Document confidence markers and typed relationships in ingest skill

skills/ingest/SKILL.md

@@ -32,6 +32,28 @@ This sequence covers a session ingesting one or more sources, listed manually.
      claim comes from a specific source when the page synthesizes 3+ sources.
    - If the source is `proposed` (not yet resolved), mark the page (or the
      relevant section) `status: proposed` / "Proposed change".
+   - **Confidence markers** (optional). Mark a claim `^[inferred]` when it's an
+     LLM-synthesized connection the source doesn't state directly, or
+     `^[ambiguous]` when sources disagree or are unclear; leave a faithful
+     paraphrase unmarked (extracted). When a claim also needs a `^[source_id]`
+     citation, stack them as independent suffixes with the confidence marker
+     first: `^[inferred]^[design-doc-3]`. If you add the page-level
+     `confidence:` frontmatter rollup, it must exactly match the inline
+     markers: count one unit per bullet line if the page has any bullets,
+     else one unit per paragraph; each unit is `inferred`/`ambiguous` per its
+     marker, else `extracted`; roll each state up as `count / total`, rounded
+     to 2 decimals round-half-up. `lint` checks this rollup for drift, so
+     recompute it by hand (or omit `confidence:` entirely) rather than
+     guessing — omitting the block is not an error.
+   - **Typed relationships** (optional). Add a `relationships:` entry to
+     enrich an existing `[[wikilink]]` already in the page body — never as a
+     duplicate link with no corresponding wikilink. Each entry is
+     `{target: "[[Page]]", type: "..."}` with `type` one of `extends`,
+     `implements`, `contradicts`, `derived_from`, `uses`, `replaces`,
+     `related_to`, and direction is always from this (the declaring) page's
+     perspective. Only pick a specific type when the source material makes
+     direction/type clear; otherwise use `related_to` or omit the entry
+     rather than fabricating one.
    - Use the [ingest-prompts](ingest-prompts.md) for writing pages.
    - **`wiki-toolkit log --action ingest --title "..." --details "..."`**
      Once per source, right after that source's page(s) are written. Appends

skills/ingest/ingest-prompts.md

@@ -0,0 +1,111 @@
+# Ingest Prompt Templates
+
+These are the mental frameworks to use when distilling a source into wiki pages.
+
+## Knowledge Extraction Frame
+
+When reading a source document, ask yourself:
+
+1. **What are the 3-5 most important ideas in this document?**
+   These become concept pages or updates to existing concept pages.
+
+2. **Who or what is mentioned that deserves its own page?**
+   People, tools, organizations, projects → entity pages.
+
+3. **What does this document teach you how to do?**
+   Procedures, workflows, techniques → skills pages.
+
+4. **What claims does this document make?**
+   Each claim needs a source attribution. If it contradicts an existing wiki claim, note the contradiction.
+
+5. **How does this connect to what the wiki already knows?**
+   This is the most important question. The value of the wiki compounds through connections.
+
+## Paper Extraction Frame
+
+For academic papers the generic frame above misses what makes a paper legible.
+Add these questions:
+
+1. **What problem does it solve, and what's new?** The one-sentence thesis + the single most important result.
+2. **What is the method?** Which figure shows the architecture/pipeline? Sketch it as a Mermaid flowchart — capture the data flow, not just the component names.
+3. **What are the core equations?** The 1–3 that define the mechanism — keep them as math (`$$…$$`), not prose.
+4. **What's the experimental setup and the headline numbers?** Datasets, baselines, and the metric table the paper is judged on.
+5. **What are the ablations and limitations?** What did they vary, and what does the method *not* do?
+
+These map onto the Paper Deep-Dive Template in `llm-wiki/SKILL.md`.
+The goal is a page a reader could study instead of the PDF; figures, equations, and results included.
+
+## Synthesis Frame
+
+When a new source covers ground that existing pages already cover:
+
+- Don't duplicate — synthesize
+- If the new source agrees with existing content, strengthen the claims with additional attribution
+- If it disagrees, create an "Open Questions" or "Debate" section noting both positions
+- If it adds nuance, weave it into the existing narrative
+
+## Cross-Reference Discovery
+
+After extracting knowledge, look for these connection patterns:
+
+- **Is-a**: "Transformers are a type of neural network" → link from transformer page to neural-network page
+- **Uses**: "RLHF uses reward models" → link from RLHF to reward-models
+- **Contrasts-with**: "CNNs vs. Transformers for vision" → mutual links
+- **Part-of**: "Attention is a component of transformers" → link from attention to transformers
+- **Created-by**: "Transformers were introduced by Vaswani et al." → link to entity page
+- **Applied-in**: "Transformers are used in GPT" → link from transformers to GPT
+
+## Confidence Markers
+
+Every claim starts as a faithful paraphrase (extracted, unmarked). Ask
+yourself, for each claim you write:
+
+- **Did the source say this, or did I connect the dots?** If you're the one
+  drawing the connection — the source never states it directly — mark it
+  `^[inferred]`.
+  > RLHF and constitutional AI both shape model behavior post-pretraining. ^[inferred]
+- **Do my sources disagree, or is one of them unclear on this point?** Mark
+  it `^[ambiguous]` instead of silently picking a side.
+  > The team disagrees on whether the migration caused the latency regression. ^[ambiguous]
+- **Does this claim also need a source citation?** Stack the confidence
+  marker first, the source citation second — they're independent suffixes.
+  > The retry queue was added to absorb the Q3 traffic spike. ^[inferred]^[incident-report-14]
+
+Don't hand-compute the page-level `confidence:` rollup by guessing — count
+your bullets (or paragraphs, if the page has no bullets), tally how many
+carry each marker, divide by the total, and round to 2 decimals,
+round-half-up (a tie like 0.125 rounds to 0.13, not 0.12). If you
+didn't track this carefully while writing, leave `confidence:` off the page
+rather than write a number `lint` will flag as drifted.
+
+## Typed Relationships
+
+Once a page has its `[[wikilink]]`s, look at each one and ask: **is the
+relationship between these two pages more specific than "related"?** If the
+source material makes it clear, add a `relationships:` entry describing it
+from this page's point of view — don't add a link that isn't already in the
+body, and don't add an entry when the source is vague about direction.
+
+```yaml
+relationships:
+  - target: "[[Reward Modeling]]"
+    type: uses          # this page's approach uses reward modeling
+  - target: "[[Legacy Auth Middleware]]"
+    type: replaces       # this page's approach replaces the old one
+  - target: "[[Constitutional AI]]"
+    type: related_to     # connected, but the source doesn't specify how
+```
+
+Type cheat sheet (always from the declaring page's perspective):
+
+- `extends` — this page builds on the target's ideas without replacing them
+- `implements` — this page is a concrete realization of the target's design
+- `contradicts` — this page's claims conflict with the target's
+- `derived_from` — this page's content was distilled from the target
+- `uses` — this page's subject depends on or invokes the target's
+- `replaces` — this page's subject supersedes the target's
+- `related_to` — connected, but none of the above fit or the source doesn't say
+
+When in doubt between a specific type and `related_to`, take `related_to`
+— or skip the entry. A fabricated `extends` is worse than no relationship
+at all.