Radar Self Eval

Self-evaluation

The radar must measure whether it is winning, not assume it. Three parts: metrics (every week), retrospective (monthly), curator proposals (every week).

1. Weekly metrics

Compute from the week's daily reports (primary source — they list ledger changes) and, where needed, git log -p -- TRENDS.md since the previous weekly commit:

• queue funnel: items added / promoted to trend / dropped / older than 14 days and still unverified (stale)
• ledger: evidence items added, stage moves (up and down)
• exploration compliance: daily runs whose logs/source_rotation.md line contains a venue-exploration entry ÷ daily runs executed
• off-axis rate: share of new queue items that do NOT match any axis in logs/strategy_notes.md (judgment call — name them)
• discovery lag: for each evidence item added this week, the days between its evidence-line date and the date it entered the ledger (commit date via git log -p -- TRENDS.md, or the daily reports). Report the median, split by channel — exploration finds vs queue promotions (backfill) — plus the backfill share of all new evidence. This is the daily-ness KPI.
• coverage (the self-check that auto-detects the listed-but-never-swept failure class): for each "swept every run" heading in SOURCES.md (lab blogs, YouTube curators, pointer/digest blogs, discovery venues, tool-discovery channels), enumerate EVERY entry under it — INCLUDING bullets nested under a sub-label (e.g. a "non-GitHub channels:" sub-block), not just top-level list entries — and diff that full list against the week's logs/source_rotation.md lines, classifying every listed source as opened, degraded, or MISSING (in SOURCES.md but never in any log line this week). MISSING = a coverage lie — NAME them; never report "0 missing" without having run the full nested list-vs-log diff. A source MISSING or degraded for the whole week is a heal-or-REMOVE candidate (see Amendments): the registry must be honest about what it actually sweeps. AMENDMENT (2026-07-11, W28, applied — proposed W27 as Proposal D, signal persisted): "opened" is not enough on its own — when SOURCES.md documents a CONFIRMED heal for a source (a specific working method, e.g. a search+extract pipeline replacing a broken feed), check whether the daily logs actually RAN that method this week, not just re-logged the source's OLD pre-heal degraded: label. A source can pass "opened" for weeks while quietly reading only a stale/broken slice of its own surface (multi-section sources like Anthropic /research+/news are the same failure class as an unused heal) — name any source where this happened as its own finding, distinct from a plain MISSING.
• routing-leak: queue items tagged with (or clearly on) an EXISTING trend's axis yet absent from that trend's evidence — on-axis primaries hoarded in the queue instead of routed to evidence. Count and name them; each is a routing miss to correct this week.
• source-discovery: from SOURCES.md → "Discovered-source candidates" and the week's reports — candidates STAGED this week / PROMOTED to a swept source / PRUNED, plus discovery-latency = runs (or days) between an org's FIRST appearance in the staging tally and its promotion. The tool-author/vendor analog of routing-leak: a recurring on-axis org that stays un-promoted (≥2 weeks in the tally) is a coverage leak — NAME it and promote-or-drop it now. Zero candidates staged for two weeks while new evidence keeps arriving from the SAME few orgs is itself a signal the capture-the-source step isn't firing.
• capture-leak (the notes-prose blind spot, complementary to routing-leak): scan every trend's notes field AND the week's reports for named primaries (arXiv ids, repo / release URLs) that appear NOWHERE as a discrete observation_queue item or an evidence line on any trend. Each is a capture leak — a primary the radar named but never captured as a first-class item, so it never reached study-pick selection or the convergence check. NAME them and queue them now. routing-leak only scans the queue, so an item that never reached the queue is invisible to it; this metric is what catches it.

Append ONE dated line to logs/calibration.md (the externalized self-eval log; the ## calibration section of TRENDS.md is now only a pointer): - YYYY-MM-DD — W<nn>: queue +a/→p/−d/stale s · evidence +e · moves m · exploration c/r · off-axis o/a · lag expl Xd / backfill Yd (Z%) · coverage <opened>/<listed> (miss <n>, degr <d>) · routing-leak <n> · capture-leak <n> · src-disc <staged>/<promoted> (lat L) and include the same numbers, readable, in the weekly report.

Interpretation thresholds (act, don't just log):

• exploration compliance < 100% → call it out; two weeks running → propose a daily prompt/skill refinement.
• stale > half the queue → scanning breadth exceeds verification capacity: propose narrowing or a verification-only day.
• off-axis = 0 for two consecutive weeks → anchoring warning, even if the tunnel-vision check passed.
• exploration-channel median lag > 7 days → the venues are stale: rotate them.
• backfill share not shrinking for two consecutive weeks (once the bootstrap queue is cleared) → the radar is doing literature review, not daily detection: propose rebalancing scan time toward exploration.

2. Monthly retrospective (first weekly run of the month: date +%d ≤ 7)

Goal: ground truth — did the radar see early what later became big?

1. Pick 2–3 items that clearly became big in the past month by BROWSING venues (open the pages — evidence rules apply): HF papers trending, two major lab blog indexes, one arXiv listing. Pick what is everywhere, not what is interesting.
2. For each, find the radar's earliest mention: git log -S"<term>" --oneline -- TRENDS.md plus the queue and reports.
3. Classify: HIT-early (tracked ≥7 days before it was everywhere), HIT-late, MISS.
4. For every MISS: add it to observation_queue now, and name the venue or axis that would have caught it — feed that into the proposals.

Log in logs/calibration.md: - YYYY-MM-DD — retro M<mm>: <item> — HIT-early|HIT-late|MISS (first seen <date or never>), …

3. Amendments (proposed → cooling period → applied)

The radar runs unattended and may change its own operating instructions — under the autonomy contract in AGENTS.md. Every week, up to 3 proposed amendments: routine edits (exact replacement text), axis drops/merges/ supersessions, exploration-budget changes. Each must cite the metric or retrospective result that motivates it.

The new coverage/routing metrics feed amendments directly (this is how the radar self-corrects the failure classes the curator used to catch by hand):

• a source MISSING or degraded all week → first a radar-source-heal attempt; if it was already healed and still fails, a heal-or-REMOVE amendment (remove it from the SOURCES.md "swept every run" list, or escalate the secret/allowlist it needs) — a first-class amendment, because a listed-but- unswept source is a coverage lie.
• a non-zero routing-leak → not an amendment but a this-week FIX: route the hoarded on-axis queue items to their trend's evidence now; if it recurs, amend the routing instruction.

Lifecycle:

• Week N: propose — log in calibration and the weekly report.
• Week N+1: apply IF the motivating signal persists AND no dated curator veto appeared in logs/strategy_notes.md (silence is consent). One dedicated commit per amendment. If the signal vanished, drop the proposal and say so.
• After applying: if calibration metrics worsen for two consecutive weeks, git revert the amendment and log the rollback.

Scope-axis changes are recorded as "radar-adopted" dated entries in logs/strategy_notes.md, explicitly naming what they supersede. Curator entries are never deleted or edited — vetoes and mission input belong to the curator alone. The immutable sections listed in AGENTS.md are out of bounds: an amendment touching them is invalid.

Boundaries

• Evidence rules apply to the retrospective: cite only opened pages.
• logs/calibration.md is append-only: never edit or delete existing lines.
• If a metric shows the same procedural failure twice, trigger a skill refinement per the maintenance policy in AGENTS.md (dedicated commit, explained in the report).