An open format for knowledge vaults

Every note is a tile that stands on its own.

The picture exists only in the whole.

Mosaix is a file-level format for vaults of atomic, self-describing notes. Machines retrieve one note at a time. People read many notes as one document.

Nothing is silently lost.

01Why a format

Three documents, three launch dates. Which one is true?

Every knowledge base has this: the last edit wins, and nobody recorded that there was a disagreement. Mosaix records it at the file level, and keeps every version where it was.

Roadmap.docxLaunch: 26 Oct 2026
edited 2026-07-22
Master index.mdLaunch: 25 Nov 2026
edited 2026-08-05
Timing rules.mdLaunch: 27 Jan 2027
edited 2026-08-29
■ current · ▢ superseded, kept in place · ☐ open question
Launch date — online platform27 Jan 2027
✅ sourced · decided with partners 2026-08-29 · rev 2b7c…
Launch date — online platform (v2)25 Nov 2026
superseded → links to the current note · kept in place
Launch date — online platform (v1)26 Oct 2026
superseded → links to the current note · kept in place
Open question #12Three dates recorded on 2026-09-03; consequence: the email schedule shifts by 90 days.
_meta/Open questions.md · R6 record, don't resolve
02What the format asks, and what it gives back

One paragraph in, three tiles out.

The same knowledge, once as prose in a document and once as Mosaix tiles. The body of each tile is the sentence you already wrote. The keys are generated by enrichment and reviewed by a person.

BEFORE · a paragraph in "Market notes.docx"
Portwest, an Irish manufacturer, sells in more than 130 countries and undercuts distributors with its own-brand line. The European PPE market is consolidating: three acquisitions in 2025 moved about 18% of distribution under two groups. A colleague heard that Portwest opened a plant in Albania in 2024, but we could not confirm it.
Portwest — player cardsummary: Irish PPE manufacturer, 130+ countries, own-brand line that undercuts distributors on price.
entities: Portwest (company) · ✅ sourced
PPE consolidation 2025summary: Three acquisitions in 2025 moved ~18% of European PPE distribution under two groups.
relations: Bunzl → owns → distributors · ✅ sourced
Open question #33 — Albaniasummary: Unconfirmed report of a Portwest plant in Albania (2024); blocking for one argument, not for a number.
_meta/Open questions.md · ⚠ to confirm
03The tile

Up close, a tile is a note you can read. And nine keys a machine reads first.

Nothing in the body changes. The keys sit above it, in plain YAML. Hover a key to find it in the note.

  • summaryOne declarative sentence, 120–240 characters.
  • keywordsSix to eight terms: synonyms, spoken phrasings, questions.
  • entitiesNamed things, typed.
  • relationsTyped links between entities. The graph lives in the note.
  • revHash of the body when metadata was computed. Mismatch means stale.
How a tile ages
2026-08-22body written · metadata computed · rev 91be7f0a3c44 ✓ fresh
2026-09-03body edited by a person · metadata untouched · ✗ stale: rev no longer matches
2026-09-03enrichment proposes new keys on a branch · reviewer sees the mismatch · merges · ✓ fresh
--- title: §6 Composed documents updated: 2026-09-03 tags: [spec, mosaix, composed-document] summary: "Defines the note type that joins atomic notes into one readable document without duplicating them." keywords: [composed document, document type, fragments, view over notes, transclusion] entities: - {name: Mosaix Format, type: project} relations: - {from: §6, type: depends on, to: §3 The note} rev: 91be7f0a3c44 --- # 6. Composed documents Atomic notes are for machines. Humans read documents. The format therefore defines … vault/spec/§6 Composed documents.md · a real note of the specification vault
04Composition

The five gold tiles from the mosaic above, read as one document.

A note of type document lists the tiles it is made of, in reading order. Hover a paragraph to see its tile. Nothing is copied.

64 tiles · 5 in fragments, reading order · one still to confirm

Market entry brief — PPE, Europe

type: document · fragments: 5 · composed 2026-09-03 · every paragraph saves back to its own note

1The European PPE market is consolidating: three acquisitions in 2025 moved 18% of distribution under two groups.← PPE consolidation 2025 · ✅ sourced

2Portwest sells in 130+ countries and undercuts distributors with an own-brand line; it is the reference competitor for price.← Portwest — player card · ✅ sourced

3Whether Portwest opened an Albanian plant in 2024 is disputed between two sources; the argument does not depend on it.← Open question #33 · ⚠ to confirm

4A+A Düsseldorf in October is the single most relevant fair; exhibitor lists confirm both direct competitors will attend.← PPE trade fairs, Europe 2026 · ✅ sourced

5Recommendation: enter through safety footwear, where Turkish manufacturing has a documented cost advantage of 22–30%.← Strategic scenarios · ✅ sourced

Every paragraph saves back to its own note.
05The rules

Eight rules. Three make the difference. Five are checked from files.

A vault that obeys them can call itself Mosaix 1.0 conformant. Conformance is a property of the files, not of any tool.

The three that no other format has
The five the checker verifies
  1. Atomicity

    One note answers one question.

  2. Self-description

    Every note carries the core keys.

  3. Connectedness

    No orphans, no broken links.

  4. Declared taxonomy

    Every tag lives in the meta note.

  5. Composition

    Read together? Composed, never copied.

06Why it matters

Small models can read this. That is the point.

When each tile carries its own contract, the size of the model stops being the bottleneck.

Gemma 4 · 2B parameters · 12 GB consumer GPU · cost: none
0/ 47

tests passed by a two-billion-parameter model, generating a 31-unit package one tile at a time. Qwen 3.5 at 9B: 40 of 40.

$0 vs $0.003A frontier model with extended reasoning, for output identical to a flash model. Three hundred times the price, the same code.
0/6 → 6/6Six cloud models on one long prompt: none gets the data types right. On tiles, all six pass 28/28. The contract does the work, not the model.
~0 tokens, flatPer unit, constant across the build. The monolithic equivalent climbs from 544 to 4,055. 5.7× overall, five replications.

Results from the code-generation study that produced the format. Measurements on knowledge vaults are in progress and will be published here. How we measured →

07Why it scales

One agent, one cluster. Any model.

A planner hands each agent one tile and its direct dependencies: about six hundred tokens, nothing else. No agent reads the whole vault; there is no shared memory to keep in sync. So each agent can run on the smallest model that passes its tile, local or cloud, general or vertical, all in parallel.

the vault · 64 tiles
the planner reads the map and picks, for each agent, one tile and its direct dependencies
cluster A
Gemma 4 · 2B · local
its tile + 2 dependencies · ~600 tok
cluster B
Qwen 3.5 · 4B · local
its tile + 1 dependency + 1 open · ~600 tok
cluster C
vertical model · legal
its tile + 3 dependencies · ~600 tok
cluster D
DeepSeek flash · cloud
its tile + 1 dependency · ~600 tok
composed result · four tiles back in the vault · fragments in reading order
no shared memory · no agent reads the vault · all four run in parallel · one model per agent
0models, from 4B to Opus, all 28/28 on the same tiles 0/20tiles right at the first attempt, three vaults 0%end to end, 56 tiles 0×faster, agents in parallel

Because the cluster is the unit of work, the model is a per-agent choice: a 2B model for the tiles it passes, a vertical model where the domain needs it, a cloud model only where nothing smaller will do. Swap one, and the others don't notice. This is the direct evolution of the system: work verticalised across different models, each on its own cluster of knowledge.

08Compatibility

Works where you already write.

Obsidian is the reference host. The format needs none.

ObsidianLogseqFoamDendronZettlrVS CodeAny Markdown editorgit
09Check a vault

Already have an Obsidian vault? You are part of the way there.

One Python file, standard library. It reports every error and warning against §10, note by note, and exits 0 when the vault is conformant. On the right: the example vault, then a real marketing vault before and after two fixes.

Download audit_reference.py
Python 3.10+ · no dependencies · read-only · example vault or the bakery vault (.zip) to try it on
vault:
# run it on your vault python audit_reference.py ~/vaults/example-vault
10What's next

Open tiles. Announced, not promised.

The fourth tile of the mark is outlined and empty: it stands for what is decided but not built. These are the open tiles of the format itself. Each is recorded in the example vault's ledger with a status, like any other open question.

Nothing on this list is a product. What produces, enriches, searches or composes vaults is out of scope by design (§8).

Not going to write YAML today?

Send an empty email. You get one line when 1.1 ships and one when the checker package is out. Nothing else.

Keep me posted →No form, no list, no tracking: it is a mailbox. Reply to unsubscribe, or to ask anything.