clearotron 0.3.1-beta.2 → 0.3.1-beta.4

package/driver/skills/knockout-frame/SKILL.md

@@ -35,7 +35,8 @@ mark) and the assessment turn; nothing you write here searches anything.
    integers only. Scope expansion is NEVER silent.
 4. **Context framing** per mark (MANDATORY): what does this name read as in THIS context — coined,
    compound, common phrase, cultural echo? ("Coop" = chicken coop for a poultry brand, not a
-   cooperative.) The assessor's rating hangs off this framing.
+   cooperative.) It tells the assessor what the name reads as. It does not set the band — a framing
+   never raises or lowers a rating, and a risk belonging to the name is a finding of its own.
 5. **Prior knowledge** from the request/instructions: anything the forwarder said about a mark
    ("client loves this one", known history) — folded in verbatim, never invented.
 6. **Execution order**: common-word / known-problem marks first (they surface the batch's hard calls

package/driver/skills/prelim-search/risk-framework-triage.md

@@ -23,17 +23,17 @@
 
 ### Manageable
 
-- **Triage read.** Minor findings, no strong claims. Flag major competitors/brand owners if relevant. The floor for any common English phrase.
+- **Triage read.** Minor findings, no strong claims. Flag major competitors/brand owners if relevant.
 
 ### Low
 
-- **Triage read.** No issues found. Available only for coined/fanciful terms (calibration rule 3).
+- **Triage read.** No issues found.
 
 ## Calibration rules (they override gut instinct — reviewer-derived)
 
 1. **Class-specific ratings are mandatory at Medium and above** — "Medium (Classes 9, 25, 28, 41)", never a bare band.
 2. **Crowded field = diluted risk, not amplified** — many small non-enforcing players typically caps risk at Medium.
-3. **Common English phrases never rate Low** (Manageable floor + a purple register-pending note); **coined/fanciful terms CAN be Low**. The test: could a reasonable person use this word in everyday speech without reference to the proposed mark? No → coined → Low available. Yes → common phrase → Manageable floor.
+3. **RETIRED — an everyday name with nothing serious found can reach the bottom band** (owner, 2026-09-15). This read "common English phrases never rate Low"; it decided a band from how ordinary the words were, which no evidence supports and which the clearance lane's own instruction forbids in terms — the applicant's own-mark read is advisory and never moves a band. How ordinary or ownable the words are is a note for the reviewing lawyer. Anything that should lift a name is a finding of its own, rated as one.
 4. **Every rating is pending register overlay** — the standing caveat rides every summary.
 5. **Enforcer profiling matters** at Medium+ — portfolio brand? opposition history? institutional filer? "No assertive enforcers" can cap a Medium at "Medium (low)".
 6. **Client's prior use mitigates, never overrides** — rate the full external landscape first; note the mitigation in a purple bullet.

package/driver/stages.mjs

@@ -1363,6 +1363,16 @@ export const STAGES = {
       // The driver STAMPS the instructed-scope section from _driver/instructed-scope.json, so the seat is
       // not asked to quote back values the driver wrote at intake. That retyping was the stage's
       // `frame_scope_missing` loop and it is gone; see matter-frame-record.mjs.
+      // RATIFIED FORMS CARRY THEIR OWN IMPERATIVE, for the reason stated above this block: a field
+      // phrased outside one was written 0 of 9 times against 74 of 74 when imperative-carried. Without
+      // this sentence the typed field is never filled, and the floor under the per-form read in
+      // validators.narrative is green because it is inert — which is the exact defect that floor exists
+      // to end, arriving one layer up.
+      //
+      // ONLY WHEN THE CLIENT RATIFIED MORE THAN ONE. The ordinary matter is one name and must stay
+      // free: a seat asked to list forms for a single-form matter will invent a second one, and the
+      // floor would then demand a read of a form nobody bought.
+      `Send \`ratified_forms\` ONLY if the instruction ratifies more than one form of the same name — the name and a variant of it, both to be searched as one matter. Send each form exactly as the client wrote it. Omit the field entirely for an ordinary single-name matter; do not invent a second form.`,
       `You are NOT asked to quote the request's marks, classes, territories or goods back — the driver stamps that section from its own intake record, so it cannot drift from what was asked.`,
       `Do NOT write or edit any file. There is no path for you to write to: the driver renders the frame from what you send, and nothing you hand-write is read.`,
       `When the tool accepts your call, return ONLY a 2-3 line summary of the frame.`,
@@ -2089,7 +2099,7 @@ export const STAGES = {
       },
       "enumerated block records carried verbatim — record_id, mark_text, classes, status, owner_name, owner_country, application_date, registration_date, expiry_date, jurisdictions, screen_verdict": {
         class: "mechanical:tool-written", tokens: ["named_band_collapsed"],
-        why: "register_enumerate already returns each record batch-screened; carrying it is transcription. `named_band_collapsed` (verify.mjs:1283, findCollapsedBands) is exactly the recall loss that transcription produces — a slice claiming total_hits with zero records reaching the band. [citation unverified]",
+        why: "register_enumerate already returns each record batch-screened; carrying it is transcription. `named_band_collapsed` (verify.mjs:1325, findCollapsedBands) is exactly the recall loss that transcription produces — a slice claiming total_hits with zero records reaching the band. [citation unverified]",
       },
       "per-axis prose digest — the SHORT AUDIT NOTE at register-units/<axis>.md": {
         class: "judgment", tokens: ["too_short"],
@@ -2655,7 +2665,7 @@ export const STAGES = {
       },
       "the escalation decision — which register axes carry a material, unresolved, genuinely closeable gap, and the one-line reason for each": {
         class: "judgment", tokens: [],
-        why: "Whether a documented coverage-limited row, a capability-gap deferral or a fresh concern on a confirmed-clean row warrants spending a re-run. The driver hands the coverage/execution truth in as a computed table (stages.mjs:1466; skepticDeferralExtra) precisely so this is a call over data rather than a re-derivation — but the call itself is nobody else's. [citation unverified]",
+        why: "Whether a documented coverage-limited row, a capability-gap deferral or a fresh concern on a confirmed-clean row warrants spending a re-run. The driver hands the coverage/execution truth in as a computed table (stages.mjs:1476; skepticDeferralExtra) precisely so this is a call over data rather than a re-derivation — but the call itself is nobody else's. [citation unverified]",
       },
       "escalation decisions — one {axis, reason} per axis that must be re-run, sent through record_skeptic": {
         class: "mechanical:code-rendered", tokens: [],
@@ -2860,6 +2870,10 @@ export const STAGES = {
         class: "mechanical:code-rendered", tokens: ["intake_ask_unanswered"],
         why: "the driver already code-builds the report's section from the ask_answers register (assembleReportMd · buildAskAnswersSection, named in stages.mjs) — the narrative section is a second authoring of a register the driver renders anyway",
       },
+      "narrative per-form read — each ratified form reasoned through the framework, or the reads stated alike": {
+        class: "judgment", tokens: ["ratified_form_unread"],
+        why: "a client who ratified two forms is choosing between them; which conflicts move between the forms is the read they bought, and the token is a floor on whether the question was answered at all rather than on the reasoning",
+      },
       "narrative coverage prose — how a coverage gap is stated": {
         class: "judgment", tokens: ["coverage_recommendation", "coverage_gap_unexplained"],
         why: "what the run could and could not clear is the lawyer's honest statement; the token polices only the forbidden 'commission a re-run' form",
@@ -2886,7 +2900,7 @@ export const STAGES = {
       },
       "findings[].meters.use.token / meters.enforcer.token — the two strengths with no quadrant axis": {
         class: "judgment", tokens: ["finding_meter_missing", "finding_meter_token_invalid", "finding_meter_unknown"],
-        why: "Judgment, and S1 does not reach them: quadrant plots goods proximity and mark similarity only (stages.mjs:1733), so nothing derives use (confirmed/not-confirmed/unknown) or enforcer strength. Deriving these from the quadrant would invent a reading the position does not carry. [citation unverified]",
+        why: "Judgment, and S1 does not reach them: quadrant plots goods proximity and mark similarity only (stages.mjs, the quadrant element), so nothing derives use (confirmed/not-confirmed/unknown) or enforcer strength. Deriving these from the quadrant would invent a reading the position does not carry.",
       },
       "findings[].meters.mark_similarity.token / meters.goods_proximity.token — the coarse 3-pip restatement of the quadrant position": {
         class: "mechanical:code-assigned", tokens: ["finding_meter_missing", "finding_meter_token_invalid", "finding_meter_unknown"],
@@ -3042,7 +3056,7 @@ export const STAGES = {
       // therapeutic-goods discipline, no-use ≠ safety (pipelines run 5-10y), practitioner/pharmacist
       // confusion. Conditional read (frameworkFor pattern); doctrine TABLES live elsewhere, untouched.
       reads(synthesisSkillReads({ profile, job })),
-      `Produce the joint synthesis narrative (dominant-element spine, cross-pollination Option-D cap N=10, actual-use check, owner workup on high-risk findings, rate each conflict net of merits defences UNDER THE FRAMEWORK YOU JUST READ — state the likelihood of confusion in plain words, then give the band those words require under the framework's own definitions (the prose and the band are one judgment and may not disagree), coverage honesty).`,
+      `Produce the joint synthesis narrative (dominant-element spine, cross-pollination Option-D cap N=10, actual-use check, owner workup on high-risk findings, rate each conflict net of merits defences UNDER THE FRAMEWORK YOU JUST READ — state the likelihood of confusion in plain words, then give the band those words require under the framework's own definitions (the prose and the band are one judgment and may not disagree), coverage honesty). WHERE THE MATTER RATIFIED MORE THAN ONE FORM OF THE NAME — a second spelling the instruction told you to read as a variant of the same candidate — REASON EACH FORM THROUGH THE FRAMEWORK ON ITS OWN, then say in the narrative which conflicts move between them and why, naming the form each read belongs to. A conflict that turns on one form exactly is not the same exposure for the other, and a reader choosing between the forms cannot act on a single undifferentiated read. Where every conflict lands the same on both, say so in one line — "the reads are the same for both forms" — as an asserted finding rather than a silence: a reader must be able to tell a matter where the forms do not differ from one where nobody looked. This changes the NARRATIVE and nothing about the report's shape: one overall band as today, and never a second one.`,
       // — the prose rung. Empty on a one-country run, and `lines` drops it.
       proseRungDirective(depth, framework?.bands ?? null),
       // doc 50 — THE RATING AUTHORITY: the framework read above (the customer's own, or the Generic default)
@@ -3702,7 +3716,7 @@ export const STAGES = {
       `Hand the shell back by calling the \`record_report_overview\` tool. There is no file to write and no front-matter to type: the driver renders the shell — every front-matter key it already holds, the \`# Actions\` section and the optional \`# Methodology\` note — from the values you send, and assembles the cards around it. Do NOT author any \`## <owner> — <MARK>\` card and do not think about a \`# Marks\` heading; each card is rendered separately. Risk scoring is unchanged.`,
       // TRANSFORM, NOT FREE-GENERATION (Fix B1 — carried from main's report-synthesis): the shell RESHAPES the
       // settled synthesis, it never re-derives. Applied here to overall_caption + # Actions.
-      `TRANSFORM, NOT FREE-GENERATION: overall_caption and # Actions draw ONLY on the settled synthesis narrative ${P.narrative} (and findings.json ${P.findings}) — introduce NO finding, owner, or claim that is not already there, NEVER re-compute a band, and never let a claim outrun its evidence (a verified-from-record fact stays fact; an inferred-from-signal claim reads AS an inference — "on the available signals…" — never flattened to bald fact).`,
+      `TRANSFORM, NOT FREE-GENERATION: overall_caption and # Actions draw ONLY on the settled synthesis narrative ${P.narrative} (and findings.json ${P.findings}) — introduce NO finding, owner, or claim that is not already there, NEVER re-compute a band, and never let a claim outrun its evidence (a verified-from-record fact stays fact; an inferred-from-signal claim reads AS an inference — "on the available signals…" — never flattened to bald fact). WHERE THE SETTLED NARRATIVE REASONED MORE THAN ONE RATIFIED FORM OF THE NAME, CARRY ITS PER-FORM READ THROUGH — the overall_caption says which form the decisive conflicts turn on, or that the reads are the same for both, in the narrative's own words. Do not flatten two reads into one undifferentiated statement: that loses the distinction the client ratified both forms to get. This is a transform of what the narrative already settled, not a second rating, and the band is unchanged — one overall band as today.`,
       // — the D2 case-law line is RETIRED HERE (it stays on `report-card`, which is where
       // enforcement-history and opposition-likelihood are actually written). Two reasons, and the second
       // is the general one:

package/driver/suite-census.json

@@ -471,6 +471,12 @@
           "skips": 1,
           "todos": 0
         },
+        "a-name-is-never-rated-above-its-worst-card.test.mjs": {
+          "tests": 4,
+          "asserts": 15,
+          "skips": 0,
+          "todos": 0
+        },
         "a-name-the-wizard-returns-is-a-setup-name.test.mjs": {
           "tests": 5,
           "asserts": 9,
@@ -843,6 +849,12 @@
           "skips": 0,
           "todos": 0
         },
+        "a-two-form-matter-owes-a-read-per-form.test.mjs": {
+          "tests": 8,
+          "asserts": 25,
+          "skips": 0,
+          "todos": 0
+        },
         "a-unit-configuration-cannot-reach-is-resolved-at-install.test.mjs": {
           "tests": 5,
           "asserts": 14,
@@ -1378,8 +1390,8 @@
           "todos": 0
         },
         "common-law-receipts.test.mjs": {
-          "tests": 36,
-          "asserts": 129,
+          "tests": 40,
+          "asserts": 141,
           "skips": 0,
           "todos": 0
         },
@@ -1439,7 +1451,7 @@
         },
         "connect-clients-are-data.test.mjs": {
           "tests": 17,
-          "asserts": 87,
+          "asserts": 91,
           "skips": 0,
           "todos": 0
         },
@@ -1505,7 +1517,7 @@
         },
         "contract-audit.test.mjs": {
           "tests": 54,
-          "asserts": 167,
+          "asserts": 168,
           "skips": 0,
           "todos": 0
         },
@@ -1725,6 +1737,12 @@
           "skips": 0,
           "todos": 0
         },
+        "deprecating-below-a-stable-touches-nothing-above-it.test.mjs": {
+          "tests": 4,
+          "asserts": 18,
+          "skips": 0,
+          "todos": 0
+        },
         "depth-ladder-loop-bounds.test.mjs": {
           "tests": 6,
           "asserts": 10,
@@ -2205,6 +2223,12 @@
           "skips": 0,
           "todos": 0
         },
+        "every-json-response-says-how-it-may-be-cached.test.mjs": {
+          "tests": 4,
+          "asserts": 14,
+          "skips": 0,
+          "todos": 0
+        },
         "every-lane-declares-its-transport-position.test.mjs": {
           "tests": 7,
           "asserts": 12,
@@ -2439,9 +2463,15 @@
           "skips": 0,
           "todos": 0
         },
+        "giving-access-to-everything-is-a-switch-not-a-point.test.mjs": {
+          "tests": 1,
+          "asserts": 4,
+          "skips": 0,
+          "todos": 0
+        },
         "grant-command.test.mjs": {
           "tests": 19,
-          "asserts": 75,
+          "asserts": 76,
           "skips": 0,
           "todos": 0
         },
@@ -2817,6 +2847,12 @@
           "skips": 0,
           "todos": 0
         },
+        "no-two-dictated-queries-fold-to-one-key.test.mjs": {
+          "tests": 5,
+          "asserts": 10,
+          "skips": 0,
+          "todos": 0
+        },
         "no-vendor-provenance-in-the-public-cut.test.mjs": {
           "tests": 4,
           "asserts": 11,
@@ -3142,8 +3178,8 @@
           "todos": 0
         },
         "portal-service.test.mjs": {
-          "tests": 124,
-          "asserts": 673,
+          "tests": 127,
+          "asserts": 683,
           "skips": 1,
           "todos": 0
         },
@@ -3665,7 +3701,7 @@
         },
         "render.test.mjs": {
           "tests": 154,
-          "asserts": 630,
+          "asserts": 632,
           "skips": 0,
           "todos": 0
         },
@@ -4377,6 +4413,12 @@
           "skips": 0,
           "todos": 0
         },
+        "taking-access-away-is-the-editor-that-gave-it.test.mjs": {
+          "tests": 9,
+          "asserts": 30,
+          "skips": 0,
+          "todos": 0
+        },
         "teardown-leaves-nothing-claimable.test.mjs": {
           "tests": 4,
           "asserts": 19,
@@ -4527,6 +4569,12 @@
           "skips": 0,
           "todos": 0
         },
+        "the-containment-rule-does-not-depend-on-the-string-tmp.test.mjs": {
+          "tests": 2,
+          "asserts": 3,
+          "skips": 0,
+          "todos": 0
+        },
         "the-corrections-section-is-an-allowlist.test.mjs": {
           "tests": 6,
           "asserts": 18,
@@ -4815,6 +4863,12 @@
           "skips": 0,
           "todos": 0
         },
+        "the-people-page-can-change-and-remove-a-person.test.mjs": {
+          "tests": 16,
+          "asserts": 80,
+          "skips": 0,
+          "todos": 0
+        },
         "the-person-who-installs-is-a-person-not-a-domain.test.mjs": {
           "tests": 18,
           "asserts": 86,
@@ -5151,6 +5205,12 @@
           "skips": 0,
           "todos": 0
         },
+        "two-ratified-forms-get-reasoned-apart.test.mjs": {
+          "tests": 7,
+          "asserts": 17,
+          "skips": 0,
+          "todos": 0
+        },
         "two-word-annotated-variant-is-refused.test.mjs": {
           "tests": 6,
           "asserts": 21,
@@ -5450,6 +5510,18 @@
           "skips": 0,
           "todos": 0
         },
+        "a-removed-person-loses-their-assistant-too.test.mjs": {
+          "tests": 4,
+          "asserts": 12,
+          "skips": 0,
+          "todos": 0
+        },
+        "a-run-list-names-its-client.test.mjs": {
+          "tests": 4,
+          "asserts": 29,
+          "skips": 0,
+          "todos": 0
+        },
         "account-audit-chain.test.mjs": {
           "tests": 10,
           "asserts": 46,
@@ -5720,6 +5792,12 @@
           "skips": 0,
           "todos": 0
         },
+        "the-portal-can-tell-who-has-connected-an-assistant.test.mjs": {
+          "tests": 10,
+          "asserts": 28,
+          "skips": 0,
+          "todos": 0
+        },
         "token-door-boot.test.mjs": {
           "tests": 7,
           "asserts": 8,
@@ -5738,14 +5816,14 @@
       "ext": "ts",
       "perFile": {
         "api.test.ts": {
-          "tests": 44,
-          "asserts": 176,
+          "tests": 45,
+          "asserts": 183,
           "skips": 0,
           "todos": 0
         },
         "askAi.test.ts": {
-          "tests": 5,
-          "asserts": 12,
+          "tests": 11,
+          "asserts": 34,
           "skips": 0,
           "todos": 0
         },
@@ -5815,6 +5893,12 @@
           "skips": 0,
           "todos": 0
         },
+        "decodePersonCarriesEveryField.test.ts": {
+          "tests": 4,
+          "asserts": 17,
+          "skips": 0,
+          "todos": 0
+        },
         "effortModelParity.test.ts": {
           "tests": 7,
           "asserts": 31,
@@ -5864,8 +5948,8 @@
           "todos": 0
         },
         "home.test.ts": {
-          "tests": 51,
-          "asserts": 169,
+          "tests": 56,
+          "asserts": 196,
           "skips": 0,
           "todos": 0
         },
@@ -5941,6 +6025,12 @@
           "skips": 0,
           "todos": 0
         },
+        "oneHeaderPerPage.test.ts": {
+          "tests": 4,
+          "asserts": 15,
+          "skips": 0,
+          "todos": 0
+        },
         "ownerNames.test.ts": {
           "tests": 9,
           "asserts": 23,
@@ -5977,6 +6067,12 @@
           "skips": 0,
           "todos": 0
         },
+        "proseIsStrippedOnce.test.ts": {
+          "tests": 3,
+          "asserts": 12,
+          "skips": 0,
+          "todos": 0
+        },
         "readiness.test.ts": {
           "tests": 6,
           "asserts": 19,
@@ -6062,8 +6158,8 @@
           "todos": 0
         },
         "useYourAI.test.ts": {
-          "tests": 13,
-          "asserts": 41,
+          "tests": 12,
+          "asserts": 38,
           "skips": 0,
           "todos": 0
         },

package/driver/verify-knockout.mjs

@@ -668,12 +668,70 @@ export function validateMergedFindings(runDir, merged, plan) {
     // would put a field in the artifact the stage never emitted.
     try { const ranked = validateKnockoutFindings(m.findings, { manifest }); if (Array.isArray(m.findings)) m.findings = ranked; }
     catch (e) { failures.push(`mark "${m.name}": ${e.message}`); }
+    const above = markRatedAboveItsCards(manifest, m);
+    if (above) failures.push(above);
   }
   const receipts = knockoutReceipts(runDir, marks);
   failures.push(...receipts.failures);
   return { ok: failures.length === 0, failures, receipts: receipts.checked };
 }
 
+// ── A NAME IS NEVER RATED ABOVE THE WORST CARD ON ITS PAGE ───────────────────────────────────────────
+//
+// Three names went out rated a band above every conflict card on their own pages. Twelve cards across
+// the three, all at the client's bottom band; three names a band above them; nothing on any page
+// explaining the gap. The reviewing lawyer read the cards as right and the names as wrong.
+//
+// Nothing tied the two together. The name's band is the model's own `rating` and no check compared it
+// with the bands underneath it — a gap the full clearance does not have, where the headline is computed
+// from the worst live finding. The doctrine that produced the lift (the everyday-words floor) is retired
+// in this change; this is the floor under it, so the same shape cannot arrive by another route.
+//
+// IT REFUSES AND ASKS AGAIN. It never lowers the band itself, and that restraint is the point: a page
+// whose badge a script moved while its basis line still argues for the old band gives a reader a
+// Manageable chip over a Medium argument, which is worse than the defect. The corrective ladder this
+// failure list feeds re-dispatches the mark with the reason, and the model reasons the band again.
+//
+// A CARD IS A RATED ITEM ON THIS NAME'S PAGE — a `findings[]` record, or a `registerReads[]` row that
+// carries a band. A register row's band is optional by design (a read with no band prints no chip), so
+// a row without one is not a card and does not count; including it would invent a floor out of silence.
+// There is no withdrawn state on a knockout finding — that is the clearance lane's concept and it does
+// not exist here, so nothing is filtered for it. Whoever brings it over adds the filter with it.
+//
+// RATING BELOW THE WORST CARD STAYS ALLOWED. It has happened, and the one that was checked was reasoned
+// — the worst card's holder touched only the edge of the request. This refuses one direction only.
+//
+// INDEX 0 IS THE WORST BAND, as `worstBand` below says in its own comment, so "above" is a SMALLER
+// index. Written the other way round this check would pass on the exact production run it was written
+// for and fire on correct ones; the arms assert the direction against a real four-band ladder rather
+// than trusting this sentence.
+const cardBands = (m) => [
+  ...(Array.isArray(m?.findings) ? m.findings : []).map((f) => f?.band),
+  ...(Array.isArray(m?.registerReads) ? m.registerReads : []).map((r) => r?.band),
+].filter((b) => b != null && String(b).trim());
+
+export function markRatedAboveItsCards(framework, m) {
+  const ladder = ladderOf(framework);
+  if (!ladder.length) return null;                       // no frozen ladder — the band checks above own that
+  const mine = bandIdx(ladder, m?.rating);
+  if (mine < 0) return null;                             // an unknown band is knockout_band_unknown's to name
+  const idxs = cardBands(m).map((b) => bandIdx(ladder, b)).filter((i) => i >= 0);
+  // NO CARDS, NOTHING TO BE ABOVE — and the code stops here deliberately.
+  //
+  // The doctrine says a name with nothing found against it takes the bottom band, and that belongs in
+  // the reasoning where it is written (calibration rule 8). Enforcing it HERE would be a second rule in
+  // code: a floor that forces a band down, which is what the ruling forbids in terms — "do not replace
+  // the removed rule with another floor, cap or word list. The two rulings are the whole requirement."
+  //
+  // The first draft of this function did enforce it, and the e2e batch is what said so: a mark with no
+  // cards, correctly rated Manageable on a five-band ladder, was refused because the bottom of that
+  // ladder is Low. That is the retired rule's own shape — a band decided by something other than the
+  // evidence on the page — arriving from the opposite direction.
+  if (!idxs.length) return null;
+  const worst = Math.min(...idxs);
+  return mine < worst ? `knockout_rating_above_cards:${m?.name}: rated "${m?.rating}", above every card on its page — the worst is "${ladder[worst]}". Rate the name from its cards, or write the risk that lifts it as a finding of its own and rate it there. Rating BELOW the worst card is allowed and wants a sentence saying which card does not carry the name.` : null;
+}
+
 // worst band across the batch, in the frozen ladder's own order (index 0 = worst)
 export function worstBand(framework, marks) {
   const ladder = ladderOf(framework);

package/driver/verify.mjs

@@ -12,9 +12,9 @@ import { dirname, join, basename } from "node:path";
 import { driverDir } from "../shared/driver-dir.mjs";   //
 import { findReceiptViolations, findGridLedgerViolations, findPlatformIdentityViolations, parsePrRiskQueries, MEANING_SEAT } from "./common-law-receipts.mjs";
 // Conversion 2 — the discriminator the two rulings above key on. PURE-ish: one existsSync-shaped read.
-import { matterFrameWasRecorded } from "./matter-frame-record.mjs";
+import { matterFrameWasRecorded, frameRatifiedForms } from "./matter-frame-record.mjs";
 import { findConnotationViolations, parsePrRiskResults, MEANING_ANGLES_RE,
-  parseDispositionForm, CONNOTATION_UNRULED_REASONS } from "./connotation-search.mjs";
+  parseDispositionForm, CONNOTATION_UNRULED_REASONS, queryKey } from "./connotation-search.mjs";
 import { formSidecarName, formSidecarPath } from "./disposition-union.mjs";
 // B — the transport's own four failure states. The audit reads the run's records; this file locates them.
 import { auditDispositionCalls, CALL_FAILURE_REASONS } from "./disposition-call-audit.mjs";
@@ -354,12 +354,54 @@ function commonLawMeaningSeat(p, c) {
   try { ledgerRaw = readFileSync(join(dirname(p), `common-law-grid.half-${MEANING_SEAT}.json`), "utf8"); } catch { /* missing → fail-closed below */ }
   if (ledgerRaw == null) return fail(`grid_ledger_missing:common-law-grid.half-${MEANING_SEAT}.json absent while _driver/grid-spec.half-${MEANING_SEAT}.json dictates the meaning sweep`);
   const dictated = Array.isArray(spec?.connotation?.queries) ? spec.connotation.queries : [];
-  let recordedQ;
-  try { recordedQ = new Set(parsePrRiskResults(ledgerRaw).map((e) => String(e?.query ?? "").trim())); }
+  // COMPARED ON A KEY, NOT ON THE RAW TEXT — `queryKey` in connotation-search.mjs, beside the parser,
+  // so there is one author of what "the same query" means. The dictated spelling is the driver's; the
+  // recorded one comes back through a provider that returns typographic punctuation. Measured on a
+  // production clearance, 2026-09-16: one query of sixty-one differed by a single right single quotation
+  // mark where the driver wrote an apostrophe, and the grid had run and recorded every one of them.
+  //
+  // THE ORIGINAL TEXT IS WHAT THE REFUSAL QUOTES, below. Somebody chasing a query that really is missing
+  // needs the spelling the driver asked for, not a flattened key.
+  let recordedRaw;
+  try { recordedRaw = parsePrRiskResults(ledgerRaw).map((e) => String(e?.query ?? "")); }
   catch (e) { return fail(`grid_ledger_unparseable:${String(e.message).slice(0, 80)}`); }
-  const dropped = dictated.filter((q) => !recordedQ.has(String(q).trim()));
-  if (dropped.length)
-    return fail(`connotation_query_unrecorded:${dropped.slice(0, 3).map((q) => abbrev(q, 40)).join(",")}${dropped.length > 3 ? ` (+${dropped.length - 3} more)` : ""}`);
+  const recordedQ = new Set(recordedRaw.map(queryKey));
+  const dropped = dictated.filter((q) => !recordedQ.has(queryKey(q)));
+  if (dropped.length) {
+    // ── THE REFUSAL SAYS WHICH OF TWO FAULTS THIS IS, because they need opposite remedies ──────────
+    //
+    // ABSENT: no recorded query resembles it, so the search was not run and the seat must run it.
+    // UNMATCHED: something close IS recorded, so the search ran and the two spellings disagree beyond
+    // what the key folds — a re-ordering, a translation, a truncation, a query the provider chose for
+    // itself. Telling the seat to "re-run the missing query" in that case asks for the one thing that
+    // cannot help, and that is what turned one attempt into four on a production clearance.
+    //
+    // NO THRESHOLD DECIDES ANYTHING (owner's ruling). The nearest recorded query is shown so a person or
+    // a seat can SEE the difference in one attempt; it never makes the gate pass. A similarity score
+    // that could pass this gate would be a score that can hide a skipped query, which is what the gate
+    // is for.
+    const nearest = (q) => {
+      const words = new Set(queryKey(q).split(" ").filter(Boolean));
+      if (!words.size) return null;
+      let best = null, bestScore = 0;
+      for (const r of recordedRaw) {
+        const rw = queryKey(r).split(" ").filter(Boolean);
+        if (!rw.length) continue;
+        const shared = rw.filter((w) => words.has(w)).length / Math.max(words.size, rw.length);
+        if (shared > bestScore) { bestScore = shared; best = r; }
+      }
+      // A single shared word is a coincidence in a sweep about one mark, where the mark's own name is in
+      // every query. Below half the words in common, say nothing rather than point at a red herring.
+      return bestScore >= 0.5 ? best : null;
+    };
+    const parts = dropped.slice(0, 3).map((q) => {
+      const n = nearest(q);
+      return n
+        ? `${abbrev(q, 40)} [unmatched; nearest recorded: ${abbrev(n, 40)}]`
+        : `${abbrev(q, 40)} [absent from the ledger]`;
+    });
+    return fail(`connotation_query_unrecorded:${parts.join(",")}${dropped.length > 3 ? ` (+${dropped.length - 3} more)` : ""}`);
+  }
   if (spec?.connotation?.disposition_required === true) {
     const recorded = parsePrRiskResults(ledgerRaw);
     const form = dispositionForm(dirname(p), spec?.connotation?.dispositions_path);
@@ -2249,6 +2291,40 @@ export const validators = {
           return fail(`intake_ask_unanswered:${asks.length - answered}:of:${asks.length}`);
       }
     } catch { /* no sidecar / unreadable — legacy run, gate off */ }
+
+    // ── A TWO-FORM MATTER OWES A READ PER FORM ────────────────────────────────────────────────────
+    //
+    // Synthesis is told to reason each ratified form through the framework and to say which conflicts
+    // move between them — or to state in one line that the reads are the same for both. Until this,
+    // nothing enforced it: the arms on that dictation match the wording, so they red when the prose
+    // moves and cannot red when a model ignores the ask. This is the deterministic floor under it.
+    //
+    // GATED ON THE FRAME HAVING FROZEN MORE THAN ONE FORM, which is why it can exist at all. A run with
+    // no frame, an archived or replayed run whose accepted call predates the field, and every ordinary
+    // single-form run reach `forms.length > 1` as false and leave here untouched — so every one of them
+    // re-verifies byte-identically. Same shape as the intake-asks gate above, and the same reason.
+    //
+    // WHAT IT ASKS FOR IS THE CHEAPEST HONEST THING: each form NAMED in the narrative, or the
+    // stated-alike line. It does not judge the reasoning — a floor that tried would be a worse version
+    // of the refutation pass, which stays the semantic backstop. It judges that the question was
+    // answered at all, which is exactly what was missing: a narrative that never mentions the second
+    // form is indistinguishable from one that examined it and found nothing, and a client who ratified
+    // two forms is choosing between them.
+    try {
+      const forms = frameRatifiedForms(dirname(p));
+      if (forms.length > 1) {
+        const text = String(c ?? "");
+        // The alternative the dictation offers, matched on its own terms rather than by a fixed
+        // sentence: a narrative may state the reads are alike instead of splitting them.
+        const alike = /\bthe reads are the same for both forms\b/i.test(text);
+        if (!alike) {
+          const unnamed = forms.filter((f) => !text.toLowerCase().includes(f.toLowerCase()));
+          if (unnamed.length)
+            return fail(`ratified_form_unread:${unnamed.length}:of:${forms.length}:${abbrev(unnamed[0], 60)}`);
+        }
+      }
+    } catch { /* no frame / unreadable — gate off, exactly as above */ }
+
     // V4-4 item 6 — narrative coverage contract (code-checked, conditional on the closure receipt):
     // a marketplace coverage gap may be stated ONLY as an attempted-and-unreachable fact, never as
     // "commission a re-run" work for a human — the driver already closed or proved-unclosable every

package/mcp-server/CHANGELOG.md

@@ -1,5 +1,16 @@
 # trademark-artifacts-mcp
 
+## 0.3.1-beta.4
+
+No changes in this release.
+
+## 0.3.1-beta.3
+
+### Patch Changes
+
+- 56a760f: Fixed: Removing somebody's access now ends their assistant's connection on its next request. Until now an assistant that had already connected kept the access it started with for up to half an hour.
+- c05e0a7: Fixed: Ask your assistant for a client's recent searches by the client's name and it finds them. Every search in the list now says which client it was for, and which project. A search whose own record cannot be read says so, instead of appearing to belong to nobody.
+
 ## 0.3.1-beta.2
 
 ### Patch Changes