cnx_vinyl/docs/catalogue-intelligence.md

Own Catalogue: Sales Map, Demand Forecast and Market Advisor

The idea (Alexander, 26 October 2026): if CNX owns a catalogue it can press, its sales by shop, region and location can drive a model that says what to press next. That needs three things: a catalogue dashboard with a map of where the records sell, a demand model that turns those sales into press decisions, and an advisor agent that points at market opportunities. This document plans all three on the systems the plant already intends to run. It is a planning position: every figure without a source is an estimate, and nothing here is in the business plan's revenue.

Why a pressing plant should own a catalogue

• It fills press time the plan does not sell. At the planned 60 % specials, specials run only on the Duplex system, and the Classic runs at about two thirds of its ~128 records an hour (revenue model). The spare third is 21 to 48 records an hour; at the plan's 60 to 70 % effectiveness, 16 hours and 300 days that is ~60,000 to 160,000 records a year (estimate), black and solid colours only. At the calculator's planning rate the launch plant is balanced and that spare time disappears, so the catalogue is a filler for idle hours, not a reason to buy presses.
• A repress is the plant's shortest and cheapest order. The father and mother of every side are archived, so a repress skips cutting and the three artist gates and takes ~7 to 11 working days; a mother gives about ten stamper sets and a father 5 to 10 mothers (production timeline). A label that has to order from a plant abroad presses big runs months ahead and carries the stock. A label inside a plant with its own metal archive can press small runs often, against measured demand. That is what makes a demand model worth having: it is asked every few weeks, for a few hundred records, not once a year for thousands.
• There is a market for Thai catalogue abroad, and it runs through Japan. EM Records (Osaka) reissues luk thung and molam compiled by Soi48, many of them first official reissues (EM1185LP, EM1198DLP); Bangkok's ZudRangMa Records co-released an Angkanang Kunchai compilation with EM Records and Soi48 that launched in Japan with the singer present (Bangkok Post). Japanese labels shipped 3.38 million records in 2025, and Japan is short of pressing capacity (revenue model). Contract pressing for Japanese labels now competes with ResurRec's new Japan plant; selling Thai catalogue into Japan does not.
• The reverse has been done. P-Vine, a Japanese label, built its own plant (VGAP, 2024) because domestic plants took six months to a year and pressing abroad about six months, too slow for new releases. P-Vine also said that, having made records themselves, it found the ~¥4,000 a new record sells for "tight" (Phile-web, 15 October 2024, BRUTUS, 2025). A catalogue earns the retail or wholesale price instead of a pressing fee, and carries stock, rights and marketing in exchange.

What counts as our catalogue

Kind	What it is	In scope
Poy Records releases	Titles whose masters the label owns; SST in Frankfurt cut lacquers for POY003, 005, 006 and 013 (equipment suppliers)	yes, the seed and the first data
Licensed reissues	Masters licensed for vinyl: territories, term, sell-off period, royalty per unit, advance	a decision (see Open decisions)
Customers' titles	Records the plant presses for artists and labels	no: their sales are their data

Two rules follow from the third row:

• Customers' data never trains our catalogue decisions. Order sizes, represses and shipping addresses of customer titles stay theirs. A customer can opt in to share aggregate sales; nobody is opted in by default.
• Customer orders come first on the presses. Catalogue runs fill idle slots and never move a customer's date. A label that suspects its plant favours its own releases goes elsewhere.

Which titles Poy Records holds, in which territories and until when, is the first list to write; nothing below can be tested without it.

The questions the system answers

1. What to press next: which title, before it runs out where it sells.
2. How many: a run sized to the demand of the next few months, since the next repress is two weeks away.
3. Which variant: black or a colour, within what the presses can make (catalogue specials compete with customer specials on the Duplex; black and solid colours go on the Classic's idle hours).
4. Where the stock goes: the plant, a distributor, or a shop on consignment.
5. Which titles to add: licensing leads for a human to research.

Data sources

Source	What it gives	Location detail	Access and terms	Lands in
Own web shop, direct shop accounts, event sales (Poy Festival merch table)	Every sale with its ship-to or the event's place	postcode or city; the event venue	ours	ERPNext: Sales Invoice, POS Invoice, Delivery Note
Distributor statements	Units per account (shop) and territory per month, returns	shop or territory	the distribution agreement must require monthly sell-through by account, not only sell-in	ERPNext: Channel Sales Report (new, below)
Consignment shops	Monthly stock count, so sell-through per shop	the shop	the consignment agreement	ERPNext: Channel Sales Report, consignment warehouse per shop
Bandcamp	Raw sales report with the buyer's ship-to city, state, postcode and country for physical items; also as JSON through the Sales API (help, API)	postcode or city	the label's own account	ERPNext: Sales Invoice per Bandcamp payout with lines, or Channel Sales Report
Record passport registrations	A buyer scans the sleeve code and, opting in, gives country and city: the end listener's place even when the record went through a shop or distributor	city	consent on the form; no IP address stored (traceability)	machine-data store, per-record rows keyed to the record key
Streaming of our own artists	Top countries and cities per artist and per song in Spotify for Artists (Spotify); country totals in the distributor's streaming statements	country, top cities	only for artists we release	machine-data store, weekly series
Interest list and pre-orders	"Tell me when it is pressed" sign-ups and campaign pre-orders on the artist platform	country, city	consent on sign-up	ERPNext: Lead, Campaign, Payment Entry
Discogs monthly data dumps	Every release, version and pressing in the Discogs database: labels, countries, formats, identifiers (data.discogs.com, CC0)	country of release	free to use (CC0)	a local index on the plant core, refreshed monthly
Discogs API: have/want counts, lowest price and copies for sale (API)	The collectors' view of a title	none useful	Restricted Data (marketplace data by name; the have/want counts come from users' collections and wantlists, which the terms also restrict): no commercial use, no storage longer than needed, never shown more than six hours old, no marketing analytics without Discogs' written permission (API Terms of Use, checked October 2026); 60 requests a minute	not stored; live look-ups by a person only, unless Discogs agrees in writing
Market totals	Country production and sales series, e.g. RIAJ for Japan (revenue model)	country	public	a context layer on the map
Events	Tours, festival slots, film or ad syncs, anniversaries, Record Store Day	city, date	entered by the team	ERPNext (a dated note on the Release)

Sell-in is not sell-through. A distributor buying 300 copies says nothing about fans until the shops sell them; returns can come months later. The model uses sell-through wherever an agreement provides it and treats sell-in as a weaker signal.

Stockouts hide demand. Weeks in which a title was out of stock at a shop record zero sales, not zero demand. Those weeks are marked censored, and the interest list and passport registrations carry the demand the till could not.

Where the data lives (ERPNext touchpoints)

The plant's standing rule holds: ERPNext is the only system of record, our tools read and write it, and only machine-speed data stays outside (digital factory).

Object	In ERPNext	Outside (machine-data store)
Release, side, item per format and variant	existing custom Release and Side documents; Item per release, format and variant	none
Rights	new Licence document in cnx_erp: licensor, territories, term, sell-off end, royalty basis and rate, advance and recoupment; Release links to it; "owned" or "licensed" on the Release	none
Royalties	royalty accrual per unit sold, posted as journal entries; a royalty statement print format per licensor	none
Places	Address gains latitude and longitude of its postcode or city centroid (custom fields, filled by the geocoder); shops are Customers in a "Shop" group with their exact location; regions are ERPNext Territories	none
Sales	Sales Invoice, POS Invoice, Delivery Note	none
Third-party sales	new Channel Sales Report: source (distributor, Bandcamp, consignment count), period, lines of item, place (shop, territory or city), units sold, returned, on hand	none
Stock	Warehouses at the plant, at each distributor and one per consignment shop	none
Metal	Serial No per father, mother and stamper with sets used per mother (already planned)	metrology raw files
Recommendations	new Catalogue Recommendation: title and variant, quantity, the forecast quantiles it rests on, the reason, status (proposed, accepted, rejected), link to the Work Order it became	the full forecast rows per title, place and month
Market opportunities	CRM Opportunity and Lead (new shop, distributor, territory, licensing lead)	none
External signals	none	weekly streams per country and city, passport registrations, Discogs dump index, map aggregates

The forecast rows and map aggregates are derived and can be rebuilt from ERPNext and the signal series at any time; nobody types into them. Decisions and their outcomes are ERPNext documents, which is what lets the model and the advisor be scored later.

flowchart LR
    subgraph S[Sources]
        SHOP[Web shop, shop accounts,<br/>event POS]
        DIST[Distributor and<br/>consignment statements]
        BC[Bandcamp sales report]
        INT[Interest list,<br/>pre-orders]
        PP[Record passport<br/>registrations, opt-in]
        STR[Streaming by country<br/>and city, own artists]
        DG[Discogs CC0 dump]
    end
    ERP[(ERPNext<br/>releases, licences, sales,<br/>stock, channel reports)]
    MD[(Machine-data store<br/>signal series, forecast rows)]
    FC[Forecast job]
    DASH[Catalogue dashboard<br/>and sales map]
    AG[Market advisor]
    TEAM[Team decides]
    SHOP --> ERP
    DIST --> ERP
    BC --> ERP
    INT --> ERP
    PP --> MD
    STR --> MD
    DG --> MD
    ERP --> FC
    MD --> FC
    FC --> MD
    FC -->|recommendations| ERP
    ERP --> DASH
    MD --> DASH
    DASH --> AG
    AG -->|drafts: recommendations,<br/>opportunities| ERP
    ERP --> TEAM
    TEAM -->|repress work order| ERP

Personal data

Buyers' names and addresses are personal data under Thailand's PDPA (B.E. 2562, fully in force since 1 June 2022; overview), and so is a sales location fine enough to point at one buyer. The map is built so that it never needs a person:

• addresses are geocoded to the postcode or city centroid, never to the house;
• a map cell shows a count only when it holds at least five buyers; smaller cells merge into the surrounding region;
• only shops and venues, which are businesses, appear as points;
• passport registrations ask for country and city, with consent, and store no name, e-mail or IP address unless the fan asks to be told about represses;
• the advisor sees aggregates only; no personal data goes into a prompt.

The catalogue dashboard

One internal tool for the label team, with five views.

1. Catalogue. Every title and variant: owned or licensed, licence territories and expiry, stock by place (plant, distributors, each consignment shop), sell-through over 4, 13 and 52 weeks, weeks of cover at the current rate, repress-ready (father and mother archived, stamper sets left on the mother from ERPNext's counts), last pressing, royalties accrued, and a flag when a recommendation is open.
2. Sales map. Sales density by city or hexagon over time, with a time slider and filters by title, genre, variant and channel. Layers: own sales; shops with our stock (pins sized by stock, coloured by weeks of cover); distributor territories; streaming listeners of our artists; passport registrations; interest-list sign-ups; country totals (RIAJ and others) as context. A white-space layer marks places where listeners, sign-ups or registrations are high and sales or stock are low: the map's version of a market opportunity.
3. Title page. The sales curve with out-of-stock weeks shaded, the forecast fan for the next months, the title's own map, its metal genealogy and its licence clock.
4. Repress planner. Proposed runs against free press slots from the factory scheduler (customer work orders already placed, catalogue in the gaps), with the Classic's idle hours as the default home for black and solid colours.
5. Advisor inbox. The advisor's notes, each with its numbers linked to the rows they came from, and accept or reject buttons that create or close the ERPNext document.

How it is built. Same stack and rules as every other tool (digital factory): Rust (axum, minijinja, HTMX) in cnx-factory, its own ERPNext integration user, deployed by the clan, internal only (office and the team over the mesh, not public). The map is open source and self-hosted, in line with the plant's open-source decision (digital factory):

• MapLibre GL JS (BSD-3-Clause) in the browser;
• vector tiles from OpenStreetMap data (ODbL, attribution on the map) as one Protomaps PMTiles file (BSD-3-Clause tooling) served from the fleet, so no Google Maps or Mapbox account;
• H3 hexagons (Apache-2.0) for binning;
• geocoding offline: postcode and city centroids from GeoNames (CC BY 4.0), with a self-hosted Nominatim only if centroids prove too coarse.

The demand model

The model grows with the data. Each stage has to beat the simpler one before it is allowed to decide anything.

Stage 0: no sales history (a new title or a reissue candidate). Demand is borrowed from comparable titles (genre, era, the artist's streaming, how many pressings already exist per the Discogs dump) and tested cheaply: a pre-order campaign on the artist platform for our own title is the experiment. If the pre-orders reach the minimum run, press; if not, refund and learn. The same mechanism the platform offers artists becomes the label's market test.

Stage 1: the first months of sales. A reorder rule, not a model: sell-through per title and place per week, with out-of-stock weeks left out of the rate. Repress when the stock that will be left at the end of the repress lead time (~7 to 11 working days, plus transport to where it sells) falls below a safety stock. Dashboard view 1 shows this from the first sale.

Stage 2: a year of data and a few dozen titles (thresholds are estimates). A hierarchical count model of weekly sales per title and place (negative binomial, with partial pooling across titles of the same genre and across nearby places), so a title with three months of history borrows strength from its neighbours. Inputs: past sales with censoring for stockouts, the artist's streaming in that place with a lag, price, variant, season (year-end, Record Store Day), events, age since release, passport registrations and sign-ups. Output: a forecast distribution (quantiles), never a single number. Runs as a nightly batch job on the plant core (Python with statsmodels or PyMC from nixpkgs, as a systemd timer).

How many to press. With a two-week repress the run only has to cover the next few months, so it is sized by the classic stocking rule: press the quantile of next-period demand at which the margin lost on a missed sale balances the cost of a record left unsold.

• Missed-sale cost: net selling price after the channel's margin, minus direct cost (1.35 to 2.70 USD a record in the revenue model), minus royalty.
• Unsold-record cost: direct cost plus storage and the money tied up, minus what the PVC is worth as regrind after the label punch (pressing line).
• Floor: a run long enough to pay the press setup (half a day per run in the production timeline) and within one stamper's life (1,000 to 2,000 pressings).

Our wholesale and retail prices for catalogue titles are not in any document yet; they are inputs to fill before the rule produces a number.

How it is checked. Rolling back-tests on our own history: do the 80 % intervals contain about 80 % of outcomes, and does following the model give fewer out-of-stock weeks and less stock unsold after a year than the Stage 1 rule did? Accepted and rejected recommendations, with what then sold, are ERPNext records, so this is a report, not a project.

The market advisor

An agent that reads the dashboard's data and writes a weekly note plus answers to questions ("where does molam sell outside Thailand?", "which shops stock out fastest?"). It proposes and never acts: it places no work order, sends no e-mail to a shop or licensor and changes no price. An accepted proposal becomes an ERPNext document created by a person.

What it looks for:

• Repress now: weeks of cover below the repress lead time plus safety, with the run size from the model.
• Move stock: a shop out of stock while another sits on months of cover.
• White space: places with listeners, sign-ups or registrations and no shop or distributor; proposed as a CRM Lead for a shop or distributor to approach.
• Variants: colours that sell faster in a region than black, within what the presses can make that month.
• Events: an artist's tour or festival date in a city where stock is thin.
• Licensing leads: titles in genres that sell for us that, per the Discogs dump, were never reissued or not for decades, with the original label named. A lead for a person to research; the advisor never says that rights are available.

How it is kept honest:

• the numbers come from the model and the database, the words from the agent; every figure in a note links to its row and date;
• it separates our own data from outside signals and says how sure it is;
• it has read-only tools: the dashboard API, the forecast rows, ERPNext reports through a read-only integration user, the local Discogs dump index, this repository's market notes, and web search through a self-hosted SearXNG for news (tours, syncs, competing reissues);
• every suggestion and its outcome is logged, and after six months it is scored against the plain reorder rule; if it does not beat it, it becomes a report instead of an agent.

Which model runs it. By the open-source rule, an open-weights language model served on a fleet machine (llama.cpp or vLLM on NixOS) is the default; the forecasting is ordinary statistics, so the language model only reads, searches and writes. No such GPU machine is in the fleet plan yet (estimate: one workstation-class GPU, a few thousand USD). A hosted model is a fallback to decide on only if the open one proves inadequate, and then only aggregates leave the fleet.

How it fits the plant

• Scheduling. Catalogue work orders enter the same ERPNext schedule as customer orders, flagged as fillers; the factory scheduler places them in idle Classic hours first and never moves a customer date.
• Metal archive. Catalogue fathers and mothers sit in the archive room with everyone else's (facility security); "repress-ready" on the dashboard is ERPNext's serial-number data.
• Traceability. Catalogue sleeves carry the record passport code like every other record, which is what makes the registration layer of the map possible (traceability).
• Artist platform. Catalogue pre-orders use the campaign machinery and its ring-fenced payments; the catalogue's own web shop is a separate storefront on the same public side (artist platform).
• Money. Not in the revenue model or the total investment. A catalogue needs working capital for stock and licence advances, which neither plan carries.

Phasing

Roadmap phase	Step
0, validate (now)	List Poy Records' titles and rights; collect its existing sales (Bandcamp report, shops, festival) into one sheet and draw the first map by hand; decide whether to ask Discogs for written permission
1 to 2, quotes and commit	Decide whether the catalogue is in scope, under which label, owned only or licensed too, and how the working capital is funded
3, build	ERPNext: Licence, Channel Sales Report, Catalogue Recommendation, geocode fields, consignment warehouses; dashboard views 1 and 2 in the factory application; geocoder and tiles on the fleet
4, launch	First catalogue represses in idle slots on the Stage 1 rule; passport registrations live; advisor version 0 (rule-based notes written up by the language model)
5, expand	Stage 2 model once there is a year of data; licensing leads; offer label customers the same dashboard on their own data (opt-in), like the open plant stack, not in the revenue

Rough cost (estimates)

Item	Effort or cost
ERPNext additions in cnx_erp (Licence, Channel Sales Report, Catalogue Recommendation, geocode fields, royalty statement)	0.5 to 1 person-month
Importers (Bandcamp, distributor and consignment statements, streaming exports)	0.5 to 1 person-month
Dashboard and sales map in the factory application	1.5 to 3 person-months
Forecast job: Stage 1 rule, then the Stage 2 model	0.5, then 1 to 2 person-months
Advisor: tools, prompts, logging, scoring	1 to 1.5 person-months
GPU machine for the open-weights model	a few thousand USD of hardware
Map tiles, geocoder, hosting	on the existing fleet

Risks

• Stock is the risk a pressing plant normally avoids. Small, frequent represses keep it small; the model's job is to keep it smaller.
• Small data. A few dozen titles is little to learn from; the stages above exist so that no complex model decides before it has beaten a simple rule.
• Data access. Distributors and shops report sell-through only if the contract says so; that clause goes into every agreement from the first.
• Platform terms. Discogs' marketplace and wantlist data are restricted; the plan uses its CC0 dumps and live look-ups by a person unless Discogs agrees in writing.
• Conflict with customers. The scheduling rule and the data rule above have to be written into the customer terms, not only into this document.
• Distraction. A label is a second business. The catalogue stays a filler for idle press time until the plant's own order book is validated (roadmap item 1).

Open decisions

1. Catalogue in scope, and under which name: Poy Records (the seed), CNX Vinyl, or a new imprint.
2. Owned titles only, or licensed reissues too. Licensing adds advances, royalty accounting and territory limits.
3. Working capital for stock and advances: how much and from where; neither plan carries it.
4. Discogs: ask for written permission to use have/want and marketplace data commercially, or work from the CC0 dumps alone.
5. Language model: open-weights on the fleet (default) or hosted.
6. Label customers: whether they are offered the dashboard on their own data, and on what terms.