bending-spoons-ontology

# Bending Spoons — a queryable world model  > **What this is** — a demonstrative exercise by **Play New**: a worked example of *how we model an organization* as a queryable ontology + intelligence. The subject is Bending Spoons, built **only from its public filings** (Form F-1, Jun 2026; 424B4, Jul 2026) **as of those dates** — a **point-in-time snapshot, not a maintained or live product**, and **not affiliated with or endorsed by Bending Spoons**. The method is the point; Bending Spoons is the example. ## TL;DR — what is this, and what can I do with it? Bending Spoons acquires proven digital products (Evernote, WeTransfer, Vimeo, AOL…), transforms and operates them, and reinvests the earnings into the next acquisition — an operating *machine*, not a portfolio of apps. **This repo is a queryable model of how that machine works** — built only from official documents — **plus the machine's own moves as tools you can run.** ### Connect it to Claude — step by step (no coding needed) You do this once; it takes a few minutes. Never set up an MCP server before? Follow all five steps. **1 · Get the code onto your computer.** With git: ```bash git clone https://github.com/Play-New/bending-spoons-ontology.git cd bending-spoons-ontology ``` No git? On the GitHub page click the green **Code** button → **Download ZIP**, unzip it, and open a terminal in that folder. **2 · Make sure you have Python 3.10+.** Check with `python3 --version`. If it's missing, install it from [python.org](https://www.python.org/downloads/) (tick "Add Python to PATH" on Windows). **3 · Install the three small dependencies** (from inside the folder): ```bash python3 -m pip install -r mcp/requirements.txt ``` **4 · Tell Claude where the server is.** Pick your app: - **Claude Code** (the terminal tool) — one command, from inside the folder: ```bash claude mcp add bending-spoons-ontology -e RADAR_CONTACT=you@example.com -- python3 "$(pwd)/mcp/server.py" ``` (`$(pwd)` fills in the full path for you.) Done — start `claude` and ask. - **Claude Desktop** (the Mac/Windows app): 1. First get the folder's full path: run `pwd` in the terminal (you're inside the folder) and copy what it prints. 2. In the app open **Settings → Developer → Edit Config** — this opens the file `claude_desktop_config.json` in your editor. 3. Paste the block below inside it, replacing the path with the one from step 1. **If the file already has an `"mcpServers"` section, add just the `"bending-spoons-ontology": { … }` part inside it** (don't create a second `mcpServers`): ```json { "mcpServers": { "bending-spoons-ontology": { "command": "python3", "args": ["/PASTE/YOUR/PATH/HERE/bending-spoons-ontology/mcp/server.py"], "env": { "RADAR_CONTACT": "you@example.com" } } } } ``` 4. **Save the file, then fully quit and reopen Claude Desktop** (not just close the window). A 🔨 tools icon appears when it's connected. > **What's `RADAR_CONTACT`?** Put your email there. The model tools — query, capabilities, hire, audit — work **without** it; only the *live* ⚡ radar tools that scout **SEC EDGAR** need it, because EDGAR requires a declared contact in every request and returns **403** otherwise. Not scouting live US data? Leave the placeholder; everything else still works. **5 · Check it worked.** Ask Claude: *"What Bending Spoons tools do you have?"* — it should list ~27. Stuck? The most common cause is a wrong path in step 4 — make sure it's the **absolute** path (starts with `/` on Mac/Linux, `C:\` on Windows) and points to `mcp/server.py`. Once connected, just ask, in plain language: - **Understand it** — *"Show me Bending Spoons' portfolio and how they transformed each business."* · *"How do they decide what to acquire, and what do they pay?"* · *"How does the acquisition flywheel work?"* - **Scout like they do** — *"Find EU-listed companies with $50M–$5B revenue I could acquire."* · *"Check the gates for Sonos and Pinterest."* · *"How strong is Evernote's brand right now?"* - **Run the deal machine** — *"Screen this target, then underwrite it against the 65/25 hurdle and show me the walk-away price."* It **proposes**; you **approve** the write (the human gate is built in — nothing is written without your yes). - **Hire like they do** — *"Draft an advisory assessment of this candidate against Bending Spoons' published hiring gates."* (decision-support only — a human decides; candidate scoring is high-risk under the EU AI Act, so it recommends and stops, it never picks) **More questions to try** — a full, tested bank (grouped by intent, one per tool, heavy on the capabilities) is in [`docs/try-it.md`](docs/try-it.md). Two rules make it trustworthy: **every fact is cited to an official document** (fact vs. thesis is never blurred), and **nothing changes the model without a human approving it.** It is not a slide deck — it is a runnable model with guardrails. The rest of this README is the how and the why. --- ## One model, one analysis A rigorous, public world model of Bending Spoons — a case study of an AI-native operating model — built only from official sources: the public F-1 and IPO record, and Bending Spoons' own published documents. It is a **world model** you can query, plus **capabilities as invocable skills**: the same engine, not an analysis handed to you. The repo separates two things that must never blur: - **the model, at the root** (`ontology.md` · `world-model/` · `capabilities/` · `interfaces/`) — Bending Spoons as the filings show it today — the acquisition machine, products-as-assets, the capital-flywheel. Every node in it is `status: confirmed`, cited to the line — except the one declared gap (`status: proposed`). - **`bending-spoons-as-an-intelligence.md`** — **the analysis**: the repo's one argued thesis — how the same subject would reorganize if the value migrates from products to knowledge (the graph, the unbundling, the knowledge-flywheel). `status: proposed`, an essay by design: a thesis in essay form cannot be mistaken for a fact in model form. The model states; the analysis argues. What connects them is the gap the model itself declares (`world-model/customer/cross-product-graph.md`). ## The reading **The work.** Bending Spoons went public in 2026; its F-1 and 424B4 are among the most detailed public descriptions of an AI-native acquisition machine that exist. This repo turns that corpus into something you can *interrogate* rather than just read: a typed world model (objects, properties, links), its operating verbs as runnable capabilities behind human gates, and an MCP server that serves both. Every fact is anchored to a line in an official source (the filings, or Bending Spoons' own published docs); every inference is marked; a runnable audit refuses to let the two blur. Two readings of the same company come out of it — one the record proves, one the repo argues — and the discipline is to keep them visibly apart. **Bending Spoons as it is — the model.** Bending Spoons is a machine for buying mature digital businesses and running them for cash. It acquires proven consumer/prosumer products — often past their growth peak, undermonetized, with loyal bases (Evernote, WeTransfer, Vimeo, Eventbrite, Brightcove, komoot, AOL…) — at disciplined prices, then centralizes them on a shared Platform of proprietary technology and a small, durable core of operators. The economics compound: revenue $387M → $1.31B (FY2023→FY2025), adjusted operating margin 36% (FY2023) → 51% (Q1 2026), while the same acquisition hurdles (65% levered / 25% unlevered IRR) held as capital deployed scaled ~10×. It is financed by debt and retained cash — no dividends, everything reinvested — so GAAP net income sits near zero under the interest load, and net debt runs ~$3.6B. Products are treated as impermanent assets (100% → 24% of revenue in eight quarters); "never sell" is a strategy, not a charter law. Discipline is the moat: margin of safety, patience, no bid escalation, a ≤4.00 leverage covenant, and founder voting control (82.71%). This is the model at the root — every node `confirmed`, cited to the line. **Bending Spoons as it could be — the analysis.** Read the same company not as a portfolio of products but as an *intelligence*: a world model of its assets and customers, a set of capabilities, an intelligence layer that composes them, and interfaces where people act. Seen that way, one thing is conspicuously missing. The 500M+ monthly users are a *sum of silos* — no unified account, no cross-product identity — so the customer signal that should feed sourcing and monetization is thrown away every turn of the flywheel. The thesis (`bending-spoons-as-an-intelligence.md`, `status: proposed`): as AI commoditizes building and operating products — the very edge the machine sells — the durable asset stops being any product and becomes the **unified, per-person knowledge no single product can hold**. The will does not change; the thing being compounded shifts from *capital across businesses* to *knowledge across people*, and the capital-flywheel's leak becomes a knowledge-flywheel. The F-1 shows no plan to do this — which is exactly why it lives in one declared essay, not a second model. ## Where this is going — the agentic org The end state this model is built for is not a document you read but an **organization that can run**: press play and the machine sources, screens, underwrites, closes, transforms, hires, and finances — each step a human-gated action over the world model, executed through the two-phase engine (`mcp/engine.py`). The **agentified capabilities** are the body; the **feedback loops** are the nervous system that lets it learn. A capability that only executes is a tool; a capability whose **outcomes feed back to improve it** is an intelligence. Every turn of the flywheel (`ontology.md §3`, loop ②) is meant to close: - every **target → deal → business** teaches which signals predict a strong business → recalibrates sourcing and underwriting; - every **business that grows** teaches which transformation levers worked → refines the playbook; - every **hire who succeeds** teaches which selection signals predicted performance → sharpens the gates (Bending Spoons' own "hiring as a science", `sources/hiring/`). Mechanically, each loop is an **outcome reading** (a function: realized-vs-expected) feeding a **calibrate step** (adjusting the capability's own parameters — the thesis, the priors, the playbook, the gates). That write-back closing the loop is what *would* turn loop ② from a diagram into a running mechanism, and make the org *self-improving* rather than merely repeatable. This is the **direction**, marked `[decision]`, not a built engine verb: it is modeled from Bending Spoons' observed outcome-informed discipline (the 65/25 IRR hurdles held stable across a ~10× scaling of deployed capital, `bsp-f1` ~L110-111/~L334; the talent process run as a science, `sources/hiring/`), and a genuinely data-driven calibration needs *longitudinal* outcome data a single-snapshot filing does not yet disclose. What is real today: the outcome-reading side (`monitor` reads held businesses vs their underwriting; `retire` consumes its drift finding). The calibrate-back-to-sourcing side is the specified next turn, and it points at the customer-signal loop the analysis argues — the **knowledge-flywheel** (`bending-spoons-as-an-intelligence.md §5`). This is the north star; the repo builds toward it capability by capability, each behind its gate. ## Layout ``` bending-spoons-ontology/ ├── README.md this ├── CLAUDE.md the modeling discipline (governing rules) + status ├── foundations/ the two contracts — ontology.md (structure) · org-as-an-intelligence.md (organization) ├── will.md the volere (SHARED): problem → mission → intent → constraints + governance ├── AGENTS.md the intelligence layer: how the model composes capabilities into action ├── sources/ immutable captures (SHARED): the F-1 (bsp-f1) + 424B4 (bsp-424b4) filings · Bending Spoons' own docs (hiring/) · press (non-fact) ├── mcp/ the Ontology SDK: server.py (27 tools — query · two-phase engine · radar · hire · audit) · audit.py + audit-contract.md (the model's service tools) ├── bending-spoons-as-an-intelligence.md THE ANALYSIS (proposed): the value-migration thesis, in essay form │ ├── ontology.md THE MODEL's index (confirmed): §1 objects · §2 actions & functions · §3 flywheel ├── world-model/ OBJECTS (+ labelled backing datasets) │ ├── company/ platform · tools/(×5) · spooners · founders · deal · facility · businesses/(×17) + the backing csvs │ └── customer/ market-of-targets (Target) + targets.csv · cross-product-graph (the gap) ├── capabilities/ KINETIC — actions/ (source·screen·underwrite·close·transform·talent·retire·finance) · functions/ (monitor·the readings) └── interfaces/ Interface ×18, one node each (+ interfaces.csv): the app the user touches, of→Business ``` ## The method — reapplicable This repo is one instance of a **reusable pipeline**: everything in `foundations/` (the two contracts, the seven-field action contract, the three-field function contract, the eleven audit invariants) is org-independent; everything else is the per-subject content it produced from one corpus. To model another subject, keep the engine, change the fuel: ``` 0. sources/ capture the corpus — immutable, line-addressable (~L) 1. foundations/ the TWO contracts (fixed): structure (the ontology grammar) · organization (the org as an intelligence) 2. will.md read who the subject is: problem → mission → intent → constraints 3. ontology.md declare the schema: §1 objects+properties+links · §2 actions+functions [only the human amends] 4. world-model/ fill: nodes + backing datasets 1:1, every fact with provenance + markers + the two clocks (as_of = period of the data · last_synced = last verified vs sources) 5. capabilities/ contract the verbs (7 fields) and the readings (3 fields) 6. audit enforce the invariants; every defect found by hand becomes a standing check 7. interfaces expose: MCP — nodes via read-only query tools, actions behind the gates 8. analysis (optional) argue the subject's next shape — as a declared essay, never a second model ``` The running system stacks like the layer diagram it is modeled on — data at the bottom, security wrapped around it, the ontology as the plane humans and agents actually work on: ``` humans + agents (Claude, MCP clients — judgment at the edge) ──────────────────────────────────────────────────────────────── THE ONTOLOGY PLANE world-model + capabilities + interfaces (moved by will.md) + the analysis ──────────────────────────────────────────────────────────────── SECURITY provenance (~L) · markers · the will's gates (foundations/ontology.md §6) ──────────────────────────────────────────────────────────────── DATA sources/ (the immutable filings) ``` The MCP server (`mcp/`) is this repo's Ontology SDK: the door through which the top layer reaches the plane. ## Where to start Read `CLAUDE.md` for the modeling discipline and `foundations/` for the two contracts it rests on (the ontology grammar; the org-as-an-intelligence model). Then `will.md` (the shared volition) and `AGENTS.md` (how it acts). Then `ontology.md` → `world-model/` → `capabilities/` for the model as it is, and `bending-spoons-as-an-intelligence.md` for the analysis of where it could go — and why that is an essay, not a second model. ## Sourcing discipline (why it is publishable) Facts come from **official sources** only, in two tiers (CLAUDE.md rule 8), each captured under `sources/`: 1. **The SEC filings** — the primary record: filing facts cite `bsp-f1` (the Form F-1) with a `~L` line reference against the full-text capture; **IPO-pricing facts cite `bsp-424b4`** (the final 424B4 prospectus) with a `~L` ref — the offer price $29.00/share, the share counts, and net proceeds all live there. 2. **Bending Spoons' own published documents** — company-authored primaries for what only they document, cited by id (e.g. `bsp-selection-process`, `bsp-talent-formula` in `sources/hiring/`, which ground the `hire` capability's selection gates). Captured as facts + short fair-use quotes, never verbatim (they are copyrighted). A few interface brand/feature facts carry a third, weaker provenance — `official-site` (a live company page, cited by URL, not line-addressable) — honestly tagged and never load-bearing for a model-critical figure. The transformation theory lives in one declared analysis file (`bending-spoons-as-an-intelligence.md`), the system's own thesis, never asserted as a fact about BS. **Third-party press is never a fact source**: the press captures (`ipo-2026-07`, `press-turnaround`) are not sources about Bending Spoons — `ipo-2026-07` is superseded for pricing by `bsp-424b4` and survives only for the first-day trading print. Anything no official source supports — the implied ~$18.4B valuation (mark `[derived]`, from price × shares), first-day trading, layoff percentages — is marked `[derived]`, `[to-validate — press only]`, or `status: proposed`, never asserted as fact. That discipline is what makes the case defensible in public. ## References Influences — not sources of the Bending Spoons facts, which come only from the official sources above. The method rests on two, one per contract: - **The ontology** — how the model is *structured* (object / property / link / action / function types; the digital twin): the **Palantir Foundry ontology**. → `foundations/ontology.md`. - **The org as an intelligence** — how the subject is *organized* (world model · capabilities · intelligence layer · interfaces): **Jack Dorsey & Roelof Botha, *"From Hierarchy to Intelligence"*** (Block + Sequoia, 2026). → `foundations/org-as-an-intelligence.md`. The **will** — the intention and direction that sits above both — is this project's own addition. ## License **MIT** (see [LICENSE](LICENSE)). It covers Play New's own software and its expression and arrangement of the model — **not** the underlying copyrighted sources. [NOTICE](NOTICE) states what that means: the facts come from Bending Spoons' public filings and its own documents, used with short fair-use quotes, and this repo is a **point-in-time snapshot by Play New, not maintained and not affiliated with Bending Spoons**. It is a demonstrative exercise, so it does not accept pull requests — see [CONTRIBUTING.md](CONTRIBUTING.md) (issues welcome; fork freely).