Papers (website and owner)

complement of /references (outbound). Every inbound edge's from_paper_id is by construction a real in-library Paper, unlike outbound references which may be dangling, so this can return PaperResponse directly with no unresolved-reference handling.

`transcripts/{paper_id}/images/{filename}` in S3 (services/pdf_extraction.py's image_s3_key). The transcript markdown references figures as `images/{filename}`, so this is the "frontend route" that key scheme's docstring anticipates -- the paper reader rewrites those relative refs to this URL.

Deliberately no DB lookup and no auth, same as /transcript: the paper id and filename fully determine the S3 key, and only figures already extracted to our bucket are reachable. Missing object -> 404, unsafe filename -> 404.

results/limitations/prerequisite concepts (services/distillation.py).

404s only if the paper itself doesn't exist. If the paper exists but hasn't been distilled yet (still pending, or its PDF extraction hasn't succeeded yet -- see distill_paper's gating), this returns a 200 with every field null, the same "async-generated content that may not have completed yet renders as a null field, not a 404" convention PaperResponse.summary already uses.

claim-extraction + citation-resolution + judgment pipeline (Section 3 -- reference validation, on-demand per-draft). Same background_tasks.add_task fan-out ingest_paper uses for distill_paper_task/embed_paper_task, rather than awaiting the full multi-LLM-call pipeline inline on the request.

Returns immediately with a "triggered" status; results land as PaperCitationValidation rows, readable via the validate_references MCP tool (app/mcp_server.py) once the run completes.

papers whose PDF extraction has already succeeded but haven't been chunked+embedded yet -- same rationale/shape as backfill-insights above, for the RAG chunk store instead of PaperInsight.

before embedding generation was added to ingest_paper. Same authenticated-but-not-owner-gated pattern as the other bulk admin endpoints on this router (fetch-all-citations, extract-all-references).

extraction has already succeeded but haven't been distilled yet -- covers both papers ingested before distillation existed, and the ordinary steady-state case where distill_paper_task's ingest-time fan-out call (services/ingest.py) ran before the separately-invoked, asynchronous pdf_extraction Lambda had produced a transcript yet, and so no-oped (see services/distillation.py::distill_paper's docstring). Same authenticated-but-not-owner-gated bounded-concurrency bulk-admin pattern as backfill-embeddings above.

scheduled Lambda runs (see infra/graph_growth.tf) -- useful for testing without waiting for the schedule.

This runs inside the backend Lambda (120s timeout), unlike the nightly one (900s), so the run is given a budget that ends it well before the Lambda would be killed.