Work.

How it works

• Structure: five signals, each scored 0-100 on its own rubric, then a weighted sum BQS = Σ wᵢ·sᵢ with Σ wᵢ = 1. Weight distribution: behaviour 0.30, cheque capacity 0.25, thesis coherence 0.20, acquisition experience 0.15, declared urgency 0.10.
• Normalisation: per signal, min-max against the trailing 90-day cohort using the 5th and 95th percentiles, sᵢ = 100·(x − p5)/(p95 − p5) clipped to [0, 100], so a single outlier cannot dominate the ranking.
• Behaviour signal: effort-weighted event count over a 30-day window (logins, deal views, document requests, meeting acceptances), each event type weighted by the effort it costs the buyer, with recency decay 0.5^(days/14).
• Cheque capacity: declared range mapped to five bands, then validated against observable proof (prior transactions, committed vehicle, verified funding). Unproven declarations are capped at 60/100.
• Thesis coherence: rubric of five binary checks (sector named, geography named, size band, timeline, exclusion criteria). Score = 20 × checks passed, so coherence comes from observable specificity and not self-assessment.
• Missing evidence: weights are renormalised over the signals that exist, wᵢ' = wᵢ / Σ available wᵢ, and the record carries a low-confidence flag when signal coverage falls below 60%.
• Output: tier A ≥ 75, B 55-74, C 35-54, D < 35, written to a CRM field. A and B route to Sales within 24 hours, C to nurture sequences, D to self-serve.

Design decisions

• Five signals capped by design, because a model an account executive can restate in one sentence is a model they actually apply.
• Tiers exposed to operators, continuous score kept underneath: a 2-point difference is noise, a tier change is a decision.
• Renormalisation instead of zero-filling, so a new account is not punished for having a short history.
• A hard cap on unverified claims rather than exclusion, so self-reported capacity can lift a buyer but never carry them alone.
• Quarterly recalibration against the closed-won base rate, validated by rank correlation between tier and conversion rather than by overall accuracy.

What I would change

• Calibrated on one market and roughly 250 accounts, so the weights are trustworthy at tier level and not at single-point differences.
• It ranks probability of closing, not quality of fit. Those are two different target variables and deserve two models.

How it works

• Two stages in fixed order. Stage 1 evaluates deterministic override rules; stage 2 runs a weighted behavioural score 0-100. If a rule fires, the score is ignored.
• Override rules: an explicit cancellation or mandate withdrawal sets risk to 100; an account in exclusivity or with a signed letter of intent is capped at 20; an account under 7 days old is routed to the onboarding read instead of the churn read.
• Behavioural score: R = Σ wⱼ·dⱼ over five dimensions, contact frequency 0.25, meeting cadence 0.25, silence duration 0.20, relational tone 0.20, unresolved conflict 0.10, each dimension expressed 0-100.
• Silence: d = min(100, 100 × days since last two-way contact / 30), so a month of quiet saturates the dimension instead of growing without bound.
• Meeting cadence: expected meetings are defined per stage, and the dimension is 100 × (1 − min(1, held / expected)), which separates a slow stage from an abandoned one.
• Conflict and tone come from logged interaction events with polarity and severity 1-5, aggregated with recency decay 0.5^(days/21).
• Late adjustment: verified external interest (a competing platform, a direct buyer) adds 10 to 15 points, applied after the core read and capped at 100.
• Bands to plays: ≥ 70 critical, owner call within 48 hours; 40-69 watch, weekly check-in plus a re-match; < 40 healthy, standard cadence.

Design decisions

• Rules before weights, because when operational truth and a smooth curve disagree, the rule is right.
• Interpretable linear weights over a black box, so any operator can reconstruct why an account was flagged from five numbers.
• Tone weighted at the same level as meeting cadence, since logged friction precedes cancellation by weeks in the observed cases.
• External interest applied as a post-adjustment rather than a sixth dimension, keeping the core read stable when that evidence is missing.
• Specified in writing before implementation, so the thresholds could be argued by people who do not read code.

What I would change

• Thresholds are calibrated on one market and one product; the 30-day silence saturation and the stage-level meeting expectations would need refitting elsewhere.
• It predicts risk well and cause only partially: the dimension tells you where to look, the operator still has to ask the right question.

How it works

• Five independent modules, each returning friction 0-100: relational tone, responsiveness, match fit, resilience after a failed process, and search complexity. Composite F = Σ wₖ·fₖ, with the weights deliberately left unset until enough labelled churn cases exist. The specification fixes the variables and the scales, not the coefficients.
• Responsiveness: median reply latency mapped on a log scale, under 4 hours = 0, 24 hours ≈ 50, 72 hours or more = 100, combined with the share of unanswered outbound messages over a rolling 30 days.
• Match fit: share of surfaced opportunities inside the declared box (sector, size band, geography, timeline). Each match is weighted by counterpart seriousness, signed mandate, financials shared, historical response rate, so twenty loose matches score below three verified ones.
• Resilience: a 30-day window after a process falls through, measured as the ratio of activity in that window to activity in the 30 days before it. Below 0.5 means the relationship, not the deal, is the thing at risk.
• Search complexity: count of simultaneous hard constraints. Above four, the friction is classified as structural, coming from the mandate itself, rather than as service friction the team can fix.
• Trend over level: every module reports the slope over the trailing 4 weeks, and intervention is triggered by a rise of 15 points or more, not by an absolute value.
• Each module resolves to a different play: expectation setting, re-matching, or a human conversation with the owner.

Design decisions

• Modules before a single composite, so a weak module can be refitted without touching the rest of the architecture.
• Friction rather than satisfaction as the organising concept, because latency, match rates and activity ratios are observable while satisfaction is reported.
• Match volume explicitly discounted by counterpart quality, since raw match counts were the metric that hid the problem in the first place.
• Resilience isolated as its own dimension, because a failed process is a fork in the relationship and behaves differently from ordinary inactivity.
• Delivered as a written specification with variables, scales and triggers rather than as a prototype, since the scarce resource was shared definitions.

What I would change

• It is a specification, not a validated model. The composite weighting stays honest only once there are enough labelled outcomes to fit it.
• Two modules depend on event data that was not captured consistently, so instrumentation has to come before prediction.

How it works

• Every logged emotional signal carries three attributes: a category (onboarding, valuation expectations, buyer meetings, process delays, commercial friction), a polarity, and a severity from 1 to 5.
• Score: S = 100 × Σ (w_c · sev_e · decay_e) / Σ (w_c · 5), normalising by the maximum reachable mass, so the scale stays 0-100 no matter how many events an account has.
• Weights w_c are defined per lifecycle phase. A category that cannot exist yet in the current phase gets weight 0 and its share is redistributed proportionally across the categories in play, so an absence never dilutes the score.
• Recency: decay = 0.5^(days/21), a 21-day half-life, with a floor of 0.25 reserved for the first relationship signal, so a bad start never disappears entirely.
• Repetition inside a category is not summed. It is smoothed with an exponentially weighted mean, α = 0.5 on the most recent event, which lets a recovery move the score instead of being buried by history.
• Bands and plays: 0-24 stable, standard cadence; 25-49 monitor, note in the weekly review; 50-74 at risk, structured call with a written next step; 75-100 critical, escalation to the lead and a recovery plan within 72 hours.

Design decisions

• Per-phase weight tables instead of one global table, because a signal in week one and the same signal mid-process are not the same event.
• A 21-day half-life chosen over a fixed lookback window, so the score cools down continuously instead of dropping when an event leaves the window.
• Normalising by reachable mass rather than by event count, otherwise a talkative account looks riskier than a silent one.
• Bands, not the raw number, shown to operators, so the response to a score is unambiguous.
• The floor on the first relationship signal added deliberately, since early distrust kept predicting late churn.

What I would change

• Severity and polarity are human input, so the model inherits the discipline of whoever logs the interaction.
• The band cut points come from judgement more than from labelled outcomes, and that calibration debt has to be paid with real cases.

How it works

• Before the call: a fixed 20-minute research checklist, six items, investment history, prior transactions, sector and geography, cheque size, likely source of funds, and three questions written specifically for that investor. If the checklist is not documented in the CRM, the meeting is rescheduled.
• During the call: 45 minutes split 15 / 20 / 10. Fifteen minutes on the thesis, using open questions and no product talk; twenty minutes walking the platform on live opportunities that match what was just said; ten minutes agreeing next steps with a date.
• Qualification criteria are explicit and captured as fields, not impressions: declared box (sector, size band, geography), decision process, timeline, and the number of deals the investor expects to review per month.
• After the call: a six-item closing checklist inside 24 hours, CRM fields updated, declared criteria recorded, contact cadence agreed, first matches sent, objections logged verbatim, next date booked.
• Measurement: portfolio retention at 90 days per operator, reviewed monthly, which is the number that moved from roughly 75% to around 89%.

Design decisions

• A protocol rather than a training session, because a protocol survives turnover and a session does not.
• Preparation made a blocking precondition rather than a recommendation, since that was the only step that reliably disappeared under pressure.
• Fixed sequence and fixed time split, free tone, so structure does not turn the call into a recital.
• Objections recorded verbatim instead of summarised, because the wording is what feeds the next re-match.
• Ten items total across the two checklists, since a longer protocol is a protocol nobody follows.

What I would change

• A protocol raises the floor more than the ceiling, and the best operators still outperform it.
• It needs revision as the product changes, otherwise the walkthrough ages badly.