/ingest-source — ingest a paper/article/talk/docs into the wiki

Encodes SCHEMA's INGEST-SOURCE operation. The discipline matters: ingest is the front door for the wiki's hypothesis-generation pipeline, and getting emphasis wrong (or skipping connection updates) propagates downstream into bad hypothesis proposals.

When to invoke

• User says "ingest this paper / blog post / talk / docs page" and provides a URL or raw/sources/ path.
• A paper or talk just landed that touches TPU perf-relevant topics (Pallas kernels, sharding partitioners, FP8 training, attention kernels, scheduler/compiler passes, etc.).
• An existing source needs re-ingestion because a meaningful update was published.
• NOT for codebases — those go through INGEST-CODEBASE (separate skill, TBD).
• NOT for per-experiment narrative — that's wiki/experiments/ and a different operations contract.

Cardinal rules

• Discuss before writing. The human owns emphasis decisions. Do NOT write the source page until you've surfaced the key claims and confirmed which ones matter.
• Update connections. Every concept, model, or codebase that the source touches gets its ## Sources updated and (where appropriate) prose extended. Skipping this step is the canonical failure — orphan source pages with no incoming links.
• Generate hypothesis candidates. If the source suggests optimizations applicable to an active model, file them as wiki/hypotheses/<slug>.md with origin: <source-slug>. Add to the model page's ranked list.
• Update index.md and log.md. Mandatory per SCHEMA's operations contract.
• Never modify raw/. The source file is immutable. If the URL was given, save to raw/sources/<year>-<slug>.<ext> first.

Step 1 — Acquire the source

The user supplies one of:

• A raw/sources/<year>-<slug>.<ext> path (already ingested locally)
• A URL (fetch + save)
• A direct file paste

If URL: fetch and save to raw/sources/<year>-<slug>.<ext>. Choose <slug> to be short, lowercase, hyphen-separated, descriptive. Choose <year> from the source's publication date.

If the source has figures the wiki page will reference, save figures to raw/assets/<year>-<slug>-<figure>.<ext>.

Step 2 — Read fully + draft key claims

Read the source end-to-end (PDF, HTML, transcript, slides). Do NOT skim. The whole point of ingestion is to extract durable knowledge; skimming produces shallow source pages that propagate misunderstanding downstream.

Draft (in your head or scratch space, NOT yet in a file):

• 1-paragraph summary
• 3–7 key claims with the exact mechanism + estimated impact when stated
• Key data points (extract tables when present)
• Techniques referenced (link to existing concept pages where they exist; flag missing concept pages)
• Gaps + caveats (what the source DOESN'T cover; what limits its conclusions)
• Connections to existing wiki pages (concepts / models / codebases / observations / other sources)

Step 3 — Discuss with the human (MANDATORY)

Surface the draft to the user. Concretely:

Read [<year>-<slug>]. Key claims I extracted:

  1. <claim 1> — impact: <if stated>
  2. <claim 2> — impact: <if stated>
  3. ...

Connections I see:
  - Concept: <existing page or "STUB needed"> — extends with <one-liner>
  - Model: <active model> — applicable hypothesis: <one-liner>
  - Codebase: <existing page> — references <subsystem>
  - Sources: <other sources this should cross-link to>

Hypothesis candidates this could generate (for which models):
  - <model>: <hypothesis statement> (estimated <gain>, effort <S/M/L>)
  - ...

Anything to emphasize or de-emphasize before I file the wiki page?

WAIT for the human's response. The human typically:

• Reorders or adds emphasis to specific claims
• Says "skip this; not relevant to our optimization focus"
• Notes contradictions with existing wiki content (handle per the [!warning] rule)
• Names additional models/concepts to connect

Don't proceed to Step 4 until you have explicit confirmation or instructions.

Step 4 — Write the source page

Path: wiki/sources/<year>-<slug>.md. Format per SCHEMA's source page template:

---
title: "<full title>"
type: source
tags: [#<type>, ...]  # one of #paper, #article, #blog, #talk, #docs
created: <YYYY-MM-DD>
updated: <YYYY-MM-DD>
---

<lead paragraph — 1-3 sentences defining the page for any reader>

## Overview

<2-4 paragraphs summarizing the source>

## Key claims

<numbered list of the agreed-upon-with-human claims, each with mechanism + impact>

## Key data points

<tables / numbers where the source provides them>

## Techniques referenced

<list with links to existing concept pages, or stubs for missing ones>

## Gaps & caveats

<what the source doesn't cover; limits of its conclusions>

## Connections

<list of wiki pages this source updates or should update>

## Sources

- raw/sources/<year>-<slug>.<ext>
- (URL if applicable)

Tags must include one of: #paper, #article, #blog, #talk, #docs.

Step 5 — Update connections

For each connection identified in Step 3:

• Concept page touched: add the new claim to the page's prose, append the source to its ## Sources list, bump updated: field. If the concept page doesn't exist, create a STUB per SCHEMA (tags: [stub] + *Stub — expand when more sources are available.* body).
• Model page touched: if the source suggests an optimization applicable to an active model, surface in the Cross-variant or Variant-specific open hypotheses section.
• Codebase page touched: extend the relevant H2 section (typically Performance-relevant surfaces) with the new mechanism + source citation.
• Other source pages cross-linked: add reciprocal links in the ## See also sections.

Step 6 — Generate hypothesis candidates

For each (model, hypothesis) pair from Step 3 the human approved:

• File wiki/hypotheses/<slug>.md per SCHEMA's hypothesis page template, with frontmatter origin: <source-slug> linking back to this source.
• Add to the model page's ranked list (Cross-variant open hypotheses or Variant-specific open hypotheses depending on scope).

If no hypotheses came out (the source was useful for concept-page enrichment but didn't suggest a directly-testable mechanism), skip this step.

Step 7 — Update wiki/index.md and wiki/log.md

wiki/index.md:

• Bump the Sources count (## Sources (N+1)).
• Add a one-line entry: - [<title>](sources/<year>-<slug>.md) — <one-line takeaway>.
• If concept stubs were created, also bump Concepts count.
• If hypothesis pages were filed, bump Hypotheses count + add ranked rows.

wiki/log.md (global; ingest is cross-cutting per SCHEMA's two-tier log convention):

## [YYYY-MM-DD] ingest-source | <title>

**Op**: ingest-source
**Pages created**: wiki/sources/<year>-<slug>.md, <any concept stubs>, <any hypothesis pages>
**Pages updated**: <list of concept/model/codebase pages touched>, wiki/index.md
**Key result**: <one-paragraph summary of what the source contributes — claims, applicability to active models>
**Notes**: <human guidance applied; contradictions resolved; specific de-emphasis decisions>

Step 8 — Return to user

Brief summary:

Ingested <year>-<slug>:
  - Created: wiki/sources/<year>-<slug>.md
             <N concept stubs / N hypothesis pages>
  - Updated: <N concept/model/codebase pages>
             wiki/index.md (+1 source)
             wiki/log.md (+1 entry)
  - Key claims: <count>
  - Hypothesis candidates: <count> filed (model: <which>)

Failure modes to handle

• URL fetch fails — surface the error; ask user to download manually and provide the raw/sources/ path.
• Source is in a format we can't read (proprietary, password-protected) — refuse with a hint; ask user to extract relevant text.
• Discussion step (Step 3) reveals the source isn't relevant — abandon ingestion; do NOT file the source page or update connections. Note in summary: "Discussion concluded source not applicable to current optimization scope; not ingested."
• Existing source page for this <year>-<slug>** — never silently overwrite. If a re-ingest is intended (source updated), prompt user: "Existing page at ; overwrite (re-ingest) or version-suffix?" Default: prompt for which.
• Contradicts existing wiki claim — use the [!warning] contradiction block per SCHEMA's "Contradictions" rule. Surface to user; they adjudicate.
• Connection ambiguity (multiple concept pages plausible) — Ask via AskUserQuestion; don't guess.

Cost expectations

• Per source: ~10–30 minutes wall time depending on source length. Most of the time is in Step 2 (reading) and Step 3 (discussion).
• Source-page write: ~5 minutes after discussion converges.
• Connection updates: ~10 minutes (varies by how many pages need touching).
• Tokens: ~30–100K depending on source length + connection breadth.

What this skill does NOT do

• Does not ingest codebases. Different operation with different concerns (Architecture / Performance-relevant surfaces / SHA-pinning). Use the future /ingest-codebase skill or follow SCHEMA's INGEST-CODEBASE operation manually for now.
• Does not silently overwrite existing source pages. Prompts on conflict.
• Does not file hypotheses without human approval. Step 3's discussion gates hypothesis generation.
• Does not commit. Wiki changes are staged; user reviews and commits per existing convention.
• Does not modify raw/. Source file is immutable once saved.
• Does not update model-optimization-index.md automatically. If the source suggests a new topic mechanism citation, that's a separate manual edit (or a future regen of the index).

See also

• SCHEMA.md — the INGEST-SOURCE operation contract this skill implements
• wiki/model-optimization-index-regenerate-prompt.md — how the index uses ingested sources at next regen
• .claude/skills/formulate-hypothesis/SKILL.md — downstream consumer (cites sources in proposals)
• .claude/skills/extract-ast/SKILL.md — sibling skill for ingested codebases (AST snapshots)