Render the dashboard (README.md) and the daily summary (SUMMARY.md)

README.md is what GitHub renders on the repo home page — it IS the radar's output surface. It is fully derived from TRENDS.md; regenerate it from scratch on every run, never patch it. Design goal: one screen, everything clickable.

Layout (exactly these blocks, in this order)

1. Title — # AI Radar.
2. Badge row — four shields.io static badges, style flat-square:

trends-<total>-3266ad, accelerating-<n>-e8590c, watchlist-<queue count>-6c757d, updated-<YYYY--MM--DD>-2f9e44 (escape - in dates as --).

1. Intro line — one sentence: what the radar tracks, for whom, and that the

page is generated from TRENDS.md.

1. Digest — **Since last scan (YYYY-MM-DD):** + 3–4 bullets MAX, in

priority order: stage moves, queue promotions/drops, strongest new evidence. Every bullet must contain at least one link (trend anchor or evidence URL). If nothing changed: one bullet, "Quiet scan — no ledger changes."

1. Pinned topics (⭐) — ## ⭐ Pinned topics then a table of ONLY the trends

with pinned: true, same columns as the main table. These are the curator's standing-watch axes; they appear here AND in the main table (where their trend label is prefixed with ⭐). Sort by last_evidence descending.

1. Trends section — stage strip on one line

(🌱 n · 📈 n · 🚀 n · 🌊 n · 🏔 n · 📉 n · 💤 n), then ONE table | trend | stage | latest signal | with ALL trends (pinned included, prefix their label with ⭐), sorted accelerating → emerging → seed → mainstreaming → saturated → declining → dormant, then by last_evidence descending:

• trend = short label (≤5 words; reuse previous README labels for

stability) linking to the trend's anchor in TRENDS.md (see anchor rule); prefix with ⭐ when pinned: true.

• stage = emoji + word.
• latest signal = the last_evidence date linking to the evidence URL

whose date equals last_evidence (first match in list order); if none matches, link the top evidence item.

1. Worth studying — the study_shelf section of TRENDS.md, newest first,

up to 12 bullets: [name](url) — one line of why (the shelf persists ~20 picks / 30 days so a reader away for a couple of weeks can still catch up).

1. Community pulse — ## Community pulse (Phase 3): up to 5 bullets distilled

from the latest radar-pulse intake — non-evidence, link-only, NEVER naming individuals (link the thread/profile, summarise). If no pulse data yet, one line: "No community pulse sampled yet." This section is clearly marked as unverified sentiment, separate from the verified trend table.

1. Output map — one line of links: TRENDS.md, watchlist (queue count) →

TRENDS.md#observation_queue, reports/ + newest daily, newest weekly ("weekly: none yet" if absent), AGENTS.md, SOURCES.md.

SUMMARY.md — the daily summary page (daily runs only)

A second derived page, for a reader who wants the run's RESULTS in ten seconds. Write it on EVERY daily run, in the same commit as README.md, OVERWRITING the previous version (history lives in reports/). Weekly runs do not touch it. It states the same facts as today's report and README digest — no new claims; every link must already be in TRENDS.md or today's report.

Exact shape — nothing above, between or below these lines:

# AI Radar — YYYY-MM-DD

- **New trend** · [<trend label>](TRENDS.md#<anchor>) — <what it is>.
- **Stage** · [<trend label>](TRENDS.md#<anchor>) <old> → <new> — [<source title>](<url>): <what it shows>.
- **New** · [<source title>](<url>) — <what it shows> (→ [<trend label>](TRENDS.md#<anchor>)).
- **Study pick** · [<source title>](<url>) — <why it is worth reading>.

Rules:

• At most 6 bullets, most significant first. Only the four kinds above; any may be absent,

any may repeat. One line each; the text after the dash is ≤ 15 words (the (→ trend) link not counted).

• RESULTS ONLY — what was found, never how it was found or what the radar did with it.

Nothing about the radar's own machinery: no sweeps, lanes, passes, sources checked or degraded, heals, queue/watchlist counts or moves, cap swaps, capture-leak, diffs, convergence flags, "Nth independent group" counts, unchanged repos, amendments, dormancy rescues, backfills, next steps.

• Verified results only: new trends, stage moves, new evidence, study picks. Unverified queue

signals are not results — leave them out. Do not list the same item twice (a study pick already listed as New is not repeated).

• No narrative: no evaluative adjectives (striking, genuine, major, finally, notable, real),

no scene-setting, no "a sign that…" interpretation. State the fact. A result that is a claim not yet vetted keeps a one-word tag at the end: (claim) or (preprint).

• Trends by their README short label (never a bare id or number); sources by their own

title, cut at its colon when the title is long.

• A run with no ledger change writes exactly one bullet: - Nothing new in this run.

Check before commit (a hit is allowed only inside a source's own title or a trend label):

test "$(grep -c '^- ' SUMMARY.md)" -le 6 || echo "TOO MANY BULLETS"
grep '^- ' SUMMARY.md | grep -v '](' | grep -v 'Nothing new in this run' && echo "BULLET WITHOUT LINK"
grep -q "^# .* — $(date +%F)$" SUMMARY.md || echo "WRONG DATE"
grep -niwE 'sweeps?|swept|lanes?|pass(es)?|queued?|watchlist|cap-swaps?|capture-leak|degraded|heal(ed)?|sitemap|diff|coverage|convergence|repo-watch|amendments?|dormancy|backfill|triage|flagged' SUMMARY.md && echo "PROCESS WORDS — rewrite"
grep -niE '[0-9]+(st|nd|rd|th) (independent )?(group|org|organi[sz]ation)' SUMMARY.md && echo "INTERNAL COUNT — drop it"

Anchor rule (GitHub heading slugs)

For a heading ### [id: slug-NNN] Title, the anchor is: lowercase the whole heading text, delete every character that is not a letter, digit, space or hyphen (no replacement — Prefill/decode → prefilldecode), then replace spaces with hyphens. Example: ### [id: pd-disagg-002] Prefill/decode disaggregation as the standard LLM serving architecture → TRENDS.md#id-pd-disagg-002-prefilldecode-disaggregation-as-the-standard-llm-serving-architecture. Anchors are recomputed at every render, so title edits stay in sync.

Checks before commit

• Badge counts = stage-strip counts = table row count = number of ### [id:

blocks in TRENDS.md.

• Every table date equals that trend's last_evidence exactly; every row has

two working relative/absolute links.

• No Mermaid, no images besides the badges, no sections beyond the blocks above.
• Pinned-topics table contains exactly the pinned: true trends; every pinned

trend is ⭐-prefixed in the main table too.

• Community-pulse items are link-only and name no individuals.
• Commit README.md together with the TRENDS.md change that triggered it.