Initial commit: lensing_wiki + README

Bootstraps the PyAutoPaper repo with the strong-gravitational-lensing
sub-wiki (~85 markdown files under lensing_wiki/), a short README, and a
.gitignore that keeps PDFs and other binaries out of the repo.

The wiki follows Karpathy's LLM Wiki pattern: raw PDFs as the immutable
local layer, compiled topical concept / entity / source pages as the
git-tracked layer, with a CLAUDE.md schema describing how an AI
assistant should read and maintain it.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Author: 2 people

.gitignore

@@ -0,0 +1,50 @@
+# PDFs are kept locally only — they are available from arXiv / journal sites
+*.pdf
+
+# Other binary/large files we don't want in the repo
+*.exe
+*.zip
+*.tar
+*.tar.gz
+*.tgz
+*.7z
+*.odt
+*.doc
+*.docx
+*.ppt
+*.pptx
+*.xls
+*.xlsx
+*.jpg
+*.JPG
+*.jpeg
+*.png
+*.PNG
+*.gif
+*.tif
+*.tiff
+
+# OS / editor cruft
+.DS_Store
+Thumbs.db
+desktop.ini
+.idea/
+.vscode/
+*.swp
+*~
+__pycache__/
+*.pyc
+
+# LaTeX build artefacts
+*.aux
+*.bbl
+*.blg
+*.fdb_latexmk
+*.fls
+*.log
+*.out
+*.synctex.gz
+*.toc
+
+# Exception — the wiki's own log.md is text and must be tracked
+!lensing_wiki/log.md

README.md

@@ -0,0 +1,44 @@
+# PyAutoPaper
+
+A personal paper library and writing context repo for
+[PyAutoLens](https://github.com/Jammy2211/PyAutoLens) and related
+science. Holds source PDFs locally (gitignored) and compiled
+LLM-readable wikis in git.
+
+## What's in here
+
+- **`lensing_wiki/`** — the strong-gravitational-lensing sub-wiki.
+  Topical concept pages, named-entity pages (surveys, lenses, software),
+  and per-topic bibliographies covering ~190 papers across
+  `Strong_Lens/`, `Substructure/`, `StrongLensCluster/`,
+  `Dark_Matter_Detection/`, `DarkMatterModels/`. Built on Karpathy's
+  "LLM Wiki" pattern: the PDFs are the raw layer, the wiki is the
+  compiled layer an AI assistant reads at query time. Start at
+  [`lensing_wiki/index.md`](lensing_wiki/index.md). See
+  [`lensing_wiki/CLAUDE.md`](lensing_wiki/CLAUDE.md) for the schema.
+- **`Strong_Lens/`, `Substructure/`, …** (gitignored) — the PDF folders.
+  Kept locally so the wiki can reference them; not pushed because the
+  PDFs are available from arXiv / journal sites.
+
+## For collaborators
+
+Clone the repo and you have the wikis. You won't see the PDFs — fetch
+them from the URLs cited in each source page if you need them. The
+wikis are self-contained scientific context: every concept page is a
+synthesis, every source page lists the file path and a one-paragraph
+stub keyed to the cited PDF.
+
+The wiki status of each source page is tracked in its frontmatter
+(`status: stub | drafted | reviewed`). In the initial build all
+per-paper stubs are `stub` — inferred from filenames and field
+knowledge, not yet verified against the PDFs. Anything contributed
+should upgrade stubs to `drafted` after the corresponding PDF has been
+read, and add an entry to `lensing_wiki/log.md`.
+
+## Roadmap
+
+- Verify and upgrade per-paper stubs from `stub` → `drafted`.
+- Add tangential folders (`WeakLensing/`, `Ellipticals/`,
+  `Deep Learning/`, `Stats/`, …) to scope.
+- Add sibling sub-wikis as the library grows (e.g. `imaging_wiki/`,
+  `cosmology_wiki/`) following the same schema.

lensing_wiki/CLAUDE.md

@@ -0,0 +1,156 @@
+# PyAutoLens AI Assistant — Strong Lensing Wiki
+
+This wiki gives a PyAutoLens AI assistant the broad scientific context of
+strong gravitational lensing. It follows Karpathy's "LLM Wiki" pattern: the
+PDFs in the surrounding `PyAutoPaper/` repo are the immutable `raw/` layer
+(gitignored — kept locally only), and this `lensing_wiki/` folder is the
+compiled, cross-linked knowledge layer the assistant reads at query time
+and that lives in git for collaborators.
+
+## Layout
+
+```
+PyAutoPaper/                  # repo root
+├── lensing_wiki/             # this folder — the compiled wiki (in git)
+│   ├── CLAUDE.md             # this file — schema + usage rules
+│   ├── index.md              # top-level navigation
+│   ├── log.md                # append-only compilation log
+│   ├── concepts/             # one topic per page — the science
+│   ├── entities/             # specific surveys, lenses, collaborations, software
+│   └── sources/              # per-topic bibliography pages (one paper = one section)
+├── Strong_Lens/              # raw/ layer — PDFs (gitignored)
+├── Substructure/
+├── StrongLensCluster/
+├── Dark_Matter_Detection/
+├── DarkMatterModels/
+└── ...                       # other PDF folders (gitignored)
+```
+
+The PDFs are the ground truth. Wiki pages are syntheses. If a wiki page and a
+PDF disagree, the PDF wins; the wiki page should be updated and the change
+noted in `log.md`.
+
+## Page types
+
+| Type        | Folder       | Scope                                                 |
+|-------------|--------------|-------------------------------------------------------|
+| Concept     | `concepts/`  | One scientific concept (e.g. mass-sheet degeneracy)   |
+| Entity      | `entities/`  | One named thing (survey, lens, code, collaboration)   |
+| Sources     | `sources/`   | All paper stubs for one topic, one section per paper  |
+| Index/log   | root         | Navigation and provenance                             |
+
+## Naming
+
+- File names are lowercase kebab-case: `mass-sheet-degeneracy.md`,
+  `slacs.md`, `time-delay-cosmography.md`.
+- One concept per concept page. If a page tries to cover two ideas, split it.
+- Source-collection pages are named by topic: `sources/dark-matter-substructure.md`.
+
+## Cross-references
+
+Use `[[page-slug]]` for wiki-internal links — for example
+`[[mass-sheet-degeneracy]]` or `[[slacs]]`. Slugs match the filename without
+`.md`. A `[[link]]` that has no target file yet is fine — it marks a future
+page to write.
+
+External references point at the PDF on disk using a relative path from
+the `PyAutoPaper/` repo root, e.g. `Strong_Lens/Suyu2016Holicow.pdf`.
+The PDFs themselves are kept locally (gitignored), so collaborators
+cloning the repo will see the wiki references but need to source the
+PDFs separately (most are on arXiv or the journal site).
+
+## Frontmatter
+
+Every wiki file starts with YAML frontmatter:
+
+```yaml
+---
+title: Mass-sheet degeneracy
+type: concept            # concept | entity | sources | meta
+topics: [degeneracies, cosmography]
+sources:                 # optional — papers most relevant to this page
+  - Strong_Lens/Suyu2016Holicow.pdf
+  - Strong_Lens/Birrer2020TDCOSMOSIVH0.pdf
+status: stub             # stub | drafted | reviewed
+---
+```
+
+Sources may be `[]` for pages that synthesize general field knowledge.
+
+## Concept page structure
+
+```
+# Title
+
+## TL;DR
+One paragraph an assistant can quote back to a user.
+
+## What it is
+The physics / definition.
+
+## Why it matters for PyAutoLens
+How this concept shows up in lens-modeling decisions a PyAutoLens user makes.
+
+## Key results from the literature
+Bullet list. Each bullet ends with `([[author-year-stub]])` so the LLM can
+follow the link to the per-paper section.
+
+## See also
+- [[related-concept-1]]
+- [[related-concept-2]]
+```
+
+## Entity page structure
+
+Same idea but the headings are "What it is / Key facts / Papers / See also".
+Use entity pages for: surveys (SLACS, BELLS, H0liCOW), specific lenses
+(Abell 1201, the Cosmic Horseshoe), software (PyAutoLens, lenstronomy),
+collaborations (TDCOSMO, Space Warps).
+
+## Source-collection page structure
+
+```
+# Sources: <topic>
+
+Bibliography of papers in `Papers/` covering this topic. Each paper has its
+own H2 section; cross-link from concept and entity pages with
+`[[sources-<topic>#author-year-slug]]`.
+
+## Author Year — short tag
+
+**File:** `Strong_Lens/Filename.pdf`
+**Concepts:** [[concept-1]], [[concept-2]]
+**Summary:** one-paragraph stub inferred from filename and field knowledge.
+Mark `(stub — verify against PDF)` if not yet read.
+```
+
+## How the assistant should use this wiki
+
+1. On a user question, first open `index.md`.
+2. Follow the relevant `concepts/` or `entities/` page.
+3. Cite specific results by linking the `[[sources-topic#author-year]]`
+   anchor. If a claim is needed and the source stub is unread, open the
+   PDF at the path in the stub's `File:` line.
+4. When a PDF is read in full, upgrade the stub's `status:` from `stub` to
+   `drafted`, replace the inferred summary with what the paper actually
+   says, and add a line to `log.md`.
+5. Never fabricate a citation. If a result is not on a wiki page, say so
+   and offer to read the relevant PDF.
+
+## Scope
+
+In-scope folders for the current wiki build:
+
+- `Strong_Lens/`           — primary
+- `Substructure/`
+- `StrongLensCluster/`
+- `Dark_Matter_Detection/`
+- `DarkMatterModels/`
+
+Other folders (`WeakLensing/`, `Ellipticals/`, `Deep Learning/`, etc.)
+contain papers that touch lensing tangentially but are out of scope until
+explicitly added — see `log.md` for the decision.
+
+This wiki is the **lensing** sub-wiki of `PyAutoPaper/`. Future sub-wikis
+(e.g. `imaging_wiki/`, `cosmology_wiki/`) may sit alongside it at the
+repo root following the same schema.

lensing_wiki/concepts/bayesian-inference-lensing.md

@@ -0,0 +1,68 @@
+---
+title: Bayesian inference for strong lensing
+type: concept
+topics: [methods]
+sources: []
+status: drafted
+---
+
+# Bayesian inference for strong lensing
+
+## TL;DR
+
+Strong-lens modelling is a high-dimensional, multimodal, often
+ill-conditioned inverse problem. The community has converged on Bayesian
+inference with nested sampling for posterior + evidence estimation.
+PyAutoLens uses the [[pyautofit]] library for this, supporting nested
+samplers like Dynesty, UltraNest, and Nautilus, and MCMC samplers like
+emcee. Pixel-source likelihoods carry an analytic source-marginalisation
+term (the "Bayesian evidence" of Suyu / Koopmans).
+
+## Why nested sampling
+
+- Returns the **evidence Z** — central for [[mass-models|model
+  comparison]] (with vs. without external shear, EPL vs. composite, etc.).
+- Handles multimodality (mirror images, parity ambiguities) better than
+  vanilla MCMC.
+- Doesn't need an initial guess close to the peak.
+
+## The evidence for pixelised sources
+
+For a pixelised source under Gaussian regularisation, the marginal
+likelihood over source pixels is analytic:
+
+```
+log Z(θ_mass, λ) = −½ χ²(ŝ) − ½ log|LᵀC⁻¹L + λH| + ½ log|λH| − ½ N log(2π)
+```
+
+This is the quantity nested sampling explores over **mass-model
+parameters and regularisation hyperparameters**, with the source itself
+already marginalised out.
+
+## Pipelines and hierarchical fits
+
+A modern fit (e.g. SLaM; [[slam-pipeline]]) chains together a sequence of
+sub-fits: first an SIE + Sérsic, then a power-law with a parametric
+source, then the same with a pixelised source, then optionally a
+composite stellar + dark fit. Each phase passes its posterior as a prior
+to the next via PyAutoFit's "search chaining". This is essential for
+robustness on real data.
+
+## Model comparison
+
+Bayesian evidence is the canonical metric. In practice:
+
+- Two-sided log-evidence differences > a few are decisive.
+- Cross-check with [[bayesian-inference-lensing|posterior predictive
+  checks]] and residuals.
+- For [[dark-matter-substructure|subhalo detection]] the
+  evidence-difference threshold is calibrated against
+  [[sources-dark-matter-substructure|Despali 2018/2022 sensitivity
+  studies]].
+
+## See also
+
+- [[regularization]]
+- [[source-reconstruction]]
+- [[pyautofit]]
+- [[pyautolens]]

lensing_wiki/concepts/bulge-halo-decomposition.md

@@ -0,0 +1,70 @@
+---
+title: Bulge–halo (stellar + dark matter) decomposition
+type: concept
+topics: [galaxy-structure]
+sources:
+  - Strong_Lens/Etherington2022BulgeHalo.pdf
+  - Strong_Lens/Etherington2022SLaM.pdf
+  - Strong_Lens/Etherington2023aBeyondBulge.pdf
+  - Strong_Lens/Shajib2021MassSlopeNFW.pdf
+  - Strong_Lens/Oldham2018StellarAndDark.pdf
+status: drafted
+---
+
+# Bulge–halo decomposition
+
+## TL;DR
+
+Decompose a lens galaxy's mass into a stellar component (mass-traces-
+light Sérsic or MGE) and a dark-matter component (NFW or gNFW). This
+splits what the elliptical-power-law model conflates into a single slope
+into physically meaningful components — stellar M/L and DM concentration
+— at the cost of more parameters and a stronger sensitivity to the
+[[radial-angular-degeneracy]].
+
+## What it is
+
+Inputs:
+
+- A high-resolution image of the lens light → MGE / Sérsic / multi-Sérsic
+  decomposition (see [[sources-bulge-halo|Etherington series]]).
+- A free overall M/L or per-component M/L.
+- An NFW (or gNFW) DM halo.
+- Optional adiabatic contraction.
+- [[external-convergence-shear|External shear]].
+
+Outputs:
+
+- Stellar M/L, often inverted to **the IMF**
+  ([[microlensing-imf|complementary to microlensing-IMF measurements]]).
+- DM concentration and mass.
+- Stellar-to-halo mass ratio inside the [[einstein-radius]].
+
+## Why it matters
+
+- Tests galaxy formation: do massive ellipticals have NFW-consistent
+  haloes or contracted ones?
+- Calibrates the **dark-matter fraction inside R_eff**, a benchmark
+  galaxy-evolution quantity.
+- Disentangles IMF from DM concentration, the long-standing
+  bulge-vs-halo degeneracy.
+
+## In PyAutoLens
+
+The SLaM pipeline ([[slam-pipeline]]) chains:
+
+1. Sérsic-mass + SIE phase.
+2. Power-law phase with parametric source.
+3. Power-law phase with pixelated source.
+4. Composite Sérsic + NFW phase — the bulge-halo step.
+
+Each phase informs the next via PyAutoFit search chaining
+([[bayesian-inference-lensing]]).
+
+## See also
+
+- [[mass-models]]
+- [[radial-angular-degeneracy]]
+- [[microlensing-imf]]
+- [[slam-pipeline]]
+- [[sources-bulge-halo]]