feat(settlement): Settlement Simplification Advisor — SGP engine, Smart Switch, Insights pipeline, and insurance pricing on top of greenroom-starter

[aliomraniH]

The problem

Greenroom's settlement tool works for flat and % of gross deals — but
62% of deals at The Crescent fall outside that coverage. Mariana, the
lead booker, settles them in a Google Sheet at 2am. The tour manager
checks the math on a phone calculator. The agent reads a PDF the next
morning and emails back with questions.

The root cause is not a missing calculator. Mariana already has one
and it works. The problem is structural: deal terms written in prose
that two people read differently at midnight, expense data scattered
across four systems, and deal structures that generate disputes by
construction — door deals that lose money 93% of the time, and vs
deals in the $1–5K bracket that dispute at 3× the rate of equivalent
flat deals on similar economics.

A more powerful tool built on top of that substrate inherits every one
of those problems. The correct move is to simplify the deal structures
first, generate clean data, and let that foundation carry AI tools
that are actually reliable. That is what this branch builds.

Why this slice

Two deal types were selected where the data is unambiguous and the fix
is actionable now:

Phase 1 — Door deals ($0–1K)
27 of 29 door deals lost the venue money over 24 months (−$29K
aggregate net). The structure pays the artist the entire pool above
expenses by construction; the venue has no upside regardless of
how the show performs. 6 deals are immediately switch-eligible;
projected savings $13,211 at ~$2,200 per converted deal.

Phase 2 — vs and % of net ($1–5K)
137 vs deals in this bracket dispute at 12.4% vs 3.8% for flat —
more than 3× the rate on substantively similar economics. 87.2% of
vs deals the percentage clause fires (mean overshoot ~$2,338/show),
so the conversion is calibrated to the actual historical payout,
not a discount from it. 34 deals immediately eligible; projected
savings $9,186.

Key data distinction: the analytics cross-tab (137 / 29 deals) is
the unconditional structural evidence. The switch-projected-grid
(34 / 6 deals) applies strict eligibility — settled + pre-show
inputs available + valid SGP suggestion — and is the actionable MVP
target. As each converted deal generates clean data, eligibility
expands and the savings compound.

Not in scope (deliberate):
— Universal Vs deal calculator: disputes are not an arithmetic
problem; Mariana already has a calculator
— AI extraction of freetext deal notes: hides ambiguity rather
than eliminating it
— $5–15K deals: complexity requires human judgment; 37-show sample
is too thin for safe suggestions
— Agentic advancing: requires clean input data; Phases 1 and 2
produce that data first

What shipped

Smart Guaranteed Price (SGP)
7-step deterministic waterfall: resolve expected gross (source
waterfall: artist-at-venue → artist-anywhere → agent-history →
cell-mean → venue-mean → capacity-proxy) → deduct 10% ticketing
fees → cap expense estimate (lower of signed cap and per-bracket
default: $1,500 for $1–5K) → compute net base → project percentage
payout → winner = max(guarantee, payout) → roundTo50.
Suggested flat ≥ contract guarantee always — converting never
underpays the signed number. Full auditJson on every suggestion.
Confidence tiers A–D (deterministic, no model) gate Smart Switch.

Smart Switch
vs/$1–5K: SGP flat at A/B confidence; guarantee_amount fallback
at C/D; tier D emits no suggestion and defers to the booker.
Door (any bucket): $500 artist floor + 60/40 split above $1,500
expense cap. Provenance badged per suggestion.

Improve Deal
Caps-only suggestions (expense cap, hospitality cap) for already-
supported deals. Does not convert deal types.

7 tabs
Shows, Artists, Reports, Deal Analysis, Needs Attention, Insights,
Settings — all computed live, nothing mocked.

Insights pipeline
Two-stage LLM: per-settlement summary (concurrency 8, idempotent)
→ per-cell complaint clustering ≤5 themes. Switch-Savings backtest
replays engine against last 12 months to compute money and time
saved per deal.

Phase 3 — settlement protection (post-MVP, optional)
Product 1 (Instant Dispute Pay): licensed embedded insurance
partner; 4.37% × $345 avg = $15.07 expected cost; $30/show
one-sided; tiered by source deal type dispute rate.
Product 2 (Suggestion Accuracy Pay): Greenroom-funded beta; Basic
$400 / Standard $1,100 / Pro $1,900 caps covering 29.5% / 57.1% /
80.1% of historical gaps; Pro required for door conversions (mean
gap $2,097 exceeds Standard cap).

Data

503 settled shows · $3.20M gross · 4.37% dispute rate (22/503)
SGP backtest: 12-month · n=156 · median gap $870 · p75 $1,569
Commit: 30ca767

Links

App: https://greenroom-starter.replit.app/
PM memo: https://claude.ai/public/artifacts/339ab5c7-8c8a-4ca8-a643-bca6ae757d8a
Supporting docs: https://claude.ai/public/artifacts/74d18ae1-0be8-4728-a88c-cb0b09118830
Demo (Loom): https://www.loom.com/share/7af891243260407bb42da7ce70a56607