenerated — only the bucket label updates.
• Product scores are regenerated whenever the source inputs change (name, tagline, description, categories, candidate alternatives) or when the scoring methodology is upgraded.

• Every LLM-written sentence must cite a specific feature, use case, or baseline data point. Marketing adjectives are banned.
• A banned-word safety filter runs on every rationale; offending text is blanked out rather than retried.
• Any sub-score below 4/10, composite confidence below 0.6, or a safety-filter replacement flags the record for admin review before publication.
• Scores compare within category, never absolute. “Below median” is factual; “bad” is not a word we use.