Intake

NKS Intake

External word — an issue, a README, a chunk of source, a conversation, an RFC — is шабда: testimony with existence outside the graph. It is not yet yours. This skill brings it in as a claim, never as a fact — mapped to the right node type, marked by provenance, deduped, anchored, and then verified by direct observation (пратьякша).

The symptom this skill prevents: external word dumped en masse as facts floods the realm with authoritative-looking garbage and false tensions, breaking «the graph is a tension with reality» and «systems grow, not get built».

Source-independence is the whole point. The source is an adapter; the intake discipline below is one. product-roadmap composes this layer with a GitHub adapter; a future verstakify seeding or a docs-intake composes it with another. Do not bake a source into the discipline.

The source adapter — its contract

Fetching external word is the adapter's job, not this skill's. Whatever the source (GitHub gh, file read, a conversation transcript), the adapter must hand the core, per item:

• content — the claim itself (title + trimmed body, or the prose chunk);
• form — what kind of utterance it is (a report of breakage, a wish, an open question, an opinion, a stated fact, a decision), so the core can map it;
• provenance — a stable back-reference to the origin (issue URL, file path + symbol, "conversation 2026-…"), so each node can arose_from it;
• authority — who said it and how directly it can be checked (a drive-by issue vs. the owner's decision vs. code you can read now). This sets the epistemic mode (step 3).

If the adapter can't supply provenance for an item, that item is not intake-able — it would land unanchored and invisible.

1. Map the form to a node type

External word is heterogeneous; one item is not one node-shape. Map by form — kind of speech, not tracker category — with the agent's judgement (propose → confirm), never a blind import:

Tracker words — bug, feature-request, backlog label — are one adapter's rendering of these rows, not forms of their own; a news item, a chronicle entry and a forwarded post map through the same six.

Naming, given_as, and the modes themselves are the writing skill's discipline — invoke it at each write. This skill decides which shape the external form takes.

2. Set the epistemic mode BY KIND — and mark provenance

This is the move agents get wrong. source_kind=shabda is provenance (this entered as external word) — orthogonal to epistemic mode (how well we know it). kalpita is not a blanket stamp; it is the mode of an unverified claim. By kind of source:

Two consequences:

• The owner's volition is not shabda-to-verify. When the external word is a decision, goal, or telos from the realm's owner (the āpta for their own intent), it does not enter kalpita awaiting verification — it lands at its volitive mode (chanda/adhimoksha), witnessed (pratyakshita). Only assertions about a state of the world take the kalpita-until-pratyaksha path.
• kalpita is never permanent. It is the entry state of the шабда→сверка→тезис (word → check → thesis) thread; step 4 graduates it.

Model the node accordingly: attrs.source_kind="shabda", the given_as the content warrants, the epistemic mode from the table.

3. Dedup before you write

Before each insert, nks_semantic_search(q=<the claim as a phrase>, realm=…) against the existing graph (the ground, not only the prior backlog). Near-zero distance = the realm already carries this — link or update, do not duplicate. Locate-before-write is mandatory: intake's failure mode is N near-identical nodes, one per restated issue.

4. Anchor every node to its source

An intake node with no origin is a rumour. Each one needs:

• arose_from → the provenance (the post, the issue, the file, the conversation) — its traceable origin;
• its type's home: a vimarsha needs vimarsha_of into the contour it's about (the holon/phenomenon/kriya the claim touches); a phenomenon needs context → its holon.

The provenance is a node, not an attribute. Bring the artefact in as a ding phenomenon — the post, the document, the file, the URL — with attrs {author, url/path, source_kind: "shabda"}, and grow every node from that item arose_from it. Dissolved into its children's attributes a source has no lifecycle: it can't be argued with, can't go badhita when it turns out false, can't be found again when the next text from the same author arrives.

Anchoring is the inquiry skill's law (an unanchored vimarsha is invisible) — honour it at intake time, not post-hoc.

Word serving more than one contour has a different home. A finding, benchmark or diagnosis bearing on several of the user's realms belongs in their personal contour (minding — where the user keeps one; verstakify's cross-project-memory answer decides), not in whichever realm happened to be open — anchored there, it is invisible to every other contour it was meant to inform.

5. Verify by пратьякша and graduate the mode

Intake is not done when the claim is written — it's done when it's been reconciled with reality (the сверка kriya). For each kalpita node, observe directly on the three levels: nks_look (recall — what's written), nks_orient(lens="tensions")/neighborhood (awareness — what pulls), nks_orient(lens="trace") (reflection — where it travels). Then graduate:

• claim matches reality → assert the thesis: nks_update the epistemic mode up (kalpita → pratyakshita/pramanita), or let it stand as confirmed;
• claim contradicts reality → that's a tension, not a fact: leave it badhita, or raise a vyabhichara/samshaya (сверка routes found counter-examples to vyabhichara, not samshaya), or fix the incident;
• mode merely drifted → update the recording mode.

A realm that ingests but never verifies is exactly the mass-dump failure with extra steps.

When пратьякша is unreachable

Observation of 1917 is impossible in principle, so by the letter above a kalpita claim about the past would never graduate. Check testimony against independent sources instead:

• sources converge → pramanita;
• it follows from documents you can read now → anumita, and it stays there — for the past anumita is terminal, not a way-station;
• sources diverge → vyabhichara (one carries a counter-example) or prati-paksha (a rival account), node left at the mode its evidence warrants. Not badhita — witnesses disagreeing is not a ground that has fallen.

Same branch for anything else out of reach: a closed system, a stated intent, a vendor's internals.

6. Selectivity is a guard, not a convenience

Never the whole source — not the whole tracker, not "all the docs", not the entire feed. Declare the cut (a date bound, a marker, the high-signal subset), cap the volume, and state what you dropped — a silent cap reads as "ingested everything". Selectivity + the shabda modes + dedup are together the mitigation of the mass-dump failure; drop any one and the guard fails.

After writing

Read the CHECKS: block each factory prints (orphan, missing anchor). A fresh intake phenomenon stays not_orphan-flagged until a kriya picks it up — wire it. Then nks_orient(lens="trace", focus=<seq>) on key phenomena: did the lifecycle connect?

What intake is NOT

• Not writing — writing creates one node from a distinction you already hold; intake creates many from external word, with the шабда provenance and the сверка loop.
• Not verstakify — verstakify projects derived config (the AGENTS.md audit, "no duplication of NKS"); intake brings in content. Different acts; verstakify may compose intake to offer засев at the end of bootstrap.
• Not product-roadmap — that is an applied skill (GitHub adapter + present-state ground modelling + assembly + render) that composes this layer; intake is the base, not the pipeline.
• Not a renderer / not a modeller of the present-state ground — those belong to the consumer.
• Not bound to GitHub — issues are one adapter among many.