Neutron knowledge bundle
Bundle OKF 0.2 · 4 conceitos · DealDeploy/sunshine-equity-brain
Open source Repository Open in the app JSON README (API)
About
# Neutron knowledge bundle
- [Documents](./documents/) — summaries, by type
- [Sources](./sources/) — the converted originals
- [Companies](./companies/) — company pages
- [People](./people/) — person pages
Details
- Kind
- OKF bundles
- Topic
- No topic detected
- Publisher
- dealdeploy
- Origin
- okf_github
- Category
- dados
- Version
- 0.2
- Last push
- 2026-09-08T14:43:06Z
- Repository state
- ativo
- Added
- 2026-09-08 02:24:34
- Updated
- 2026-09-08 02:24:34
- Origin id
DealDeploy/sunshine-equity-brain:neutron/index.md
README
# The firm brain repository
Everything an agent needs to know to operate inside Sunshine Equity Partners, and everything it runs. One repo, organised by **what the reader is about to touch**: a person, a company, a system, a task.
Names and figures in this repository are demo data.
## The tree
```
README.md # this file
AGENTS.md # agent entry point (CLAUDE.md points here)
INDEX.md # one when-to-read line per file — start here
overview.md # mandate, funds, systems of record, the agent persona
glossary.md # firm terms — prefer these over improvised PE jargon
investment-criteria.md # what we invest in
signal-criteria.md # what counts as evidence worth capturing
data-handling.md # confidentiality, approved channels, AI use
agent-conduct.md # conduct rules for any agent — read once per session
team/ # one page per person
portfolio/ # one page per holding
theses/ # firm opinions, loosely held — one file per theme
systems/ # everything about one system in one place
├── affinity/ # rules.md · field-guide.md · pipeline-stages.md · mechanics.md
├── sharepoint/ # document-map.md · mechanics.md
├── dealdeploy/ # executor.md · screener.md · feedback.md
├── files/ # docx.md · xlsx.md · csv.md
├── outlook.md granola.md grata.md exa-websets.md
skills/ # one folder per skill — what agents execute
└── <name>/ # SKILL.md (+ templates/, examples/, scripts/ as needed)
vdr/ # one indexed map per active data room
neutron/ # generated · document bundle published from the wiki
```
## What each part is
- **Root files** — firm-wide facts and rules that don't belong to one system or one task. Short, flat, findable by name.
- **`team/`, `portfolio/`** — curated pages: who someone is, what we know about a holding. `neutron/people/` and `neutron/companies/` are the generated evidence trails for the same names; cross-link, never merge.
- **`theses/`** — one file per theme with stance, `status` (`exploring / active / cooling / retired`), an as-of date, the conditions the view rests on, and a dated revision log that only grows. Agents adopt these as priors and surface contradicting evidence — challenging a thesis is the intended use. Beliefs aren't machine-refreshed, so discount anything long unreviewed and say so.
- **`systems/`** — one folder (or file) per system the firm uses. Rules on how to behave in it, the IDs and enumerations it exposes, what those mean, and the mechanics of driving it through the agent's tools all sit together, so an agent about to touch Affinity opens `systems/affinity/` and nothing else. Two provenance rules inside: enumerations (field IDs, option sets, list inventories) are pulled from the live system with an as-of date, never hand-typed; semantics (what a stage *means*, what counts as "active") cite a named human's answer. An ID with no semantic entry is *undefined* — say "meaning unclear" rather than guess. `mechanics.md` files and the `dealdeploy/` and `files/` docs describe tool behaviour and are maintained with DealDeploy; never quote them in human-facing answers.
- **`skills/`** — one folder per skill. `SKILL.md` carries the trigger frontmatter, a *Read first* list of the system docs the procedure depends on, and the procedure itself in task language ("search the Outlook calendar", "update the expert tracker"). Templates to fill, gold-standard examples to imitate, and firm-owned scripts that coevolve with those templates live in the same folder. Prescriptive but defeasible: follow by default, deviate with a stated reason. `.claude/skills` symlinks here so Claude-Code-family harnesses pick skills up automatically. Observed facts (a list name, a field, a count) are hoisted into `systems/` and linked; only *chosen* values — caps, defaults, thresholds — stay inline.
- **`vdr/`** — one indexed map per active data room: deal snapshot, section table, reading spine, and the full listing. Grep the listing; don't read it end to end. Maintained by `skills/vdr/`; retired rooms move to `vdr/retired/` with a closing note.
- **`neutron/`** — every external document the firm holds, published from the wiki. Check here first for anything about a portfolio company, a deal, a meeting, or a fund. Start at `documents/<type>/` summaries; open `sources/` when a summary isn't enough. Summaries are generated and unratified — when a claim is load-bearing, read the source. Regenerated wholesale on every publish; never hand-edited.
## Filing rules
1. **Index or invisible.** Every file outside `neutron/` gets a when-to-read line in `INDEX.md`.
2. **One fact, one home.** Every number, rule, and definition lives in exactly one file; everything else links to it.
3. **File by what it's about.** A fact about Affinity goes in `systems/affinity/`, whatever kind of fact it is. A procedure goes in its skill. Only firm-wide, system-agnostic material sits at the root.
4. **One question per file, one to two pages.** Filenames are the retrieval interface: kebab-case names the firm already uses, never invented taxonomy.
5. **Provenance in `systems/`.** An enumeration is a machine pull with an as-of date; a semantic claim cites a named human. Content with neither waits for ratification.
6. **Everything is ratified.** Agents draft; a human owner merges. There is no scratch space in the repo.
7. **Every file carries an owner and a review date.** Frontmatter: `owner`, `last-reviewed`, and for theses `status`. Agents discount anything long unreviewed — and say so.
8. **Never hand-edit `neutron/`.** It is regenerated on every publish.