zettel-note

Zettel note

A note in this vault is not a summary of a source — it is one claim you can defend, written in your own words, and wired to its neighbors so the graph thinks for you later. The companion model is reading-with-a-Zettelkasten: you read, you understand, you close the book, and you write the idea as if explaining it to yourself. If you cannot write it without looking, you have not understood it yet.

All mechanics here are fixed by [[conventions|Vault Conventions]] (§3 linking, §4 relationship vocabulary, §5 naming, §6 frontmatter, §7 internal structure). This skill is the discipline that makes a note worth linking. It is the routing target for claims drawn out in [[experiential-kolb-teaching]]: when a learner reconstructs a mechanism, it becomes an evergreen note written this way.

The five rules

1. The title is a declarative claim

The title states something true that the note then defends. Not a topic, not a question, not a noun phrase — a sentence with a verb and a stance.

Bad title (topic)	Good title (claim)
Heritability	Heritability is a property of a population in an environment, not of a trait
BLUP	BLUP shrinks each estimate toward the population mean in proportion to how little data supports it
Marker density	Adding markers beyond linkage-disequilibrium saturation adds noise, not predictive signal

The slug is the title slugified: lowercased, spaces to hyphens, apostrophes and punctuation stripped, sentence structure kept (§3, §5). Long slugs are expected and fine. The title matches the note's # H1 and its aliases.

2. One claim per note (atomicity)

A note holds exactly one idea. The tell is the word and in a title joining two stances, or a second paragraph that introduces a new claim instead of developing the first. When you catch it, split into two notes and link them — usually Builds-on or Contrasts-with.

Drafting additive-variance-drives-selection-response-and-dominance-variance-does-not? That's two claims. Split: selection-response-is-driven-by-additive-genetic-variance and dominance-variance-does-not-contribute-to-the-breeders-equation-response, linked Contrasts-with.

Atomic notes are reusable: a single claim can be cited by many posts and many other claims. A compound note can be cited by none cleanly.

3. Own words, no block quotes

Write the claim as you would explain it out loud. No copied passages, no block quotes from the source (§7). If a phrase from the source is doing real work, paraphrase it and cite the source — the source lives in the YAML source: field and as a Source: [[slug|Title]] link directly under the H1 (the YAML field is invisible to the graph; the body link is what the graph sees — §3). Two to four paragraphs that develop the single claim. An optional **Open questions:** list is where honest edges go.

4. Link 3-6 times, every link labeled by relationship

Every internal link uses the piped form [[slug|Display Title]] — never bare [[Title]], which breaks because the file on disk is the slug (§3). In the ## Links section, label each link with its relationship from the controlled vocabulary (§4):

Label	Means	Example use
Confirms	same claim, independent source	a second paper finding the same accuracy ceiling
Contradicts	opposing claim, productive tension	a result that says more markers did help
Extends	builds on / deepens / applies	from heritability to the breeder's equation
Context	background, framing	what additive variance is, for a note about response
Prerequisite	must hold this claim first	linkage disequilibrium, before genomic prediction
Builds-on	this claim assumes and develops that one	GEBV builds on the kinship matrix
Applies	this claim is put to work in a project/method	heritability applied in a simulation lab
Example-of	a concrete instance of a general claim	a maize case as an example of G×E
Contrasts-with	related but importantly different	narrow- vs broad-sense heritability

Target 3-6 outbound links (§4). Fewer than three usually means the claim is orphaned or you haven't searched; more than six usually means the note is not atomic. Pick the relationship that is most true, not the most generic — Context is the lazy default; reach for Prerequisite, Builds-on, or Contrasts-with when they fit.

5. Search before you create

Before writing a new note, search the vault for the claim. The graph rots fast with near-duplicates. Run search-corpus / dedupe-against-corpus (or a plain grep over evergreen/) on the key terms and the claim's verb.

• If an identical claim exists: don't create. Add your new source as a Confirms link on the existing note, or sharpen its prose.
• If a near claim exists: decide whether yours is genuinely distinct. If it's a refinement, link Extends/Builds-on and make the boundary explicit. If it's the same claim said differently, merge.
• If nothing exists: create, and wire it into the closest existing notes immediately — an unlinked note is invisible.

Worked example: routing one claim

A Kolb session lands the claim that prediction accuracy stops improving past a marker-density threshold. Write it:

# Adding markers beyond linkage-disequilibrium saturation adds noise, not predictive signal

Source: [[wientjes-2016-marker-density-thread|Wientjes et al. — Marker Density and GS Accuracy]]

Once marker density is high enough that every causal locus is in tight linkage
disequilibrium with at least one marker, the genome is effectively "covered."
Past that point, extra markers cannot tag any new causal variation — there is none
left untagged. What they do add is estimation burden: each new column is another
effect to estimate from the same finite training set, so the model spends data
fitting noise. Accuracy plateaus, then sags.

This is why my prediction model got *worse* after I doubled the marker panel: I was
past saturation and bought only variance.

**Open questions:**
- Does the saturation point move with effective population size? (It should — Ne sets LD decay.)

## Links
- Prerequisite: [[linkage-disequilibrium-decays-with-recombination-distance|LD decays with recombination distance]]
- Builds-on: [[genomic-prediction-borrows-information-across-relatives-via-marker-tagged-ld|Genomic prediction borrows information via marker-tagged LD]]
- Contradicts: [[denser-panels-improve-accuracy-in-low-ld-populations|Denser panels help in low-LD populations]]
- Context: [[the-bias-variance-tradeoff-governs-predictive-model-error|The bias-variance tradeoff governs predictive error]]

Four labeled links, one atomic claim, own words, a source link under the H1, an honest open question. That note can now be cited by a post via [[experiential-kolb-teaching]]'s routing step, and surfaced for spaced review. The graph got smarter.

Common defects to catch

• Topic title ("Genomic selection") -> rewrite as a claim or it can never be confirmed/contradicted.
• Bare link [[Genomic prediction borrows information…]] -> breaks; use the slug-piped form.
• Context everywhere -> you skipped the harder, truer relationship label.
• Block quote smuggled in -> paraphrase; the note must be yours.
• Two claims, one note -> split and link.
• Created without searching -> likely a duplicate; merge or link instead.