Tesserae Demo Corpus: 3D Reconstruction (2016–2024)

A curated, synthetic research log used to populate the Tesserae GitHub Pages demo at <https://ca1773130n.github.io/Tesserae/>. The corpus covers 3D reconstruction broadly — neural radiance fields, Gaussian splatting, visual SLAM and multi-view stereo, diffusion-based 3D generation, neural-implicit surface and mesh extraction, dynamic / 4D reconstruction, and feed-forward generative 3D representations (LRMs).

What's real and what's not

• 50 paper abstracts — verbatim from arXiv, mirrored under arXiv's CC0-equivalent abstract policy. See LICENSES.md for the full provenance ledger and INVENTORY.md for the per-sub-topic index.
• ~12 OSS repo READMEs (Phase 3, not in this commit) — mirrored from public projects with license attribution preserved. Most are MIT / Apache-2.0 / BSD; a few canonical era-defining repos (Inria's gaussian-splatting, NVlabs's instant-ngp) ship under research-only licenses and we mirror only the README with attribution.
• 6 daily digests + 2 weekly syntheses + 3 open questions (Phase 4, not in this commit) — hand-written narrative glue, original work for the demo. They cite real corpus papers/repos but the prose is fabricated.
• 5 agent session transcripts under .agent-sessions/ — scripted demonstrations of Tesserae's MCP query workflow. Tool names are real (verified against tesserae/mcp_server.py); the conversations themselves are written for the demo. These ship as file-tree artifacts in the corpus only — the Tesserae harness session importer (tesserae/harness_sessions.py) ingests native Claude Code / Codex transcripts (from ~/.claude/projects/... and ~/.codex/sessions/...), not the synthetic showcase format used here. Visitors who clone the repo can read them directly; they do not appear in the deployed site's /sessions/ index.

The goal is to make the GitHub Pages demo a believable "your literature review could look like this" experience without claiming any of the editorial work (digests, syntheses, sessions) is genuine research output.

Why 3D reconstruction

3D reconstruction is a hot, fast-moving area with:

• a clearly identifiable canonical paper lineage (NeRF → instant-NGP → 3D Gaussian Splatting → variants),
• dense cross-citation that produces an interesting knowledge graph,
• a large pool of widely-cited public papers with permissive abstracts,
• many real public OSS implementations whose READMEs make natural paper↔repo bridges.

That gives the deployed graph view a recognizable dense cluster around each sub-topic, plus visible bridges between Gaussian Splatting and SLAM, between NeRF and surface reconstruction, and between diffusion-based generation and feed-forward LRMs.

Sub-topic distribution

See INVENTORY.md for the full per-paper listing.

Reading order for a first-time visitor

1. README.md (you are here) — framing.
2. INVENTORY.md — the map of the corpus, grouped by sub-topic, with links to each abstract.md.
3. LICENSES.md — the provenance ledger.
4. data/research/papers/arxiv-2308-04079/abstract.md — sample paper page (3D Gaussian Splatting).
5. (Once Phase 4 ships) data/research/weekly/*/synthesis.md — example weekly synthesis.
6. (Once Phase 5 ships) .agent-sessions/*/transcript.jsonl — example agent session.

File layout

examples/demo-corpus/
├── README.md           # This file
├── LICENSES.md         # Provenance ledger (every external source)
├── INVENTORY.md        # Per-sub-topic paper index
└── data/
    └── research/
        └── papers/
            └── arxiv-YYYY-NNNNN/
                └── abstract.md   # YAML frontmatter + verbatim CC0 abstract

Phases 2–5 add paper.md summaries, data/research/repos/, data/research/daily/, data/research/weekly/, data/research/questions/, and .agent-sessions/.

Phasing

This corpus is built in six commits following docs/superpowers/specs/2026-05-15-demo-corpus-plan.md. This commit is Phase 1 — license hygiene, the inventory, and 50 CC0 abstracts. Phases 2–6 (paper bodies, repo READMEs, digests, syntheses, sessions, CI wiring) follow in subsequent commits.

Snapshot date

Everything in this corpus is pinned to 2026-05-15. It is not a live feed.