Deep Dive: Architecture & Operations

my-wiki/
│
├── CLAUDE.md              ← Schema: conventions + workflows (bạn + LLM co-evolve)
│
├── raw/                   ← IMMUTABLE sources (bạn curate, LLM chỉ đọc)
│   ├── papers/
│   │   ├── attention-2017.pdf
│   │   └── gpt3-2020.pdf
│   ├── articles/
│   │   └── scaling-laws-explained.md
│   └── notes/
│       └── meeting-2026-04-10.md
│
└── wiki/                  ← LLM-OWNED content
    ├── index.md           ← Catalog: tất cả pages + one-line summary
    ├── log.md             ← Append-only chronological log
    │
    ├── entities/          ← Concept/entity pages (1 file = 1 concept)
    │   ├── attention-mechanism.md
    │   ├── transformer-architecture.md
    │   ├── bert.md
    │   └── gpt-3.md
    │
    ├── comparisons/       ← Cross-cutting comparisons
    │   └── bert-vs-gpt.md
    │
    ├── synthesis/         ← High-level synthesized views
    │   └── timeline-llm-breakthroughs.md
    │
    └── _inbox/            ← Raw observations, pending processing
        └── session-2026-04-11.md

# Wiki Schema

## Overview
This is a personal research wiki on [YOUR TOPIC].
You are the wiki maintainer. I curate sources; you do all the bookkeeping.

## Directory Structure
- raw/ — immutable source documents. NEVER modify.
- wiki/ — your domain. Create, update, maintain all files here.
- wiki/index.md — catalog of all pages (update after every ingest)
- wiki/log.md — append-only log (update after every operation)
- CLAUDE.md — this file

## Entity Page Format
```markdown
# [Entity Name]
**Type**: Concept | Person | Project | Library | Decision | Event
**Summary**: One sentence. Must be standalone (no "it" or "this").
**Tags**: #tag1 #tag2
**Sources**: [[raw/paper1.pdf]], [[raw/article1.md]]
**Last Updated**: YYYY-MM-DD
**Confidence**: 0.85 (2 sources, confirmed 2026-04-10)

---

## Overview
2-3 paragraph explanation for someone unfamiliar.

## Key Properties / Arguments
Bullet list of concrete, verifiable claims.

## Relationships
- **uses**: [[Entity A]] — why/how
- **depends_on**: [[Entity B]]
- **contradicts**: [[Entity C]] — what they disagree on
- **supersedes**: [[Entity D]] (as of 2026-04-10)

## Open Questions
Things still unclear or needing more sources.

## References
- [[raw/paper1.pdf]] — specific claim from here
```

## Ingest Workflow
When I say "ingest [source]":
1. Read the full source from raw/
2. List entities you found (confirm with me before writing)
3. For each entity:
   a. Check if page exists in wiki/entities/
   b. If yes: update, mark changed sections, update confidence, log supersessions
   c. If no: create new page using format above
4. Update wiki/index.md (add new pages, update summaries of changed pages)
5. Append to wiki/log.md: \`## [DATE] ingest | SOURCE TITLE\`
6. Note: one source may touch 10-15 pages — be thorough

## Query Workflow
When I ask a question:
1. Read wiki/index.md to find relevant pages
2. Read those pages (and their linked pages if needed)
3. Synthesize answer with citations to wiki pages
4. If answer is valuable: offer to file as new wiki page

## Lint Workflow (run every ~20 new pages)
Check for:
- Contradictions between pages → flag both, propose resolution
- Stale claims (confidence < 0.4) → flag for re-verification
- Orphan pages (no inbound links) → suggest where to link
- Concepts mentioned but lacking own page → propose creation
- Missing cross-references → add them
Report findings before making changes.

## Conventions
- Always use [[wiki-links]] for cross-references (Obsidian format)
- Confidence = (number_of_supporting_sources * 0.3) * (1 - decay_factor)
- Decay: 0.02 per week since last confirmation
- When superseding: mark old claim with ~~strikethrough~~ + link to new
- Contradictions: add ⚠️ badge on both pages + log entry

You are maintaining my research wiki. I'm uploading [N] sources on [TOPIC].

For each source:
1. Identify all key entities (concepts, people, systems, papers)
2. Create one markdown entity page per concept in wiki/ format (see template)
3. Add [[wiki-links]] connecting related concepts
4. Note any contradictions between sources
5. Create wiki/index.md listing all pages with one-line summaries
6. Create wiki/log.md with ingest entry

Use this entity page format:
[paste format from CLAUDE.md]

After writing all pages, give me a brief summary of what was created/updated.

Ingest this new source into my existing wiki.

First, read wiki/index.md to see what pages already exist.
Then read the source.

For each entity in the source:
- If page exists: UPDATE it (note changes, update confidence, log supersessions)
- If page doesn't exist: CREATE it

Important: this source may contradict existing claims.
If you find contradictions, flag them with ⚠️ on both pages.

After ingesting, update index.md and append to log.md.

# After receiving a good synthesis answer:
"This comparison is valuable. Please file it as wiki/comparisons/bert-vs-gpt.md
and add it to index.md with summary: 'Comparison of bidirectional vs autoregressive
pre-training: BERT vs GPT architectures'"

# After a research session:
"File the key insights from this session as wiki/synthesis/session-2026-04-20.md"

Run a health check on my wiki.

Read wiki/index.md and check all pages for:
1. CONTRADICTIONS: Claims on different pages that disagree
   → Flag both pages, propose which is more likely correct and why
2. STALE CLAIMS: Claims with confidence < 0.4 or last confirmed > 60 days ago
   → Mark for re-verification, suggest what source to look for
3. ORPHAN PAGES: Pages with no inbound [[wiki-links]]
   → Suggest where they should be linked from
4. MISSING PAGES: Concepts mentioned in passing but lacking own page
   → Propose page titles and stubs
5. BROKEN LINKS: [[wiki-links]] pointing to non-existent pages
   → Fix or flag

Report findings as a list. DON'T make changes yet.
After I review, I'll say "apply fixes" for each category.

# Wiki Index — [Your Topic]
Last updated: 2026-04-20 | Pages: 47 | Sources: 12

## Entities

### Architectures
| Page | Summary | Sources | Confidence |
|------|---------|---------|------------|
| [[transformer-architecture]] | Encoder-decoder with self-attention (Vaswani 2017) | 4 | 0.95 |
| [[bert]] | Bidirectional masked LM pre-training | 3 | 0.90 |
| [[gpt-3]] | Autoregressive LM with few-shot emergence | 2 | 0.85 |

### Concepts
| Page | Summary | Sources | Confidence |
|------|---------|---------|------------|
| [[attention-mechanism]] | Scaled dot-product attention over sequence | 5 | 0.98 |
| [[scaling-laws]] | Chinchilla: compute-optimal token/param ratio | 3 | 0.88 |

## Comparisons
- [[bert-vs-gpt]] — Bidirectional vs autoregressive pre-training tradeoffs
- [[rlhf-vs-dpo]] — Reward model vs direct preference tradeoffs

## Synthesis
- [[timeline-llm-breakthroughs]] — Chronological view of key developments
- [[open-questions]] — Unresolved questions and gaps

## ⚠️ Flagged Issues
- [[scaling-laws]]: Contradiction with [[chinchilla-paper]] on token ratio
- [[bert]]: Confidence 0.35 (last confirmed 2026-01-10, needs refresh)

## [2026-04-20] ingest | Attention Is All You Need (Vaswani et al. 2017)
Created: transformer-architecture.md, attention-mechanism.md, positional-encoding.md
Updated: index.md (+3 entries, updated architecture section)
Flags: None

## [2026-04-20] ingest | BERT: Pre-training Deep Bidirectional Transformers (2018)
Created: bert.md, masked-language-modeling.md, next-sentence-prediction.md
Updated: transformer-architecture.md (added BERT as derivative)
Flags: ⚠️ BERT contradicts GPT-* on directionality claim → flagged in both pages

## [2026-04-20] query | "How does multi-head attention differ from single-head?"
Filed as: wiki/comparisons/attention-single-vs-multi.md
Index updated: Yes

## [2026-04-20] lint | routine health check (47 pages, 12 sources)
Issues: 2 orphans linked, 1 contradiction flagged, 0 broken links
Actions: Applied all fixes

# Last 5 entries
grep "^## \[" log.md | tail -5

# All ingests
grep "^## .*ingest" log.md

# All flagged issues
grep "Flags: ⚠️" log.md -A1

# Count operations by type
grep "^## \[" log.md | grep -o "ingest\|query\|lint" | sort | uniq -c