Add the 500-1000 word narrative and the evidence checklist the rules require

The official submission email lists a standalone written narrative as its own
deliverable, with a specific brief the docs did not meet: how the business runs
day to day, what the AI does versus a human, and the jobs and economic
opportunity it creates or enables beyond the founding team, both actual and
potential. That last dimension was entirely absent. The honest version is that
Bloom has no employees but me and creates protected livelihood rather than
headcount: the retained salon revenue that local service jobs depend on, plus
Bloom's own future hiring. Written as 696 words, first person, no em-dashes.

Also adds the evidence pointers the email calls for and the docs lacked: a Stripe
export or bank statement as revenue proof (the refunded test charges match the
$0-net story and prove the path is live), a "not incorporated" note for the
corporate-ID field, a customer-contact placeholder that refuses to fabricate, and
an expanded confirmations checklist covering the video, narrative, and private
judge-notes fields.

Author: JonathanSolvesProblems

SUBMISSION.md

@@ -199,6 +199,77 @@ sync so the book refreshes itself instead of waiting for an upload.
 
 ---
 
+## Written narrative (500 to 1000 words)
+
+> The submission form asks for a standalone narrative on how the business was
+> built and how it runs day to day, explicitly covering what the AI does versus
+> what a human does, and the jobs and economic opportunity the business creates or
+> enables beyond the founding team, both actual and potential. This is that piece,
+> ready to paste. It is about 720 words.
+
+I set out to build an AI that writes marketing for local businesses, and I was
+wrong about the problem. Content is free now: any salon owner can get a week of
+captions out of a chatbot in the time it takes to make a coffee. I was selling
+something the market already gives away.
+
+The real problem was hiding in the retention research. In a salon or barbershop,
+nobody cancels. They just quietly stop coming, and the owner finds out months
+later, from a gap in the appointment book that nobody had time to notice. About
+40% of new clients are gone within a year. A first-timer who does not rebook
+within 30 days has roughly a one-in-five chance of ever returning. A regular is
+worth $1,200 to $3,000 a year. That is the largest single profit leak most shops
+have, and it stays invisible because it is an absence, spread across hundreds of
+people who each keep their own rhythm.
+
+Bloom reads that rhythm. An owner exports a booking-history CSV from whatever they
+already use (Fresha, Square, Vagaro, Booksy, even Google Calendar) and uploads it.
+Nothing to migrate. Bloom takes each client's own median gap between visits and
+measures them against themselves, not against a generic rule. The homepage makes
+the whole case with two clients: Aisha and Jane were both last in 44 days ago.
+Aisha comes every eight weeks, so she is fine and Bloom stays silent. Jane comes
+every four, so the identical gap means she is leaving. Any tool that flags "no
+visit in 60 days" gets one of those two wrong, every time.
+
+**What the AI does, and what a human does.** This is the part I am most deliberate
+about. A rule decides WHO is at risk, because that call has to be exact,
+reproducible, and auditable: the output is "she is worth $1,295 a year and she is
+leaving," and a model that hallucinated a rhythm would cost a real owner a real
+client. Gemini decides WHAT TO SAY, because writing a note that sounds like the
+owner remembered someone is a judgement over messy human context, which is exactly
+what a rule cannot do and an LLM can. Gemini 2.5 Flash runs on Google Cloud's
+Vertex AI for every draft. A separate weekly content agent runs fully
+autonomously: it picks the angle, writes, scores its own draft, rewrites when it
+falls short, and sends. The win-back is deliberately owner-triggered, one client
+at a time, because it sends email in a real business's name to a customer they
+care about, and I will not ship unsupervised sending at scale to a stranger's
+salon before it has earned it. Every action is logged to a public feed that shows
+the agent working without ever naming a client.
+
+**How it runs day to day.** The owner sets up once. The radar is free, because
+showing someone their own losses should not cost anything. When they act, Bloom
+drafts and sends from their verified domain, never twice, never to anyone who
+opted out. When a client books again, the owner's next export proves it, and Bloom
+counts the save from their own data rather than claiming it.
+
+**Jobs and economic opportunity beyond me.** Today the honest answer is that Bloom
+has no employees but me. What it creates is not headcount, it is protected
+livelihood. A neighbourhood salon runs on thin margins and a handful of chairs,
+and every regular who drifts away unnoticed is revenue that would otherwise pay a
+stylist. When Bloom keeps 20 clients a year who would quietly have left, worth on
+the order of $25,000, that can be the difference between a shop cutting a chair and
+keeping one, or hiring the next. The economic opportunity Bloom enables is the
+retained revenue base that local service jobs actually depend on, for people well
+beyond any founding team: the stylists, barbers, and front-desk staff whose hours
+exist only if the shop stays full. As Bloom grows, its own hiring follows the same
+logic, salon-partnership and customer-success roles whose whole purpose is keeping
+more of those chairs occupied. I would rather state that plainly, actual versus
+potential, than inflate a number I do not have.
+
+**Where it honestly stands.** The product is live in production. The gap is
+arms-length paying customers, which I am closing now by getting it into real
+salons. That is the one thing engineering cannot do for me, and it is what I am
+spending the last month of the hackathon on.
+
 ## The uniqueness claim
 
 The one sentence to gate every scope decision against, and the one to lead with in
@@ -598,7 +669,37 @@ Upload a simple P&L (template: https://bit.ly/4w3DvwL) reflecting the figures
 above: revenue $0 net, expenses ~$3, so a small net loss of about $3 for the
 period. Formats allowed: pdf, xls, xlsx, csv, png, jpg.
 
+### Revenue proof to attach (from the official email)
+
+The form wants documentary proof, not just the figures above:
+
+- **Stripe export or bank statement.** Export the Payments list from the Stripe
+  dashboard (Payments to CSV) covering the period. It will show the two July test
+  charges and their refunds, which matches the $0-net story exactly. That honesty
+  is an asset, not a liability: it proves the payment path is real and live.
+- **P&L** using the template (https://bit.ly/4w3DvwL): $0 net revenue, ~$3
+  expenses, ~$3 net loss.
+- **Corporate ID:** none. Bloom is not incorporated; it is a solo project. State
+  that plainly rather than leaving the field blank.
+
+### Customer evidence to attach (from the official email)
+
+The form asks for real customer contact info (name, email, phone) and any
+testimonials. As of now there are none, and I will not fabricate any.
+
+```
+[UPDATE: none yet. As real salons come in, add name, email, phone, and any
+written feedback here. Even one real owner saying "yes, she is exactly the client
+I would have lost" is worth more than any amount of the engineering already done.]
+```
+
 ### Confirmations
 
-- [ ] GitHub repo shared with testing@devpost.com and judging@hacker.fund
+- [ ] GitHub repo public, or shared with testing@devpost.com and judging@hacker.fund
+- [ ] 3-minute (or shorter) video, public on YouTube/Vimeo, showing the agent live
+- [ ] Written narrative pasted (see the 720-word section above)
+- [ ] Revenue proof attached (Stripe export/bank statement + P&L)
+- [ ] Product evidence attached (see docs/EVIDENCE.md + a Vertex AI usage screenshot)
+- [ ] Demo URL and password put in the private notes-to-judges field
 - [ ] No single customer represents more than 40% of revenue (true: revenue is $0)
+- [ ] Related-party revenue disclosed (two refunded founder test charges, $0 net)