Patent-landscape-style foresight on the AI frontier. It began as foresight on AI-safety research — treating papers the way patent analysts treat filings (what is worked on, how fast, by whom, and crucially what isn't) — and has grown into a full emerging-risk pipeline: from the research-trend signal, through harm/misuse vectors and a scored, evergreen risk register, to 6–24-month scenarios and a real-world incident benchmark.

The original spine is the capability-vs-safety divergence: for paired topics (e.g. agentic/autonomy capability ↔ agentic monitoring), it measures whether capability research is accelerating while the paired safety work stays flat. On top of that, a Claude-powered Foresight layer turns the signal into novel, web-verified, scored risk calls — re-classified by which real-world misuse they could enable, developed into scenarios, benchmarked against actual incidents, and exportable as analyst-ready intelligence-estimate and tabletop packs.

🧭 Analyst's note — read this first

signal-lag measures research attention, not research success. A spike in a topic can mean a breakthrough or a field thrashing against a wall — those look identical in volume. So treat this as a triage instrument that shows you where to investigate, not what to conclude. The Sentiment layer (share of critical / limitation-focused papers) exists specifically to help tell those two cases apart. Everything shown is real data — there is no synthetic or demo content anywhere.

▶️ Live demo: https://signal-lag-aw6rmrmp65nit9m9f8wmcq.streamlit.app

The live dashboard runs on real arXiv + Semantic Scholar data, refreshed weekly by a GitHub Action that pulls fresh papers, re-runs the analysis, and publishes a snapshot the app reads (see How the live data works). It only ever shows a real, published snapshot — never synthetic or demo content. Hosting is free on Streamlit Community Cloud; to run your own copy, see Deploy your own.

1. Ingestion — pulls papers from the arXiv API (cs.AI, cs.LG, cs.CL, cs.CR, cs.CY, cs.CV) over a configurable window, stratified by quarter for even time coverage, and enriches them via Semantic Scholar (citation counts, outgoing references, author ids + affiliations, TLDRs, influential-citation counts, venue), OpenReview (venue papers + peer-review scores), and lab/blog RSS (a capability-leading signal). (OpenAlex is configured but currently unreachable from the CI runner; Semantic Scholar supplies its signals instead.) Cached in local SQLite; rate-limited with backoff. All sources are fail-soft — one being down just omits its signal.
2. Topic modeling via embeddings — embeds every abstract (sentence-transformers all-MiniLM-L6-v2, with a TF-IDF+SVD fallback), then discovers emergent topics (HDBSCAN, k-means fallback) and tags papers against a supervised taxonomy of 8 safety + 6 capability topics via cosine similarity to topic centroids — with per-topic thresholds, a high/medium/weak confidence tier per tag, and a weekly LLM-judged precision audit per topic (Methodology tab) so over-inclusive topics are visible and correctable instead of silently inflating volumes.
3. Velocity analysis — submission rate per topic per quarter; flags acceleration/deceleration inflections and newly-forming clusters (the current incomplete quarter is dropped from trend math), plus a statistical warning layer: CUSUM persistent-shift detection, change-point quarters, capability→safety lead-lag cross-correlation per pairing, and trend-forecast ranges with deviation alerts.
4. Sentiment / confidence layer — the share of critical / limitation-focused papers per topic (embedding recall, LLM-verified precision on the recent window), whether it's rising (early confidence-erosion warning), a volume×critique quadrant map, Wilson confidence intervals, and false-confidence alerts (capability growing while self-critique falls and paired safety stays flat).
5. Citation layer — verifies cross-domain borrowing via real references (Semantic Scholar): which capability papers actually cite core safety work, not just share its vocabulary — plus a capability×safety cross-pollination matrix, a bridge-paper detector, and a safety-impact leaderboard (the safety papers capability builders actually cite). Citation velocity (movers/sleepers) is rebuilt from signal-lag's own weekly snapshots of per-paper citation totals, since OpenAlex's yearly series is unreachable from the CI runner.
6. Divergence layer (the headline product) — per configured capability↔safety pairing, flags where capability is accelerating but the paired safety topic is flat — with volume-balance vs growth-balance made explicit, high-confidence volumes next to raw, a confidence-adjusted gap (growth weighted by each side's self-critique posture), cumulative monitoring-debt curves, and the lab-announcement → safety-response lag: how long the paired safety literature takes to measurably answer each lab announcement (median lag, unanswered-after-N-weeks).
7. Research ecosystem — institutions from Semantic Scholar author affiliations: industry-vs-academia split, top and fastest-growing (institution, topic) pairs, plus experimental capability→safety author migration tracking.
8. Weekly Claude analysis — once per refresh, Claude (claude-opus-4-8) reads the computed metrics + real abstracts and writes the analytical headline, a per-tab read, and a what-it-does / why-it-matters note per driving paper (baked into the snapshot).
9. Foresight layer (Claude) — the emerging-risk engine, all baked into the snapshot:
   • Foresight Gap — crosses the research signal with a living societal-context file to surface novel cross-domain risks, web-checks each for novelty + disputes, and backfills for quality (methodology §10).
   • Harm/misuse vectors — re-classifies the same papers by which real-world misuse they could enable, on a 0–24-month horizon (§13).
   • Scored risk register — every risk scored by severity × likelihood × exposure × trajectory plus confidence / evidence-strength / actionability tie-breakers (a true forced ranking), accumulated into an evergreen register with watchlist statuses (open / strengthening / weakening / dormant), a persisted counterevidence trail, filters, a per-risk drill-down dossier, and a calibration panel (§14). Each risk carries epistemically-labeled claims (observed / inferred / speculative), falsification conditions (upgrade / downgrade / invalidate if), and a "so what" action map (eval to run, benchmark to watch, mitigation, owner).
   • Scenarios + exports — 6–24-month scenarios from the top risks, plus one-tap intelligence-estimate and tabletop-exercise packs (§15).
   • Incident benchmark — real, web-sourced incidents (credibility-graded on AI-involvement / attribution / source quality, cleaned + de-duplicated) crossed against the research signal into a leading-vs-lagging 2×2, with an accruing early-warning calibration record (§16).
   • Enablement map — capability topics → harm vectors → real incidents as a Sankey (edges = papers co-tagged to both tracks; association, not proof).
   • Plain-terms briefing — the top risks explained for non-specialists (evidence, context, the gap, the tool's own skepticism, bottom line), as a first-class view.
10. Output — a Streamlit dashboard: 8 tabs led by a self-contained Weekly Summary, an always-visible BLUF sidebar (each judgment with an analyst so-what), a per-tab "since last refresh" delta panel, and exports: markdown briefs / intelligence estimate / tabletop pack / one-page brief, plus structured CSV/JSON (register, incidents, velocity, lab-lag, citation matrix, full snapshot).
11. Operations — tiered LLM usage (a cheap model for high-volume classification, the strong model for synthesis/verification), a 21-day novelty-verification cache keyed by risk fingerprint, and idempotent per-refresh histories (register, benchmark, citation counts) committed by the weekly workflow.

signal-lag pulls from several free sources (all config-driven, all fail-soft — a source being down or rate-limited just omits its signal):

• arXiv — the papers themselves (title, abstract, authors, dates), from the cs.AI, cs.LG, cs.CL, cs.CR, cs.CY, cs.CV categories (extendable in config).
• Semantic Scholar — the primary enrichment source: citation counts, outgoing references (for citation-flow verification and the cross-pollination matrix), stable author ids + affiliations (the research-ecosystem view), TLDR summaries, influential-citation counts, venue, and fields of study.
• OpenAlex (configured, currently dormant) — unreachable from the CI runner; Semantic Scholar supplies its signals instead. Citation velocity is rebuilt from signal-lag's own weekly snapshots of per-paper totals.
• OpenReview (optional, config-gated) — venue papers (e.g. ICLR/NeurIPS) added as records with peer-review scores, a quality/heat signal papers-only sources lack.
• Lab/blog RSS (optional, config-gated) — posts from major labs (Anthropic, OpenAI, DeepMind, …) as a capability-leading signal, shown separately from paper velocity since they aren't papers.
• Claude (Anthropic API) (optional, config-gated) — not a data source but an analysis layer: once per refresh the computed metrics + the real abstracts are sent to claude-opus-4-8, which writes the analytical headline, the per-tab read, and the per-paper what-it-does / why-it-matters. Baked into the snapshot; fail-soft (skipped with no ANTHROPIC_API_KEY).

Each source is an isolated client; new ones slot in alongside arxiv_client.py without touching the rest of the pipeline.

What this covers well: arXiv is where the large majority of frontier AI/ML/NLP research appears first, so coverage of the fast-moving preprint literature is high.

Temporally-stratified sampling. arXiv publishes hundreds of papers per day, so naively pulling "the newest N" would only span days. Instead ingestion samples up to max_per_period papers per category per quarter across the whole window, giving even time coverage. Velocity therefore tracks each topic's share of activity per quarter (the trend that divergence relies on), not raw absolute counts.

What it does not cover (yet): further arXiv categories (cs.RO, stat.ML, …), and venues outside arXiv + the configured OpenReview conferences (ACL Anthology, PMLR, journals). So treat it as high coverage of the AI preprint literature, not "everything published." The arXiv categories and OpenReview venues are config-driven (extend them in config/settings.yaml); adding a genuinely new source means a small ingestion client alongside arxiv_client.py.

Two complementary layers, both semantic (embedding-based), not keyword matching:

1. Supervised tagging against the taxonomy. Each topic in taxonomy.yaml has a few seed phrases. Those are embedded and averaged into a centroid vector per topic. Every paper's title+abstract is embedded into the same space, and its cosine similarity to each centroid is computed; if it clears tag_threshold (default 0.28) the paper gets that tag (up to max_tags_per_paper). Because it's semantic, a paper about "models that strategically hide their objectives" tags to deceptive alignment even without those literal words.
2. Unsupervised clustering for emergent topics. All paper embeddings are clustered with HDBSCAN (auto-discovers the cluster count, marks outliers as noise). Each cluster is auto-labeled by its most distinctive terms (c-TF-IDF). This surfaces directions that aren't in the predefined taxonomy.

Embeddings use all-MiniLM-L6-v2 (sentence-transformers); offline it falls back to a TF-IDF + SVD vectorizer. Categorization quality depends on the seed phrases and threshold — both tunable in YAML.

• Why embeddings, not keywords: keyword filters can only find topics you already named. Embedding clusters surface emergent directions; the supervised taxonomy is an overlay on top, not a replacement.
• Negative/sentiment signal: a "negativity" centroid is built from limitation/ failure seed phrases (negativity_seeds in taxonomy.yaml). A paper is critical when its abstract embedding is close to that centroid; per topic we track the critical share and its quarter-over-quarter trend. Rising critical share (esp. with flat volume) is flagged as eroding confidence. Hybrid LLM precision: the embedding centroid alone mistakes academic negation ("we overcome the catastrophic failures of prior methods") for criticism, so the recent-window papers it flags critical are re-checked by Claude and false positives are downgraded before the trend is computed (sentiment.llm_verify; bounded to that subset, fail-soft).
• Citation-flow verification: the cross-silo "borrowing" claim is checked against real citation references (each paper's outgoing bibliography, by arXiv id, from Semantic Scholar — OpenAlex is unreachable from the CI runner) — a capability/applied paper that actually cites a core safety paper, not one that just shares its vocabulary. Positive-only: a verified citation is strong evidence; absence is inconclusive, never "they ignore safety work" (citation_flow.enabled).
• Author migration (experimental): using Semantic Scholar author IDs, authors who were capability-dominant historically and whose recent papers enter a safety/oversight topic are flagged as a capability→safety talent flow — a leading indicator. Noisy by construction (stratified sample + imperfect IDs): it informs the brief, never gates an alert (analysis.author_migration).
• Labs-lead signal: lab/blog posts are embedded and tagged to topics, then shown against the paired safety topic's velocity — "labs announce → safety responds on a delay → the delay is the risk window."
• Velocity & inflection: counts are bucketed by calendar quarter. An inflection compares the mean of the last N quarters against the prior N (default N=2); a relative change beyond ±30% is an acceleration/deceleration.
• Divergence metric: for each pairing, gap = capability_growth − safety_growth. A pairing is flagged "lagging" when the gap exceeds the threshold and capability growth is positive. volume_ratio shows how lopsided the absolute volumes are now.
• Citation heat: rebuilt from signal-lag's own weekly snapshots of per-paper Semantic Scholar citation totals (movers = biggest week-over-week gains). "Sleepers" have a low early citation share but a high recent share — quiet papers now heating up.
• Everything is config-driven. The taxonomy, capability↔safety pairings, date range, caps, thresholds, and clustering choice all live in config/*.yaml — no code edits needed to retune.

A second weekly Claude pass (claude-opus-4-8, using the Claude API exactly like the analysis layer in “Data sources” — same client, same fail-soft, same baked-into-snapshot, no page-load calls). Its job is to surface novel, not-yet-in-the-news risks that live in the seam between AI research and broader societal forces — risks no single community is tracking because they sit between domains.

How it works:

1. Signal digest — it pulls the strongest signals already computed this week: flagged capability-vs-safety divergences, velocity inflections, rising critical-share (eroding-confidence) flags, quadrant emerging/white-space topics, citation sleepers & rapid-growth papers, new emergent clusters, recent lab activity, and what changed week-over-week (so it weights movement, not just static state).
2. Scanning framework — a fixed STEEP/PESTLE-plus taxonomy of domains (Social, Technological, Economic, Environmental, Political, Legal/Regulatory, Security/Geopolitical, Demographic) so the synthesis is comprehensive by construction and never tunnels on technology alone. It defines dimensions, never specific trends.
3. Living societal context — config/context.md, a user-maintained file where you paste the current real-world state across those domains and keep it updated week to week. It is not baked into code, so it never goes stale; any examples in it are illustrative of the format only — not a prescribed or exhaustive list, and the synthesis is explicitly told never to treat them as the only factors that matter. If the file is missing/empty the pass still runs (just without the societal layer).
4. Live web brief (current ground truth) — because fast arXiv data crossed with a hand-maintained context.md risks anchoring on stale facts, an optional pre-synthesis web search (analysis.foresight.live_context) pulls the current, dated status of the flagged topics' real-world developments. The synthesis verifies any date/policy claim against this live brief and prefers it over the standing context where they conflict — complementing, never replacing, the analyst's file. Fail-soft; shown in its own expander.
5. Synthesis — Claude returns 2–4 candidate risks, each with a fixed six-part structure: risk statement · derived-from (citing the actual digest signals, so it's traceable) · why it's under-discussed · mechanism · leading indicator · calibration · extrapolation (an honest flag of what goes beyond the data). It's instructed to ground every claim in the provided signals and to refuse to restate well-known AI risks. It also gets the citation-verified borrowings (use as evidence, not vocabulary; absence inconclusive) and the experimental author-migration signal (soft corroboration only).

The synthesis is tuned to the tool's real strength:

• Research-trend anchor (the proprietary edge). Each risk leads with a signal only this tool has — a safety subfield whose velocity is decelerating or whose critical share is rising — then crosses it with the societal context. The research signal makes it novel; the societal cross makes it a real-world risk.
• Cross-silo seams. It prioritizes risks where two distinct expert communities each track one half and nobody connects them, and names which community sees which half.
• Framing inversions. It rewards risks that invert a trend everyone treats as simply good or bad (e.g. transparency regulation freezing in an unsound standard).
• Calibrated on contested ground. When a risk leans on a disputed or inferential claim — including over-reading the tool's own trend metric as causation — confidence is lowered explicitly, never laundered into a confident claim.

Novelty verification. After synthesis, each candidate is run through a web search (Claude's server-side web search) that looks for both confirming and disputing coverage, returning a prior-coverage check, a verified novelty rating (genuinely unsurfaced / partially anticipated / already widely discussed), disputing sources, and a recalibrated confidence. Already-discussed risks are tucked into a collapsed expander (flagged, not hidden). Quality over quantity: if too few candidates survive verification as fresh, the synthesis automatically re-runs for different seams (up to max_rounds) and verifies those too — so a week where several turn out already-covered still surfaces genuine ones. Verifications run in parallel and are baked into the snapshot (cached — never at page load). This is the calibrated posture: generate candidate risks, then check them against current coverage before surfacing them — distinguishing a genuine seam from something that just isn't in the index yet.

These are AI-surfaced candidate hypotheses for an analyst to pressure-test — not predictions. The model widens the aperture; human judgment goes on top. The tab shows the digest, context, framework, and per-risk prior-coverage check, so the reasoning is fully transparent. Config lives under analysis.foresight in settings.yaml (enable, number of risks, context-file path, verify_novelty, web-search tool version). Like the rest of the Claude layer it needs the ANTHROPIC_API_KEY repo secret; without it the tab shows an honest "unavailable" message.

The quarterly view is a slow-moving baseline (3 years of quarterly trends). Two layers add a timely and a longitudinal dimension:

• "This week" toggle (Summary + Foresight Gap). Alongside the overall view, a toggle analyzes only the papers submitted in the last window_days (default 7): topic counts (safety/capability), a focused Claude "what landed this week" summary, notable papers, and a full web-verified this-week Foresight Gap. It's anchored on the quarterly research- trend signal (as backdrop) then crossed with this week's papers + the societal context. The quarterly charts are untouched. An extra recent-window arXiv pull (recent_topup_days) guarantees the 7-day set is complete even if a category exceeds the quarterly cap.
• History tab (📜). Every refresh appends a compact briefing — headline, top foresight gaps (overall + this week), rising sentiment, and what-changed counts — to data/history.json (idempotent per date). The tab renders a metrics-over-time chart (safety-lag alerts per refresh) plus a reverse-chronological list. Records are compact by design (no abstracts/raw data) so the file stays lean; full per-week data is recoverable from the git history of data/snapshot.json.

Config lives under analysis.weekly in settings.yaml (enabled, window_days, recent_topup_days, and lighter max_risks/min_surfaced/max_rounds for the this-week foresight). Fail-soft like the rest of the Claude layer.

# Core: runs the whole pipeline + dashboard using built-in fallbacks
pip install -r requirements.txt

# Optional: best-quality backends (sentence-transformers + HDBSCAN; heavier)
pip install -r requirements-full.txt

sentence-transformers and hdbscan are the preferred backends. If sentence-transformers (or its model download) is unavailable, the embedder automatically falls back to a local scikit-learn TF-IDF + SVD vectorizer; if HDBSCAN is unavailable it falls back to k-means. The whole pipeline therefore runs even with only the core install, and even fully offline. The heavy backends are kept out of the default install so the hosted app stays lightweight.

The repo ships a synthetic fixture dataset so you can run the entire pipeline with no network access:

# 1. Load the bundled offline sample (1k+ synthetic papers)
python -m signal_lag.cli ingest --use-fixtures

# 2. Run the analysis and write the markdown brief
python -m signal_lag.cli analyze            # -> data/foresight_brief.md

# 3. Print the BLUF findings to the terminal
python -m signal_lag.cli signals

# 4. Launch the dashboard
python -m signal_lag.cli dashboard          # or: streamlit run signal_lag/dashboard/app.py

Run where arXiv and Semantic Scholar are reachable:

python -m signal_lag.cli ingest             # pull live arXiv + Semantic Scholar enrichment
python -m signal_lag.cli analyze
python -m signal_lag.cli dashboard

ingest is idempotent (upserts by arXiv id), so re-running extends the cache rather than duplicating. Use --no-enrich to skip enrichment, and enrich later to backfill:

python -m signal_lag.cli ingest --no-enrich
python -m signal_lag.cli enrich

Regenerate the fixtures (deterministic) with:

python scripts/generate_fixtures.py

• 📋 Weekly Summary — the self-contained briefing, in inverted-pyramid order: the Claude-written analytical headline (what the widest gap means and why it matters) with the driving papers, then the week's best (web-verified novel) foresight gaps, then what changed since the last refresh (new safety-lag alerts, newly accelerating topics, new citation sleepers), then a plain-language read of every other tab so you needn't open them, plus the labs-announce→safety-responds view. The former Signals tab is folded in here (full ranked list + downloadable brief).
• ⚖️ Divergence — for every capability↔safety pair, the recent growth rate of each side as horizontal bars (long capability + short safety = safety lagging). The table separates volume balance from growth balance, shows high-confidence volumes next to raw, ±1σ bands on growth, and a confidence-adjusted gap; below it, cumulative monitoring-debt curves and the 🛰️ lab-announcement → safety-response lag (median weeks to a measurable safety-literature response, per capability, with unanswered-after-N-weeks). This is the core product.
• 📈 Velocity — papers per quarter per topic over time, an inflection table, a this-week momentum vs expected table (Poisson z-scores), 🧪 statistical detectors (CUSUM shifts, change-point quarters, capability→safety lead-lag correlation, forecast-range deviations), and the strategic map (recent volume × growth: emerging / hot / cooling / white-space). Momentum (attention, not success).
• 🔬 Sentiment — the negative-signal layer: the share of critical / limitation-focused papers per topic (embedding recall, LLM-verified precision on the recent window) with Wilson 95% intervals, whether it's rising, 🟣 false-confidence alerts (capability growing + self-critique falling + paired safety flat), and a volume × critique quadrant map (growing-&-straining / growing-&-confident / contracting-&-critical / fading).
• 🔮 Foresight — a second Claude pass that anchors on the tool's own research-trend signal and crosses it with broader societal forces to surface novel, not-yet-in-the-news risks living in the seam between domains. Each candidate is then web-checked against current coverage and given a verified novelty rating (genuinely unsurfaced / partially anticipated / already widely discussed) with a prior-coverage check and disputing sources — already-discussed ones are flagged and demoted, not hidden. Each is a candidate hypothesis (not a prediction) naming which communities see which half of the problem, with mechanism, leading indicator, calibration, and an explicit extrapolation line. See methodology section 10. Exportable as a brief. A ⚠️ Harm vectors toggle on this tab adds the dual-use lens: the same frontier papers re-classified by which real-world misuse they could enable (cyber-offense, bio/chem uplift, influence ops, scams, agentic misuse, surveillance, model-weight exfiltration, jailbreak/guardrail-evasion, child safety, harassment), each with its own velocity — "which harms is the literature quietly making easier, and how fast?" over a 0–24 month horizon. Accelerating harm vectors feed the synthesis to frame capability→harm enablement risks. A foresight signal over research, not on-platform abuse telemetry. Harm taxonomy: config/taxonomy.yaml (harm_topics). Each risk also carries epistemically-labeled claims (🔵 observed · 🟠 inferred · ⚪ speculative), falsification conditions (what would upgrade / downgrade / invalidate it), and a "so what" action map. A 🕸️ enablement map (Sankey) connects capability topics → harm vectors → real incidents. A 📋 Risk register view scores every surfaced risk on severity × likelihood × exposure × trajectory (priority = severity × likelihood, 1–25) with confidence / evidence-strength / actionability tie-breakers (a true forced ranking; stale risks downgraded), accumulates them into an evergreen data/risk_register.json with watchlist statuses (open / strengthening / weakening / dormant), a persisted counterevidence trail, triage filters, a per-risk drill-down dossier, and a calibration panel — rendered as a priority matrix + ranked table. A 🧩 Plain terms view explains the top risks for non-specialists. A 🎬 Scenarios view develops 6–24 month scenarios from the top risks (drivers, leading indicators, branch points, mitigations, ICD-203 estimative likelihood), and two one-tap exports turn it all into analyst-ready products: a Strategic Intelligence Estimate and a Tabletop-exercise pack. A 🌐 Incidents view adds the all-source layer: real, already-occurred AI-misuse incidents (web-sourced from the AI Incident Database / OECD / news, verifiable + dated) tagged to the harm vectors and benchmarked against the research signal into a leading-vs-lagging 2×2 (🎯 foresight lead · 🔴 materializing · 🟠 active/known · ⚪ quiet).
• 🔍 Sources — the receipts: actual arXiv papers behind each topic (each with a what-it-does / why-it-matters note and a 🟢/🟡/⚪ tag-confidence badge, with a hide-weak-matches filter), 📈 citation velocity from signal-lag's own weekly count snapshots (movers + sleepers), the 🔗 citation cross-pollination section (capability×safety heatmap, bridge papers, safety-impact leaderboard), and the 🏛️ research ecosystem (industry-vs-academia split, top + fastest-growing institutions, from Semantic Scholar affiliations).
• 📖 Methodology — how every layer works, on the live data behind the current snapshot — including the weekly 🧪 tagging-precision audit (per-topic LLM-judged precision; the basis for the per-topic thresholds in config/settings.yaml) and 📤 structured CSV/JSON exports.
• 📜 History — a running record of past weekly briefings (headline, top foresight gaps, rising sentiment, what-changed) plus a metrics-over-time chart — the longitudinal view a single snapshot can't show.

The Summary, Divergence, Velocity, Sentiment, and Foresight Gap tabs carry a 🆕 This week ⟷ 📊 Quarterly toggle (defaulting to This week): alongside the 3-year quarterly view, "This week" analyzes only the papers from the last 7 days — topic counts, the this-week capability/safety attention split, this-week critical-share, a focused Claude read, notable papers, and a web-verified this-week Foresight Gap. The quarterly trend math (velocity/divergence/quadrant) is unchanged and clearly labeled this quarter — Quadrant stays quarterly-only (a 7-day volume×growth map isn't meaningful).

Two cross-tab layers ride on top: a 🎯 BLUF sidebar (widest safety lag, top register risk, top new risk, biggest weekly anomaly, harm vector to watch, median lab-response lag — each with an analyst so-what, plus one-tap downloads including a one-page brief), and a 🔄