[Atlas] Hypothes.is annotations on SciDEX wiki and hypothesis pages done

Effort: deep

Goal

Hypothes.is offers a battle-tested W3C-Web-Annotation overlay any
researcher already uses — installing the browser extension makes
every URL annotatable, and the Hypothes.is API exposes
group-scoped public annotations as JSON. SciDEX wiki pages and
hypothesis pages are stable URLs (~17K + ~310 today, headed to
100K + 10K) that researchers will increasingly want to mark up
with cross-references, corrections, and "interesting!" pins.
Surface those annotations in-page so the social trace is visible
to non-extension users, and feed annotation density back into the
quality-gate pipeline.

Acceptance Criteria

CREATE TABLE hypothesis_is_annotations (
        id TEXT PRIMARY KEY,
        target_url TEXT NOT NULL,
        scidex_artifact_id TEXT REFERENCES artifacts(id),
        author_account TEXT,
        author_display TEXT,
        text TEXT,
        quote TEXT,
        tags JSONB,
        group_id TEXT,
        created_at TIMESTAMP,
        updated_at TIMESTAMP,
        first_seen_at TIMESTAMP DEFAULT NOW()
      );
      CREATE INDEX ix_hypothesis_is_annotations_artifact ON
        hypothesis_is_annotations(scidex_artifact_id);

hits https://api.hypothes.is/api/search?uri.parts=
scidex.ai&order=desc&sort=updated&limit=200 every 15 min,
pages forward, upserts rows. Uses an etag/cursor (last
updated) so re-poll fetches only deltas. "Public annotations (N)" expander that lists annotations
with quote + author + date. No JS framework — use the
same vanilla pattern as the comments overlay. hypothesis_is_annotation_count to
quality_dashboard.py's rollup — high public-annotation
density on a hypothesis is positive evidence of community
engagement; tagged error or wrong annotations downgrade
the quality score. has a one-click button that links to the via-Hypothes.is
proxy https://via.hypothes.is/<page-url> so users without
the extension can still annotate. mirrored as an ai.scidex.annotation.claim AT Protocol
record (lexicon already exists in atproto/lexicons/)
published by a SciDEX-owned bot DID, so the same trace lives
on both surfaces. (Out of scope: full firehose ingest;
simple push-only here.) Hypothes.is response with 5 annotations across 3 URLs;
assert artifact resolution by URL parsing, dedup on
re-poll, AT Proto record emission shape.

Approach

Poller first — simpler than embedding the H. extension or

mirroring its full firehose.

URL → artifact resolver: parse the path components against

wiki slug / hypothesis id; fall back to leaving
scidex_artifact_id NULL.

AT Proto mirror runs as a deferred job; failures don't block

the poll.

Quality-gate signal is bonus — add behind a feature flag so

we can validate it doesn't perturb the existing leaderboard.

Dependencies

• q-perf-deferred-work-queue — runs AT Proto push as a job.
• atproto/lexicons/ai.scidex.annotation.claim.json — mirror

schema.

Dependents

• q-integ-bluesky-publish-pipeline — annotations are an

evidence source for Bluesky-targeted summaries.

Work Log

2026-04-27 — Implementation (task:3c70b9b8-0137-437c-a74b-8f5b82b4bbe4)

• Schema: migrations/20260427_hypothesis_is_annotations.sql — hypothesis_is_annotations

table with all spec-required columns plus GIN index on tags and target_url/updated_at indexes.

• Poller: scidex/atlas/hypothesis_is_poller.py — polls

api.hypothes.is/api/search?uri.parts=scidex.ai, uses system_settings cursor for delta
fetching, upserts with ON CONFLICT DO UPDATE, resolves /wiki/ /hypothesis/ /artifact/
/analyses/ URL paths to artifact IDs, mirrors new annotations to AT Proto
ai.scidex.annotation.claim via env-configured PDS credentials (no-op if absent).

• In-page overlay: _build_hyp_is_overlay_html() function in api.py; wired into

wiki_page(), hypothesis_detail(), and artifact_detail() — renders "Public annotations (N)"
<details> expander with quote/text/author/date/tag badges. Includes "Annotate on Hypothes.is"
via.hypothes.is link in the expander summary.

• Quality-gate signal: scidex/senate/quality_dashboard.py — added hypothesis_is_annotation_count,

hypothesis_is_annotated_pages, hypothesis_is_negative_count, hypothesis_is_new_7d
to headline and new community_annotations section; graceful fallback if table not yet migrated.

• Tests: tests/test_hypothesis_is_poller.py — 22 tests covering URL resolver, AT Proto

record shape, HTTP fetch stubs, upsert dedup, dry-run mode, AT Proto call counts.