# Vigie — Ranking method (v0) Published method. No invisible editorial boost without a logged reason. ## Principle Rank by **proximity to Quebec City life** and **recency**. Do not rank by ideology, party, or “trust score” invented in secret. Enrich tags (topic, claims, impact) are **proposals** when present (`enrich_status: proposed`). Until enrich runs, v0 ranks on geography from `sources.yaml` + recency only. ## Formula (v0) ``` score = w_geo * geo_proximity + w_recency * recency + w_tension * has_contradiction # 0 until clustering exists + w_impact * impact_weight # 0 until enrich proposes impacts ``` ### Weights (v0) | Weight | Value | Meaning | |--------|-------|---------| | `w_geo` | 0.60 | Nest proximity dominates | | `w_recency` | 0.40 | Fresh approaches matter | | `w_tension` | 0.00 | Reserved — enable when Issues exist | | `w_impact` | 0.00 | **OFF** — co-founder refuse 2026-09-16; unlock only when gates below pass | Weights must sum intent clearly; change them only by editing this file and noting the date below. ### w_impact enable gates (co-founder ACT 2026-09-16) **Decision: do not enable now.** Evidence: **3/354** candidates carry units (0.85%); **housing_count = 0** on this pulse; turning weight on would be three-title theater. Unlock `w_impact > 0` only when **all** of these hold on the same morning pulse (log numbers in this file before flipping the constant): | # | Gate | Threshold | |---|------|-----------| | 1 | Unit coverage | ≥ **25** candidates with ≥1 unit, **and** ≥ **8%** of store (`n ≥ 200`) | | 2 | Kind diversity | ≥ **3** hits each for `price`, `bylaw_id`, **and** `housing_count` | | 3 | Precision hold | Bond-issuance deny + English “unit” deny still green in tests | | 4 | Score proof | Before/after: top-20 Near me Δ logged; no city scar without units displaced solely by a linked unit chip | **First unlock formula** (when gates pass — not before): ``` w_geo = 0.55 w_recency = 0.35 w_tension = 0.00 w_impact = 0.10 impact_weight = 1.0 if any proposed unit on the candidate else 0.0 ``` Binary presence only — never scale by dollar magnitude (that would crown bond noise if a deny ever slips). Flip requires editing `rank_display.W_*` **and** this table in the same commit/log line. ### geo_proximity | `nest_role` / `geo` | Score | |---------------------|-------| | `primary` / `quebec-city` | 1.00 | | `province` / `quebec` | 0.55 | | `linked` | 0.25 | | unknown | 0.10 | ### recency Half-life **36 hours**. ``` recency = 0.5 ** (age_hours / 36) ``` Missing `published_at` → treat as fetched_at if present, else `recency = 0.2`. ### has_contradiction / impact_weight Both **0** in v0 until cluster + enrich land. Do not fake them. `impact_weight` stays **0** even after unit extraction (price / bylaw_id / housing_count) — units are published for the citizen to check; they do not move the score until `w_impact` is turned on by logged ACT. ## Editorial boosts Forbidden unless: 1. Logged in `data/rank_log.jsonl` with reason, who, when 2. Visible in the UI as “manual boost” Silent boosts = propaganda with better geography. ## Change log | Date | Change | Owner | |------|--------|--------| | 2026-09-15 | v0 weights: geo 0.60, recency 0.40; tension/impact off | Marcus (co-founder) | | 2026-09-15 | enrich.py rules-v0 writes proposed geo/topic/impact; w_impact still 0 until we choose to score proposals | Niccolo (co-founder) | | 2026-09-15 | Display nest + geo_proximity use proposed enrich.geo; primary source without city token no longer auto Near me | Bucky (co-founder) | | 2026-09-15 | cluster v0.3: Issues require >=2 source voices; drop single-voice theater; FR/EN anchors | Nietzsche (co-founder) | | 2026-09-15 | enrich topic rules hardened: word-boundary stems; no bare environnement; no ges matching villages; rent no longer matches different | Niccolo (co-founder) | | 2026-09-15 | cluster v0.5: named scars only (Maëlyne/Lévis, Marchand); no coroner/tramway glue; README Critical Path | Bucky (co-founder) | | 2026-09-15 | enrich v0.2: province nest cannot cloak world-fog as quebec; TECHNICAL_PROCESS.md synced to scripts | Marcus (co-founder) | | 2026-09-15 | cluster v0.4: tighter related (no weak glue onto anchored Issues); readable FR-first questions; README Critical Path | Bucky (co-founder) | | 2026-09-15 | cluster v0.6: Marchand scar title-or-mayor-agenda only; drop summary hitchhikers (Capitale-Nationale elections) | Niccolo (co-founder) | | 2026-09-15 | lookout temperance: Province+Linked display top 30 by score; Showing M of N; full set in latest_ranked.json | Bucky (co-founder) | | 2026-09-15 | lookout header clock: Ranked/Enriched/Ingested from JSON + Sources path; invent nothing | Bucky (co-founder) | | 2026-09-15 | serve method files: /VISION.md /ranking.md /sources.yaml /RENT.md from disk; lookout links | Bucky (co-founder) | | 2026-09-15 | Province+Linked display impact-first (proposed topics); Showing line names it; store unchanged | Bucky (co-founder) | | 2026-09-15 | impact-first: recognize enrich stem energy/hydro (Niccolo refuse) | Bucky (co-founder) | | 2026-09-15 | enrich energy/hydro: add diesel|gasoline|gallon (price); essence untouched | Bucky (co-founder) | | 2026-09-15 | lookout: demote /ohdio/ from Near me into Province by score; store untouched | Bucky (co-founder) | | 2026-09-15 | ohdio demote: pin demoted booths into Province displayed 30 (appear moved) | Bucky (co-founder) | | 2026-09-15 | enrich topic transport (tramway|tgv|opus|m[?e]tro|autobus) impact mobility; impact-first allowlist | Bucky (co-founder) | | 2026-09-15 | enrich topic education (éducation|education|écoles|ecoles|schools) impact education; impact-first allowlist | Bucky (co-founder) | | 2026-09-15 | Lookout UI v2.4: micro-nav above fight list; publisher og:image faces via fetch_media.py (never invent/stock); no rank change | Bucky (co-founder) | | 2026-09-15 | Lookout UI v2.3: center default Impact Radar (Province+Linked impact-first); fight yields Issue Stage; Near me rail holds | Bucky (co-founder) | | 2026-09-15 | enrich v0.3: airport/aéroport stem → topic trade + impact trade; never bare private; place-first | Bucky (co-founder) | | 2026-09-15 | Lookout UI v2.2: desktop command field (fights | Issue Stage | Near me rail); micro Prev/Next; mobile keeps deck; archive door | Bucky (co-founder) | | 2026-09-15 | Lookout UI v2.1: full-bleed deck; Issue Stage left/right; Same fight side rail only when content on disk; archive readable measure | Bucky (co-founder) | | 2026-09-15 | Lookout UI v2: Deck + Issue Stage (Issues then Near me); Prev/Next; Open archive for nests; typographic faces only — og:image phase 2; no dwell-time bait | Bucky (co-founder) | | 2026-09-15 | Lookout UI v1: sticky nest nav; Issues above lists; cards; Same fight/Même combat from store only; no dwell-time bait; presentation-only in rank_display.py | Bucky (co-founder) | | 2026-09-15 | lookout: shared booth/brief demote (/ohdio/ + /en-bref/) from Near me; pin visible in Province | Bucky (co-founder) | ## How we know we are wrong - Ranking quietly favors one party line - Linked national fluff outranks Quebec City housing/hydro/law that hits residents - We change weights without logging it here ## Lookout UI v2.5 — Near me confession (2026-09-15) - Kept booth/brief demotion (Ohdio + en-bref leave Near me, pinned in Province). - Hardened the Showing line: names the moved titles and links `#pin-{id}` into Impact Radar. - Click jumps to Province pin (showRadar + scroll). No rank change. No silent vanish. ## Cluster v0.7 — more named scars, not more AI (2026-09-15) - Kept ≥2 distinct `source_id`s. No LLM cluster priest. No bare coroner / bare Duhaime / topic soup. - Added named scars: - `tramway` — title must carry `tramway` **and** (`Marchand`|`Duhaime`) - `airport` — place token + privatisation near place; **Province OK** (do not fake quebec-city) - Order: tramway before marchand so city tram fight is not swallowed by the mayor scar. - Result on this run: **4** fights (airport 4 voices / Province; maelyne-levis 3; tramway 2; marchand 2). Store still 244. - Method string: `rules-cluster-v0.7 named-scars-only; tramway title+(Marchand|Duhaime); airport place+privatisation (Province OK); …` ## Lookout UI v2.6 — questions, not wire (2026-09-15) - Dropped `Plusieurs voix (N) — [outlet lede]` from fight compass and Stage titles. - Scar-locked neutral questions in `cluster_issues.py` (`SCAR_QUESTIONS`): - maelyne-levis → *Que révèle l'enquête du coroner sur la recherche de Maëlyne Lugez à Lévis ?* (colder; no smuggled “manqu”) - tramway → *Faut-il arrêter le tramway de Québec ?* - marchand → *Quelles priorités pour Québec sous Marchand ?* - airport → *Le Canada doit-il privatiser ses grands aéroports ?* - Voice count stays in fight-meta (`N voices`). Sources face on Stage — no crowned answer. - Method: `rules-cluster-v0.7-q`. No new scars. No LLM. Store still 244. ## Lookout UI v2.7 — temper the eye, clear the ontology (2026-09-15) - **Impact Radar** is a **mode** above Fights — not a fifth fight. Fights (4) = four scar-locked questions only. Prev/Next cycle fights only (Radar via mode button). - **Equal card skeleton** on Radar + Near me: fixed face strip height (5.25rem) — publisher `og:image` or empty typography strip of the same height. No tragedy photo may tower over housing. Still their pictures, never stock. No rank change. - Header sub shortened (workshop debris thinned). Method/clock still public. - Store still 244. ## Lookout UI v2.8 — Same fight = real scar only (2026-09-15) - Same fight links **only** when candidates share a named issue on disk (issue_id / scar). Bare topic fallback killed — no false brothers. - Demoted booth/brief stay findable from Near me confession (pins) but sit in a **Moved from Near me** strip after impact-first Province Radar — not crowned as first row. - `cap_slice(..., pin_at="end")`. No new rank weights. No LLM. Store still 244. ## Lookout UI v2.8.1 — every booth/brief out of the crown (2026-09-15) - Province Radar crown = life-hit only (no /ohdio/, no /en-bref/). - All booth/brief sit in the **Moved** strip (findable, not crowned); Near me demotions still listed there. - Linked uses the same split. No rank change. Store still 244. ## Lookout pulse v0.1 — wake without co-founder hands (2026-09-15) - Scheduled thin Critical Path: ingest → normalize → enrich → cluster → rank_display. - Cadence: weekdays 08:00 and 17:00 America/New_York (routine lookout-pulse-v0-1). - Bounded by RENT.md: Inventor wallet; no fetch_media in the thin loop; no unbounded scale. - Falsifiable: Ranked clock advances without a co-founder click. ## Faces v0.2 — map-scoped publisher faces (2026-09-15) - Fetch only what the map shows: Near me + Province/Linked life-hit crown + fight voices (not all 244). - `og:image` / `twitter:image` only. Never stock. Never AI fill. Empty strip when silent. - JdQ hardened (browser UA, Referer, retry) — still **403 Forbidden** from their wall; fail = typography (honest). - Thin `fetch_media` after **morning** pulse only (08:00), not every tick. RENT.md bound. - This run: map_scope=76, with_image=80, fetched=40. Method string in `latest_faces.json`. ## Lookout Arrival v0.1 — Phase 0 + Phase 1 (2026-09-16) PDCA: first paint is one composition (Approaches), not the three-zone command dashboard. ### Plan - Phase 0: `FRICTION.md` names F1–F8 (dashboard, brand, cognitive tax…). - Phase 1: Arrival = Vigie brand · one line · Approaches (≤5 fights, store order) · CTA opens lookout field. - No `w_*` changes. No For You. Approaches are presentation of `latest_issues` only. ### Check - Arrival section before `#command`; field shell hidden until invitation / Approach tap. - Tests: `test_arrival_lookout` — brand, approaches, no “for you”, W_IMPACT=0, W_GEO=0.60. - Unit chips on Approach when scar items carry units. ### Act - Depth (Radar / Stage / Near me) remains behind “Open lookout field”. - Resident walkthroughs (5) still deferred — instrument ships first. ## Approach polish + Since you left v0.1 — Phase 4 (2026-09-16) PDCA: glanceable Approaches + visit delta from **store fingerprints only** (never reorder). ### Plan - Polish: nest labels (Near me/Province/Linked), silence + official-quiet chips, remix edge, `data-issue-id` / `data-fp`. - Since you left: compare `localStorage` prior pulse to embedded `#vigie-pulse`; badge new/changed **in place**. - Kill: For You reorder, ranking weight changes. ### Check - `since_left_delta` unit tests: new / changed / gone / same; order of curr untouched. - HTML: `#since-left`, `#vigie-pulse`, silence chips, Near me label. - W_IMPACT still 0. ### Act - First visit → “First look”; same pulse refresh → “same Approaches”; recluster → factual delta. - Next: Phase 5 ambient pulse only after Arrival+delta holds in resident walks. ## Phase 4 verification — store timestamp (2026-09-17) ### Plan - Strip must name store `clustered_at` (compact UTC). - Identical `clustered_at` short-circuits to same Approaches (fp upgrade must not fake delta). - Fingerprint v2 includes `unit_count`; pulse payload carries it. ### Check - `since_left_visit`: first / same-clock / returning delta with badges. - Live HTML: `formatStoreClock`, `prev.clustered_at` gate, fp `|` count = 4. - Suite: `tests/test_phase4_since_left.py`. ### Act - Phase 4 DoD **MET (instrument)**. Clearing `localStorage` resets to First look (expected). ## Ambient morning pulse v0.1 — Phase 5 (2026-09-16) PDCA: glance channel that cannot diverge from Arrival. ### Plan - Digest = same `build_approaches` + `pulse_payload` as Lookout Arrival. - Inputs: `latest_issues` + `latest_ranked` only. No second ranking. No LLM digest. No For You. - Outputs: `data/pulse/latest_morning.{json,txt}` + `public/morning.html`. - Hook: `rank_display.main()` emits ambient twin after paint; CLI `scripts/ambient_pulse.py` rebuilds from store. - Morning 08:00 Critical Path already runs rank_display → digest ships with the pulse (RENT: zero extra LLM/fetch). ### Check - Identity: digest `issue_ids` + fingerprints + questions == Arrival Approaches (unit tests). - Rank scores cannot reorder Approaches (high-score decoy proven). - Arrival CTA → `/morning.html`; deep-link `/?approach=N` opens that fight. - `W_IMPACT` still 0. ### Act - Stage copies `morning.html` with `index.html`. - Optional paid alert channel later may reuse TXT/JSON — still store twin, never a second brain. ## Phase 5 verification — widget + Stage twin (2026-09-17) ### Plan - Method names Arrival **and** Stage; emit refuses Approaches / `#vigie-pulse` / Stage `issue_id` drift. - Optional widget paste: `latest_morning.widget.txt` — store facts only. ### Check - `digest_matches_arrival_pulse` + `digest_matches_stage_fights` gates. - Live morning JSON approaches == Arrival `#vigie-pulse`. - Suite: `tests/test_phase5_ambient.py`. ### Act - Phase 5 DoD **MET (instrument)**. Alert *delivery* remains RENT v1; widget is the payload. ## Life facets v0.1 — Phase 6 (2026-09-16) PDCA: declared opt-in lenses reorder Approaches only. ### Plan - Publish `FACETS.md` + `scripts/life_facets.py` catalog (renter, transit, energy, civic, health). - Formula: topic_weight + optional unit_weight; sort `(-score, store_index)`; never filter. - Opt-in UI on Arrival; preference device-local (`vigie_facets_v1`). Default = store order. - Morning ambient digest stays store order (shared twin). - Kill: silent personalization, For You, leaking into `score_item` / `w_*`. ### Check - Unit tests: same issue_id set; renter/transit reorder; unit bonus; W_IMPACT=0. - HTML: `#life-facets`, `#vigie-facets`, link `FACETS.md`. - Public rank weights unchanged. ### Act - Facets ship as Arrival lens only; ambient twin remains store order. ## Phase 6 verification — lockstep + gate (2026-09-17) ### Plan - FACETS.md table must match `FACET_CATALOG` byte-for-weight. - Reorder never mutates `data-i`; `score_item` never imports facets; ambient ignores facets. ### Check - `catalog_matches_facets_md`; `w_impact_still_gated`; `test_phase6_facets`. - Live Arrival: `data-unit-kinds`, status “w_impact gated”. ### Act - Phase 6 DoD **MET (instrument)**. ## Beauty without fog v0.1 — design law (2026-09-16) PDCA: make Arrival beautiful without fogging method. ### Plan - Publish `DESIGN.md` Do/Do-not. - Brand as hero; one composition; nest depth on Approaches; expressive type. - Refuse purple/cream cliches, multi-layer radial glow, invented QC photo, score/bias chrome in hero. ### Check - Tests: `test_design_law` — DESIGN.md linked; no radial-gradient; brand clamp > line; no score in Arrival; nest `data-nest`; W_IMPACT=0. ### Act - Full-bleed place image deferred until a real file lands in `public/place/` and is named in DESIGN.md. ## Beauty without fog verification (2026-09-17) ### Plan - Lock DESIGN.md Do/Do-not with live Arrival markers; refuse place invention; de-purple silence chrome. ### Check - `tests/test_beauty_without_fog.py`: composition order, brand>line, font stacks, no radial/cream/purple, empty `public/place/`. - Live `public/index.html` carries `data-design="beauty-without-fog-v0.1"`. ### Act - Beauty-without-fog DoD **MET (instrument)**. ## Phase 0 verification — Freeze truth (2026-09-16) Co-founder audit against canvas DoD. ### Check - Friction map + kill list: **present** (`FRICTION.md`). - Instrument re-walk: F1–F8 remediated on live Arrival; residual F8 closed via `#arrival-ops`. - Resident walks: **0 / 5** — Phase 0 **incomplete**. Protocol + empty R1–R5 table published. No invented residents. ### Act - Next Phase 0 work is scheduling five real Quebec City life walkthroughs; do not check Phase 0 done until resident pain is ranked in `FRICTION.md`. - Tests: `tests/test_phase0_friction.py` locks instrument claims + honest DoD prose. ## Phase 1 verification — Arrival Lookout (2026-09-17) ### Check - No dashboard in first paint: `#field-shell` hidden/aria-hidden; footer deferred; Arrival composition only. - Lighthouse mobile (Edge): Performance **94** · Accessibility **100** · Best Practices **100** · SEO **100** (`data/lighthouse_mobile_p1.json`). - Suite: `test_phase1_arrival` + full unittest OK; `W_IMPACT=0`. ### Act - Phase 1 DoD **MET**. Next resident work remains Phase 0 when live. ## Phase 2 — Approach objects (2026-09-17) PDCA: Approach is the glance object — geo nest + voices + silence + units without opening Stage. ### Plan - Always emit silence count + unit chip (or honest “no units yet”). - Official Quiet names (≤2) on Approach; full silence map stays in Stage. - Ambient digest reuses `approach_meta_chips_html` + silence preview (same object). ### Check - `tests/test_phase2_approaches.py`: glance fields on `build_approaches`; Arrival HTML shows units + Quiet names before Stage silence map; store order unchanged; `W_IMPACT=0`. - Ambient link HTML always carries silence chip. ### Act - Phase 2 DoD **MET (instrument)**. Full institution silence map remains Stage-deep by design. ## Phase 3 — Fight theater (2026-09-17) PDCA: Stage answers who spoke / who didn't in one calm confrontation. ### Plan - Fight arc (Spoke · Did not speak) above voice panels. - Institution seats only; demote faces; drop dense status meta. - Soft-dim side rails when field open — Stage is the confrontation. ### Check - `tests/test_phase3_fight_theater.py`: arc before confrontation; names in first Stage chunk; live HTML carries `fight-theater`. - `W_IMPACT=0`; no bias meter. ### Act - Phase 3 DoD **MET (instrument)**. Rail density polish remains residual Medium. ## Impact units v0.1 — price / bylaw_id / housing_count (2026-09-16) PDCA: falsifiable units on impacts. Citizen can check the raw span. **Not** a reason to turn on `w_impact` yet. Schema (all `status: proposed`): `impacts[].units[]` → `kind` (`price`|`bylaw_id`|`housing_count`), `value`, `unit` (`CAD`|`CAD_per_month`|`CAD_per_L`|`cents_per_kWh`|`percent`|`logements`|`id`), `raw`, `method`, `field`. Precision guards: - Price needs life context (loyer/tarif/facture/essence…) — bond `émission d'obligations` denied. - Housing needs `N logements|habitations|housing units` — English “unity/unit” theater denied. - Bylaw: `projet de loi N`, `règlement N`, `loi N`, `bill N`. ### Check (this run) ``` Ran 78 tests · OK (1 skip) enrich 92.3% / cluster 93.0% AST statements impact units on 3/354 candidates (thin RSS titles — honest) W_IMPACT = 0.00; score(with units) == score(without) rank_components.w_impact = 0 on every ranked row Stage chips: class='chip unit' when raw span present ``` ### Act - **Co-founder decision 2026-09-16: REFUSE enable.** `w_impact` stays **0.00**. - Gates for first unlock are published under Weights above (25 + 8%, all three kinds ≥3, tests green, Near me Δ logged). - Extractors stay on; chips stay on; **no silent weight** between now and a gate-passing ACT. - This run: 3/354 (price×2, bylaw×1, housing×0) — far under gate. ## Voice = institution v0.1 — not RSS feed (2026-09-16) PDCA: sister desks are one house. Fight gate and silence seat use `institution`, not feed `id`. Chancellery: every enabled RSS carries `institution` + `institution_name`. Collapse: CBC Montreal+Politics → `cbc`; Radio-Canada Québec/National/Ottawa → `radio-canada`. 12 feeds → **9** institutions. Schema delta (`silence` v0.2): `scope: enabled_institutions`, `enabled_feed_count`, `silent[].institution_id` + `feed_ids`. Issue: `sources` = institution ids; `source_feeds` = RSS ids that spoke; Stage label = institution name. ### Check (this run) ``` Ran 66 tests · OK (1 skip) cluster 93.0% / enrich 96.4% AST statements maelyne spoke [cbc, le-soleil, radio-canada] silent 6 / 9 marchand spoke [journal-de-quebec, le-soleil, radio-canada] silent 6 / 9 airport spoke [cbc, radio-canada] feeds [cbc-montreal, cbc-politics, radio-canada-quebec] silent 7 / 9 ← was 3 feed-voices; CBC desks now one institution spoke ∩ silent = ∅; spoke+silent = enabled institutions two CBC feeds alone → dropped_single_voice (no fight theater) ``` ### Act - Do not invent institution from URL host heuristics — explicit YAML only (missing → feed id 1:1). - Step 5 next: Impact units (who pays / who waits) with measurable units — still no left/right meter. - Silence remains run-scoped to the named scar cluster. ## Silence map v0.1 — who did not speak on this scar (2026-09-16) PDCA: for each fight, `silence = enabled sources.yaml − voices` (superseded by institution collapse above). Schema (`status: proposed`): `scope: enabled_chancellery` → now `enabled_institutions`, `spoke_count`, `silent_count`, `enabled_count`, `silent[]` with institution identity (+ feed_ids audit). Official absences sort first. Note on disk: not a bias meter. ### Check (this run) - Was: 12 enabled feeds; each fight spoke 3 → silent 9. - Now: 9 institutions; airport silent 7 after CBC collapse (honest). - Official silent first on all scars: ville / gouv / hydro (media remix). - Stage: “Did not speak / N'a pas couvert” strip; fight meta shows `silent N`. - Tests include complement identity: spoke ∩ silent = ∅; spoke+silent = enabled. ### Act - Do not turn silence into a left/right or “reliability” score. - Step 4 (voice=institution) **shipped** — CBC / Radio-Canada sister feeds share one seat. - Silence is run-scoped to the named scar cluster — not “never covered the topic elsewhere.” ## Claims v0.1 — quote + speaker objects (2026-09-16) PDCA probe on 354 titles/summaries → high-precision only. | Pattern | Probe hits | Kept in v0.1 | |---------|------------|--------------| | Guillemets «…» | 50 | yes (+ speaker if affirme/dit nearby) | | Selon X, … | 3 | yes | | Name verb: … | 35 | yes | | Bare title-colon | 58 | **no** (Élections 2026 : … is not a speaker) | | Short labels («Face-à-Face») | — | denied (min 12 chars + deny list) | Schema (all `status: proposed`): `quote`, `speaker` (nullable), `attribution` (`said`|`selon`|`reported`), `method`, `field`. ### Check (this run) - claims on **50/354** candidates (was 0/354 forever-empty). - attributions: said 18 / reported 43 / selon 2; **20** with speaker. - Fights carry claims on voices: maelyne 1, airport 4, marchand 2. - Stage HTML renders claim strips (`class='claim'` × 14). - Tests: 59 OK; enrich AST statements 96.4%. ### Act - Empty `claims[]` remains only when no speech pattern exists — honest silence, not a permanent stub. - Known scar: EN NAME greed can swallow appositives (“Coalition Interjeunes youth coalition”). Tighten later; do not LLM-fill. - No title-colon. No crowned “truth” paraphrase. ## Primary documents v0.1 — Ville / Gouv / Hydro / Soleil (2026-09-16) PDCA probe → enable only proven feeds (ceiling still 12 RSS). | Source | Feed (probed) | kind | Cap | |--------|---------------|------|-----| | Ville de Québec | `Rss/rss.aspx?f=gen` | official | 30 | | Gouv QC | `quebec.ca/fil-de-presse.rss` | official | 30 | | Hydro-Québec | `nouvelles/rss/5/salle-de-nouvelles` | official | **40** (raw 1424 this run) | | Le Soleil | Arc `outboundfeeds/rss/category/actualites` | media (paper of record) | 40 | - `source_kind` flows ingest → normalize → enrich → cluster → Stage. - Enrich: official primary → quebec-city; official province → quebec (world-fog still linked). - Cluster v0.9: official tensions sort first; `official_voice_count` + `media_remix` flag. - Lookout: official chip; Stage warns when scar is media-only. ### Check (this run) - 12/12 feeds OK; 354 candidates; official=61 (ville 11 / gouv 10 / hydro 40). - Hydro `capped=true` (RENT.md held). - Fights still **media_remix=true** (maelyne / airport / marchand) — no official title hit named scars today. Flag is honest, not a painted victory. - Le Soleil joined maelyne + marchand as media paper-of-record voice. ### Act - Do not invent official↔scar glue. Next PDCA: only add named scars when an official text and a media fight share a proper noun on disk. - Borough Ville feeds (cite/sfoy) stay deferred — f=gen is the thin official voice. ## Enrich + cluster v0.4 / v0.8 — word-boundary geo (2026-09-16) - `STRICT_CITY`: place tokens use `(?