The Cubed Score, in full detail.

The Cubed Score is a deterministic, weighted, stage-aware composite of 101 underlying factors organised into 9 thematic cubes. It is designed to behave like a venture analyst with a fixed methodology: same inputs produce the same score, every factor must cite evidence, every rating has a defined mathematical impact.

It is not a black-box ML model. The LLM (Gemini 2.5 Flash via the Lovable AI Gateway) is used only to classify each factor as ontrack / attention / critical / unknown and to extract supporting evidence. All arithmetic — weighting, capping, penalties, aggregation, valuation, forecast — is executed in code. The LLM provides perception; the engine provides judgement.

Design principles

• • Deterministic math — score is a closed-form function of factor ratings.
• • Stage-gated activation — only factors relevant to a company’s maturity are scored; others are excluded, not zeroed.
• • Asymmetric penalties — Critical issues cap and penalise; on-track signals never inflate above 1.0×.
• • Evidence-bound — every factor must carry a reason and an action; missing evidence collapses to ‘unknown’ rather than guessed.
• • Reproducibility — same scraped corpus + uploads → same score within ±1 pt (temp 0.2).

Revenue Model and Financials carry the heaviest weight (13% each) because at Seed they are the highest-signal predictors of survival to Series A. People, Product, Market and Customers each carry 12% as the ‘four pillars’ of fundamentals. Personality and Sentiment carry 8–10% — informative but easier to manipulate. Economy is held at 8% because macro is mostly an exogenous filter, not a differentiator.

Step 1 — RAG multiplier

µ(ontrack)    = 1.0
µ(attention)  = 0.6
µ(critical)   = 0.2
µ(unknown)    = excluded from denominator
µ(N/A)        = excluded (stage-gate inactive)

Step 2 — Cube score

S(c) = 100 · ( Σ_{f∈F_c} µ(σ_f) ) / |F_c|

if ∃ f ∈ F_c with σ_f = critical  →  S(c) = min(S(c), 65)   # critical cap

Step 3 — Overall score

Score_raw = Σ_c w(c) · S(c)
N_crit    = total count of critical factors across all cubes
Score     = Score_raw − 5 · [ N_crit ≥ 3 ]               # systemic-risk penalty
Score     = clip(Score, 0, 100)

Step 4 — Verdict band

Score ≥ 70       → PROGRESS (Green)
40 ≤ Score < 70  → WATCH    (Amber)
Score < 40       → PASS     (Red)
Also PASS if the late-stage override fires (see §7).

The 1.0 / 0.6 / 0.2 curve is convex-downward: Attention recovers 60% of value (solvable), Critical recovers only 20% (structurally damaging). The 65-cap prevents ‘one disaster averaged away by nine green lights’. The −5 penalty triggers at the 3-critical concentration threshold, which empirically correlates with failed Series A bridges in our backtest corpus.

A · Healthy Seed B2B SaaS

38 active factors · 30 ontrack · 6 attention · 2 critical (both in Customers)
Sum µ = 30·1.0 + 6·0.6 + 2·0.2 = 34.0

Cube avgs:  Economy 82 · Personality 78 · Sentiment 75 · People 80
            Product 77 · Market 72 · RevenueModel 70 · Financials 74 · Customers 62
Weighted = 0.08·82 + 0.10·78 + 0.08·75 + 0.12·80 + 0.12·77
         + 0.12·72 + 0.13·70 + 0.13·74 + 0.12·62 = 74.0
N_crit = 2  →  no systemic penalty
Final Score = 74  →  PROGRESS

B · Critical cap kicks in

Same company but Customers cube avg = 78 with one Critical factor
Customers raw = 78  →  capped to 65
Weighted impact: 0.12·65 = 7.80  (vs 0.12·78 = 9.36)
Δ = −1.56 points on overall

C · Systemic-risk penalty

4 Critical factors spread across People, Product, Financials, Customers
N_crit = 4  →  −5 applied to overall; each affected cube also capped at 65
A 72-pre-penalty company → final 67 → still PROGRESS
A 44-pre-penalty company → final 39 → PASS

D · Late-stage override

Engine detects 'Series B announced March 2025'  →  finalStage = 'Series B'
Cubed Fund I mandate = Seed only  →
   verdict.label    = PASS
   verdict.headline ≈ 'Outside mandate — Series B'
   fund_fit_criteria flagged 'fail — outside Cubed Seed mandate'
Cube scores still computed (used by RevCube / OpCube)

Signals: funding announcements, press releases, Crunchbase / PitchBook mentions, headcount, ARR. If detected stage > recorded stage, finalStage is rewritten.

If detected stage ≥ Series A, Cubed Fund I mandate fails and verdict collapses to PASS regardless of score — late entry violates ownership economics (≥20% target at $15–25M pre).

V_pre(stage, score) = V_low(stage) + ( V_high(stage) − V_low(stage) ) · ( score / 100 )

The forecast view applies the stage multiplier to projected ARR to chart enterprise value over 5 years (see CompanyForecastCharts.tsx). Cubed Ventures Fund I writes $3–5M for a minimum 20% ownership, implying $15–25M pre-money at Seed.

g(score)     = 0.15 + (score/100) · 0.85        # organic CAGR
δ(score)     = 0.10 + (score/100) · 0.15        # cubed uplift
ARR_org(t)   = ARR_0 · (1 + g(score))^(t/12)
ARR_cubed(t) = ARR_0 · (1 + g(score) + δ(score))^(t/12)
# at year ≥ 2 add +0.05 second-cheque boost to cubed track

Floor ARR: if reported ARR is 0 / unknown, ARR_0 defaults to $100k (Pre-Seed / Idea) or $500k (other) so charts remain readable; flagged ‘estimated’ in the UI.

Valuation overlay: V(t) = ARR(t) · M_stage · decay(t), where decay(t) = 1 − 0.03·t (organic) or 1 − 0.02·t (cubed) — reflects multiple compression as companies mature.

Rows (7): Immediate Focus (3 month) · Mid-Term Focus (Next QTR) · Long-Term Focus (12 months) · People · Growth · Efficiency · Accuracy.

Columns (12): Full Year · Current Quarter · Revenue Protection · Revenue Growth · Strategy · Enabled · Pipeline Gen · Qualification · Deal Growth · Deal Acceleration · Competitor Positioning · Cost.

Scoring: the LLM populates each cell from the same evidence corpus as the cubes. For early-stage companies most cells will be unknown (grey) — correct behaviour, not a flaw. As the company matures and the engine ingests CRM/ERP data, cells flip to RAG.

Reading: ‘Immediate Focus × Pipeline Gen = red’ means pipeline coverage is the single most urgent fix in the next 90 days. ‘Efficiency × Cost = amber’ means cost discipline needs a plan but is not on fire. The chessboard turns the score into action.

The cube-level People score remains weighted at 12%, but the A-Player view feeds OpCube’s ‘who to keep / coach / replace’ recommendation and FundCube’s key-person risk assessment.

Cubed Fund I has 8 fixed criteria; each returns pass / attention / fail with a one-line detail. A company can score 78 (PROGRESS) but still fail fund-fit (wrong stage, wrong geography, conflict, ESG exclusion). Fund-fit failure overrides verdict → PASS. This is the bridge between the universal score and a specific mandate.

The 9-cube score is thesis-neutral (universal company quality). The TAI layer is thesis-specific. Both run on the same evidence corpus.

FitScore(company, thesis) =
    Σ_k  v(k) · m_k(company)            # weighted match score
  − Σ_h  H_h · [ hard_fail_h ]          # hard-rule penalties

where
  v(k)   = investor weight for criterion k (Σ = 100)
  m_k    = match function ∈ [0, 1] per criterion type
  H_h    = penalty for breaching a hard rule (default = ∞ → PASS)

Criterion types

Thesis schema (JSON contract)

{
  "thesis_id": "fund_xyz_2026",
  "investor": "XYZ Capital",
  "criteria": [
    { "k": "sector",          "type": "categorical", "allow": ["fintech","saas","ai"],         "w": 15 },
    { "k": "stage",           "type": "range",       "min": "Seed", "max": "Series A",          "w": 15 },
    { "k": "geography",       "type": "categorical", "allow": ["UK","US","EU"],                 "w": 10 },
    { "k": "cheque",          "type": "range",       "min": 1e6, "max": 5e6, "tol": 0.25,       "w": 10 },
    { "k": "ownership",       "type": "threshold",   "gte": 0.10, "tol": 0.05,                  "w": 10 },
    { "k": "gross_margin",    "type": "threshold",   "gte": 0.70, "tol": 0.10,                  "w": 10 },
    { "k": "product_quality", "type": "score",       "cube": "Product & Tech", "gte": 70,       "w": 10 },
    { "k": "thesis_fit",      "type": "semantic",    "statement": "AI infra for regulated industries", "w": 20 }
  ],
  "hard_rules": [
    { "k": "sanctions",   "type": "boolean",     "fail_if": true },
    { "k": "esg_exclude", "type": "categorical", "deny": ["tobacco","weapons","gambling"] }
  ],
  "verdict_bands": { "strong": 75, "watch": 50 }
}

Multi-thesis intelligence

Because the universal score is computed once, an LP can layer N theses and instantly see how the same pipeline ranks under each lens. This enables co-investment matching: ‘which 3 of our 12 LPs would this company suit?’ becomes a sort, not a meeting.

Calibration

Each investor’s weight vector v can be learned from historical decisions: feed in past invest / pass calls, regress on criteria values, and recover the implicit weights. This converts gut-feel committee behaviour into an auditable thesis vector.

Cross-product flywheel: each product writes evidence back to the central company record (RevCube updates Forecast Accuracy when actuals load; OpCube updates Employee NPS when a survey closes). The next score recomputation absorbs those updates automatically — the score sharpens as the company is operated through the stack.

{
  "overall_score": 74,
  "projected_arr": "$2.5M",
  "detected_stage": "Seed",
  "investment_verdict": {
    "label": "PROGRESS",
    "headline": "Strong Seed-stage B2B AI play, clean GTM",
    "bull": "...", "bear": "...", "recommendation": "...",
    "fund_fit_criteria": [
      { "criterion": "Stage", "status": "pass", "detail": "Seed" }
    ]
  },
  "cubes": [
    { "name": "Economy", "score": 82, "factors": [
      { "name": "Strong", "status": "ontrack",
        "reason": "Sector CAGR 22% per S&P 2026 outlook",
        "action": "Maintain macro watch" }
    ]}
  ],
  "research_insights": [
    { "type": "news", "title": "...", "url": "...", "date": "..." }
  ],
  "driving_metrics": [
    { "name": "Forecast Accuracy", "target": "80%+", "status": "unknown", "value": null }
  ],
  "operational_chessboard": [
    { "category": "Immediate Focus (3 month)",
      "cells": [ { "label": "Pipeline gap", "status": "attention" } ] }
  ]
}

Validation rules (enforced server-side)

• • cubes.length == 9 and names match the canonical set.
• • Each factor.status ∈ {ontrack, attention, critical, unknown}.
• • Overall score recomputed from cubes and rejected if delta > 2 from LLM-claimed score.
• • operational_chessboard.length == 7, cells.length == 12 per row.
• • Every factor must have non-empty reason and action.

Pipeline

1. Scrape (website, LinkedIn, news) → corpus.
2. Merge uploaded docs + follow-up answers.
3. Build prompt with corpus + 101-factor instructions.
4. LLM call → JSON.
5. Salvage layer.
6. Validate.
7. Server-side recompute of overall_score from cubes (LLM’s claimed score is a check, not source of truth).
8. Persist.

JSON salvage

LLM output is often truncated or contains stray markdown. The salvage layer strips code fences, balances braces, repairs trailing commas, retries with a stricter system prompt (‘Return ONLY valid minified JSON, no commentary, all strings <200 chars’) if parse fails, and finally falls back to a stub object with exceptions[] populated so the UI can flag ‘partial data’.

Determinism guarantees

Same corpus + same prompt + temp 0.2 → score reproducible within ±1 pt. The math layer is fully deterministic — drift comes only from LLM classification noise on borderline factors, which are flagged in the audit trail for human review.

Backtest metric

Brier(verdict) = mean( ( P(PROGRESS) − 1[outcome = positive] )² )

Cubed Fund I target: Brier ≤ 0.18 on 24-month-forward outcomes.
Each new outcome event updates the running Brier; weight-vector review every 25 new outcomes.

Weight learning (when enabled)

w* = argmin_w  Σ_i ( Σ_c w(c) · S(c, i) − y_i )²
              s.t.  Σ_c w(c) = 1,  w(c) ≥ 0

Initialised at the curated 8 / 10 / 8 / 12 / 12 / 12 / 13 / 13 / 12 vector.
Drift > 15% on any cube triggers a human-in-the-loop committee review
before deployment — never auto-deployed.