{
 "name": "Grace – all test series",
 "date": "2026-10-03",
 "license": "Published by Parsimic for anyone to check. Please cite the source.",
 "selfMeasured": true,
 "countingRules": {
  "run": "One agent (Claude alone, Claude with Grace, Grace alone, Claude finishing Grace's result, or a third-party tool) working on one task once, with token and cost readings.",
  "validRun": "A run that finished and was measured as planned. Crashed, blocked, cut-off or misconfigured runs are listed with their reason and never count toward any result or total.",
  "calibrationMeasurement": "A single model call without a real task, used to calibrate the instruments (limit reading, base load, work lists). Counted separately from runs.",
  "analysis": "Re-judging or re-checking existing end states. No new runs; listed as a series only.",
  "distinctTasks": "Each task counted once even if it appears in several series; when unsure whether two tasks are the same, counted once.",
  "tasksPerSeries": "The tasks of a series are the distinct tasks with at least one valid run in it (the same rule as for the total). Series without runs of their own (analyses, calibration measurements, notes, series not run) have none (null).",
  "versions": "Results from different Grace versions are never averaged. Series with released versions of Grace: R83 and R84 (release of 28 Sep 2026), R87 (0.3.6), R88 and R89 (0.3.7), R90 (0.3.8, release candidate with the same behaviour as the release) and R91 (0.3.8); all other Grace series are development builds.",
  "baseline": "Unless a series says otherwise, \"Claude alone\" means Claude Code with the same model on the same task; in the release series with the two token-saving plugins caveman and context-mode."
 },
 "summary": {
  "firstDay": "2026-09-13",
  "lastDay": "2026-10-03",
  "seriesListed": 91,
  "seriesNotRun": 3,
  "seriesNoteOnly": 1,
  "seriesWithRuns": 75,
  "seriesCalibration": 4,
  "seriesAnalysisOnly": 7,
  "runsStarted": 2142,
  "runsValid": 2013,
  "runsInvalid": 129,
  "runsValidWithGrace": 1401,
  "runsValidClaudeAlone": 579,
  "runsValidOther": 33,
  "graceRunsValidByVersion": {
   "alt": 363,
   "nachbesserung": 569,
   "uebernahme": 165,
   "zwischenstand": 82,
   "umbau": 49,
   "release": 46,
   "v036": 30,
   "v037": 37,
   "v038": 60
  },
  "calibrationMeasurementsStarted": 112,
  "calibrationMeasurementsValid": 107,
  "allMeasurementsValid": 2120,
  "distinctTasks": 273,
  "distinctTasksBreakdown": {
   "newBuildCatalog": 156,
   "changesOldArchitecture": 86,
   "changeTasksExistingCode": 15,
   "wholeChangeRequests": 5,
   "learningSystemVariants": 8,
   "firstSingleTasks": 2,
   "smallTask": 1
  },
  "distinctTasksWithValidGraceVsClaudePair": 173,
  "release": {
   "runsStarted": 276,
   "runsValid": 271,
   "runsValidWithGrace": 173,
   "runsValidClaudeAlone": 98,
   "tasks": 71,
   "validPairs": 96
  },
  "headlineSeries": [
   "R82",
   "R83",
   "R84",
   "R89",
   "R90",
   "R91"
  ],
  "headlineNote": {
   "de": "Die Ergebniszahlen der Messseite stammen aus R90 und R91 (Grace 0.3.8, 02.–03.10.2026: 5-Stunden-Limit und Qualität über 60 Aufgaben; die Zeit nur aus R91, gleichzeitig gemessen); als frühere Messungen stehen dort R89 (Grace 0.3.7), R84 (Fülltest), R83 (Großtest) und R82 (API-Zwischenstand). Die Gesamtzahl der Messläufe belegt die Ersparnis nicht; sie zeigt, wie viel insgesamt gemessen wurde – auch alles, was nicht geklappt hat.",
   "en": "The result figures on the measurements page come from R90 and R91 (Grace 0.3.8, 2–3 Oct 2026: 5-hour limit and quality over 60 tasks; time only from R91, measured at the same time); R89 (Grace 0.3.7), R84 (fill test), R83 (large test) and R82 (API interim build) are shown there as earlier measurements. The total number of runs does not prove the saving; it shows how much was measured in total – including everything that did not work."
  }
 },
 "kinds": {
  "api_kosten": {
   "de": "API-Kosten zu Listenpreisen",
   "en": "API cost at list prices"
  },
  "abo_limit": {
   "de": "Abo-Limit (5 Stunden / Woche)",
   "en": "Subscription limit (5-hour / weekly)"
  },
  "qualitaet": {
   "de": "Qualität",
   "en": "Quality"
  },
  "tempo": {
   "de": "Tempo",
   "en": "Speed"
  },
  "streuung": {
   "de": "Streuung / Wiederholbarkeit",
   "en": "Spread / repeatability"
  },
  "lernsystem": {
   "de": "Lernsystem",
   "en": "Learning system"
  },
  "bezug": {
   "de": "Claude allein in anderen Einstellungen",
   "en": "Claude alone in other settings"
  },
  "konkurrenz": {
   "de": "Andere Werkzeuge (anonym)",
   "en": "Other tools (anonymous)"
  },
  "eichung": {
   "de": "Eichung der Messmittel",
   "en": "Calibrating the instruments"
  },
  "aufbau": {
   "de": "Aufbau- und Funktionsproben",
   "en": "Harness and function checks"
  }
 },
 "versions": {
  "keine": {
   "de": "ohne Grace",
   "en": "no Grace"
  },
  "alt": {
   "de": "Entwicklung: alte Architektur (13.–25.09.)",
   "en": "development: old architecture (13–25 Sep)"
  },
  "nachbesserung": {
   "de": "Entwicklung: Claude bessert nach (25.09.)",
   "en": "development: Claude finishes (25 Sep)"
  },
  "uebernahme": {
   "de": "Entwicklung: Übernahme / Liste / Strang (25.–26.09.)",
   "en": "development: hand-over / list / chain (25–26 Sep)"
  },
  "zwischenstand": {
   "de": "Entwicklung: Zwischenstände v1–v5.1 (26.–27.09.)",
   "en": "development: interim versions v1–v5.1 (26–27 Sep)"
  },
  "umbau": {
   "de": "Entwicklung: Umbau für Abo, API, Tempo (28.09.)",
   "en": "development: rebuild for subscription, API, speed (28 Sep)"
  },
  "release": {
   "de": "Release-Fassung (28.09.2026)",
   "en": "release version (28 Sep 2026)"
  },
  "v036": {
   "de": "Grace 0.3.6 (01.10.2026)",
   "en": "Grace 0.3.6 (1 Oct 2026)"
  },
  "v037": {
   "de": "Grace 0.3.7 (01.10.2026)",
   "en": "Grace 0.3.7 (1 Oct 2026)"
  },
  "v038": {
   "de": "Grace 0.3.8 (02.–03.10.2026; R90 als Freigabekandidat)",
   "en": "Grace 0.3.8 (2–3 Oct 2026; R90 as release candidate)"
  }
 },
 "blocks": {
  "A": {
   "de": "Erste Versuche, alte Architektur",
   "en": "First attempts, old architecture"
  },
  "B": {
   "de": "Große Reihen, alte Architektur",
   "en": "Large series, old architecture"
  },
  "C": {
   "de": "Claude bessert Graces Ergebnis nach",
   "en": "Claude finishing Grace's result"
  },
  "L": {
   "de": "Lernsystem",
   "en": "Learning system"
  },
  "E": {
   "de": "Neubau: Übernahme, Liste, Strang",
   "en": "New builds: hand-over, checklist, chain"
  },
  "F": {
   "de": "Zwischenstände v1–v5.1",
   "en": "Interim versions v1–v5.1"
  },
  "H": {
   "de": "Umbau für Abo, API und Tempo",
   "en": "Rebuild for subscription, API and speed"
  },
  "R": {
   "de": "Release-Fassung",
   "en": "Release version"
  },
  "G": {
   "de": "Vergleiche ohne Grace",
   "en": "Comparisons without Grace"
  },
  "D": {
   "de": "Messlatten und Eichungen",
   "en": "Yardsticks and calibration"
  }
 },
 "statuses": {
  "valid": {
   "de": "gültig",
   "en": "valid"
  },
  "partly_valid": {
   "de": "teilweise gültig",
   "en": "partly valid"
  },
  "invalid": {
   "de": "ungültig",
   "en": "invalid"
  },
  "incomplete": {
   "de": "abgebrochen / unvollständig",
   "en": "stopped / incomplete"
  },
  "not_run": {
   "de": "nicht durchgeführt",
   "en": "not run"
  },
  "calibration": {
   "de": "Eichung / Auswertung",
   "en": "calibration / analysis"
  },
  "not_counted": {
   "de": "nur Notiz, nicht gezählt",
   "en": "note only, not counted"
  }
 },
 "units": {
  "lauf": {
   "de": "Messlauf",
   "en": "run"
  },
  "eichruf": {
   "de": "Eichmessung",
   "en": "calibration measurement"
  },
  "analyse": {
   "de": "Auswertung vorhandener Ergebnisse, keine neuen Läufe",
   "en": "analysis of existing results, no new runs"
  },
  "keine": {
   "de": "nicht durchgeführt",
   "en": "not run"
  },
  "notiz": {
   "de": "nur Notiz ohne Rohdaten, nicht gezählt",
   "en": "note without raw data, not counted"
  }
 },
 "series": [
  {
   "id": "R01",
   "from": "2026-09-13",
   "to": "2026-09-14",
   "block": "A",
   "title": {
    "de": "Erste Einzelaufgaben: Kosten, Plugins, Streuung",
    "en": "First single-task runs: cost, plugins, spread"
   },
   "kinds": [
    "api_kosten",
    "streuung",
    "bezug"
   ],
   "taskType": "aenderung",
   "version": "alt",
   "build": {
    "de": "Grace-Terminal, alte Architektur (13.–21.09.)",
    "en": "Grace terminal, old architecture (13–21 Sep)"
   },
   "unit": "lauf",
   "tasks": 2,
   "runsStarted": 47,
   "runsValid": 43,
   "runsValidWithGrace": 9,
   "runsValidClaudeAlone": 26,
   "runsValidOther": 8,
   "countedInTotals": true,
   "status": "partly_valid",
   "result": {
    "de": "Claude + Graces Paket bis +100 % Kosten, Grace + Stapel +194 % gegen Claude allein; 0 von 6 Grace-Läufen voll abgenommen; derselbe Aufbau streut 21–48 %.",
    "en": "Claude + Grace package up to +100 % cost, Grace + stack +194 % vs Claude alone; 0 of 6 Grace runs fully accepted; the same setup varies by 21–48 %."
   },
   "invalidReason": {
    "de": "2 Läufe durch Netzabbruch ausgefallen und wiederholt, 2 Vorläufe verworfen. Eine Plugin-Aussage wurde zurückgezogen (Arme ungleich ausgestattet); die Läufe selbst bleiben gültig.",
    "en": "2 runs lost to network errors and repeated, 2 pre-runs discarded. One plugin finding was withdrawn (arms not equally equipped); the runs themselves stay valid."
   },
   "headline": false,
   "caveat": null
  },
  {
   "id": "R02",
   "from": "2026-09-18",
   "to": "2026-09-18",
   "block": "A",
   "title": {
    "de": "Übergabepaket mit/ohne",
    "en": "Hand-over package with/without"
   },
   "kinds": [
    "api_kosten"
   ],
   "taskType": "aenderung",
   "version": "alt",
   "build": {
    "de": "Grace-Terminal, alte Architektur (13.–21.09.)",
    "en": "Grace terminal, old architecture (13–21 Sep)"
   },
   "unit": "lauf",
   "tasks": 4,
   "runsStarted": 10,
   "runsValid": 8,
   "runsValidWithGrace": 0,
   "runsValidClaudeAlone": 8,
   "runsValidOther": 0,
   "countedInTotals": true,
   "status": "partly_valid",
   "result": {
    "de": "Als Vergleich ungültig. Weil beide Arme ohne Paket liefen, sind die 8 Läufe Claude allein gegen Claude allein: sie streuen um 4,4 %.",
    "en": "Invalid as a comparison. Since both arms ran without the package, the 8 runs are Claude alone vs Claude alone: they vary by 4.4 %."
   },
   "invalidReason": {
    "de": "Der Arm „mit Paket“ bekam das Paket wegen eines Skriptfehlers nie; 2 weitere Läufe verworfen (Fehlstart, an einer Rückfrage hängen geblieben).",
    "en": "The \"with package\" arm never received the package (script bug); 2 more runs discarded (false start, stuck on a question)."
   },
   "headline": false,
   "caveat": null
  },
  {
   "id": "R03",
   "from": "2026-09-18",
   "to": "2026-09-19",
   "block": "A",
   "title": {
    "de": "Prüfarten kurz / spar / keine",
    "en": "Review modes short / cheap / none"
   },
   "kinds": [
    "api_kosten",
    "qualitaet"
   ],
   "taskType": "aenderung",
   "version": "alt",
   "build": {
    "de": "Grace-Terminal, alte Architektur (13.–21.09.)",
    "en": "Grace terminal, old architecture (13–21 Sep)"
   },
   "unit": "lauf",
   "tasks": 4,
   "runsStarted": 17,
   "runsValid": 16,
   "runsValidWithGrace": 12,
   "runsValidClaudeAlone": 4,
   "runsValidOther": 0,
   "countedInTotals": true,
   "status": "valid",
   "result": {
    "de": "Bis −74 % Kosten gegen Claude allein, aber nur durch ein billigeres Modell (+37 % Token); Grace ganz allein nur bei 2 von 4 gleichwertig.",
    "en": "Up to −74 % cost vs Claude alone, but only via a cheaper model (+37 % tokens); Grace entirely alone matched on only 2 of 4."
   },
   "invalidReason": {
    "de": "1 Lauf abgebrochen und nachgeholt.",
    "en": "1 run aborted and redone."
   },
   "headline": false,
   "caveat": null
  },
  {
   "id": "R04",
   "from": "2026-09-19",
   "to": "2026-09-19",
   "block": "D",
   "title": {
    "de": "Eichung: Limit-Anzeige gegen Kosten und Token",
    "en": "Calibration: limit reading against cost and tokens"
   },
   "kinds": [
    "eichung",
    "abo_limit"
   ],
   "taskType": "eichung",
   "version": "keine",
   "build": {
    "de": "ohne Grace",
    "en": "no Grace"
   },
   "unit": "eichruf",
   "tasks": null,
   "runsStarted": 6,
   "runsValid": 5,
   "runsValidWithGrace": 0,
   "runsValidClaudeAlone": 0,
   "runsValidOther": 5,
   "countedInTotals": true,
   "status": "calibration",
   "result": {
    "de": "Das Limit folgt eher den Kosten als den rohen Token (286 000 Opus-Token ≈ 1 006 000 Haiku-Token ≈ 1 Punkt).",
    "en": "The limit follows cost more than raw tokens (286,000 Opus tokens ≈ 1,006,000 Haiku tokens ≈ 1 point)."
   },
   "invalidReason": {
    "de": "Auflösung ganze Prozentpunkte, Cache-Anteil nicht variiert; 1 Vorlauf nicht gewertet.",
    "en": "Resolution whole points, cache share not varied; 1 pre-run not counted."
   },
   "headline": false,
   "caveat": null
  },
  {
   "id": "R05",
   "from": "2026-09-20",
   "to": "2026-09-20",
   "block": "D",
   "title": {
    "de": "Grundlast eines Claude-Aufrufs",
    "en": "Base load of one Claude call"
   },
   "kinds": [
    "eichung"
   ],
   "taskType": "eichung",
   "version": "keine",
   "build": {
    "de": "ohne Grace",
    "en": "no Grace"
   },
   "unit": "eichruf",
   "tasks": null,
   "runsStarted": 11,
   "runsValid": 11,
   "runsValidWithGrace": 0,
   "runsValidClaudeAlone": 0,
   "runsValidOther": 11,
   "countedInTotals": true,
   "status": "calibration",
   "result": {
    "de": "Ein leerer Aufruf kostet ~22 700 Token; ohne Werkzeuge ~7 500.",
    "en": "An empty call costs ~22,700 tokens; without tools ~7,500."
   },
   "invalidReason": null,
   "headline": false,
   "caveat": null
  },
  {
   "id": "R06",
   "from": "2026-09-20",
   "to": "2026-09-20",
   "block": "A",
   "title": {
    "de": "Schlanke Prüfung mit Reparaturrunde",
    "en": "Slim review with repair round"
   },
   "kinds": [
    "api_kosten",
    "qualitaet",
    "abo_limit"
   ],
   "taskType": "aenderung",
   "version": "alt",
   "build": {
    "de": "Grace-Terminal, alte Architektur (13.–21.09.)",
    "en": "Grace terminal, old architecture (13–21 Sep)"
   },
   "unit": "lauf",
   "tasks": 4,
   "runsStarted": 16,
   "runsValid": 16,
   "runsValidWithGrace": 12,
   "runsValidClaudeAlone": 4,
   "runsValidOther": 0,
   "countedInTotals": true,
   "status": "valid",
   "result": {
    "de": "−89 % Kosten, 4 von 4 Endständen gleich wie Claude allein – aber nur 4 kleine Aufgaben in einem Projekt; auf 30 neuen Aufgaben (R16) nicht bestätigt.",
    "en": "−89 % cost, 4 of 4 end states identical to Claude alone – but only 4 small tasks in one project; not confirmed on 30 new tasks (R16)."
   },
   "invalidReason": null,
   "headline": false,
   "caveat": null
  },
  {
   "id": "R07",
   "from": "2026-09-21",
   "to": "2026-09-21",
   "block": "A",
   "title": {
    "de": "Probelauf der Messkette",
    "en": "Dry run of the test harness"
   },
   "kinds": [
    "aufbau"
   ],
   "taskType": "aenderung",
   "version": "alt",
   "build": {
    "de": "Grace-Terminal, alte Architektur (13.–21.09.)",
    "en": "Grace terminal, old architecture (13–21 Sep)"
   },
   "unit": "lauf",
   "tasks": 5,
   "runsStarted": 10,
   "runsValid": 6,
   "runsValidWithGrace": 1,
   "runsValidClaudeAlone": 5,
   "runsValidOther": 0,
   "countedInTotals": true,
   "status": "partly_valid",
   "result": {
    "de": "Nur Aufbauprüfung, kein Ergebniswert.",
    "en": "Harness check only, no result value."
   },
   "invalidReason": {
    "de": "4 von 5 Grace-Läufen durch einen Fehler im Wächter abgebrochen.",
    "en": "4 of 5 Grace runs killed by a watchdog bug."
   },
   "headline": false,
   "caveat": null
  },
  {
   "id": "R08",
   "from": "2026-09-21",
   "to": "2026-09-21",
   "block": "A",
   "title": {
    "de": "Nachweis-Lauf aller Arme",
    "en": "Proof run of all arms"
   },
   "kinds": [
    "aufbau"
   ],
   "taskType": "aenderung",
   "version": "alt",
   "build": {
    "de": "Grace-Terminal, alte Architektur (13.–21.09.)",
    "en": "Grace terminal, old architecture (13–21 Sep)"
   },
   "unit": "lauf",
   "tasks": 0,
   "runsStarted": 5,
   "runsValid": 0,
   "runsValidWithGrace": 0,
   "runsValidClaudeAlone": 0,
   "runsValidOther": 0,
   "countedInTotals": true,
   "status": "invalid",
   "result": {
    "de": "Aufbauprüfung nicht bestanden.",
    "en": "Harness check failed."
   },
   "invalidReason": {
    "de": "Speichermessung fehlte in 5 von 5 Läufen.",
    "en": "Memory reading missing in 5 of 5 runs."
   },
   "headline": false,
   "caveat": null
  },
  {
   "id": "R09",
   "from": "2026-09-21",
   "to": "2026-09-21",
   "block": "A",
   "title": {
    "de": "Leichte Reihe (mechanisch geprüft)",
    "en": "Easy series (mechanical check)"
   },
   "kinds": [
    "qualitaet"
   ],
   "taskType": "aenderung",
   "version": "alt",
   "build": {
    "de": "Grace-Terminal, alte Architektur (13.–21.09.)",
    "en": "Grace terminal, old architecture (13–21 Sep)"
   },
   "unit": "lauf",
   "tasks": 18,
   "runsStarted": 33,
   "runsValid": 31,
   "runsValidWithGrace": 18,
   "runsValidClaudeAlone": 13,
   "runsValidOther": 0,
   "countedInTotals": true,
   "status": "partly_valid",
   "result": {
    "de": "Grace allein löste 4 von 18, Claude allein 12 von 13.",
    "en": "Grace alone solved 4 of 18, Claude alone 12 of 13."
   },
   "invalidReason": {
    "de": "Nach 18 von 70 geplanten Aufgaben abgebrochen (Festplatte voll); 2 Läufe ohne Ergebnis.",
    "en": "Stopped after 18 of 70 planned tasks (disk full); 2 runs without result."
   },
   "headline": false,
   "caveat": null
  },
  {
   "id": "R10",
   "from": "2026-09-21",
   "to": "2026-09-21",
   "block": "A",
   "title": {
    "de": "Streuungsreihe (je 3 Wiederholungen)",
    "en": "Spread series (3 repeats each)"
   },
   "kinds": [
    "streuung",
    "api_kosten"
   ],
   "taskType": "aenderung",
   "version": "alt",
   "build": {
    "de": "Grace-Terminal, alte Architektur (13.–21.09.)",
    "en": "Grace terminal, old architecture (13–21 Sep)"
   },
   "unit": "lauf",
   "tasks": 3,
   "runsStarted": 18,
   "runsValid": 18,
   "runsValidWithGrace": 9,
   "runsValidClaudeAlone": 9,
   "runsValidOther": 0,
   "countedInTotals": true,
   "status": "valid",
   "result": {
    "de": "Dieselbe Aufgabe streut im Mittel um 15 % (höchstens 27 %). Grace schlank: −96 % Token, −92 % Kosten; Qualität nicht bewertet.",
    "en": "The same task varies by 15 % on average (27 % at most). Grace slim: −96 % tokens, −92 % cost; quality not judged."
   },
   "invalidReason": null,
   "headline": false,
   "caveat": null
  },
  {
   "id": "R11",
   "from": "2026-09-21",
   "to": "2026-09-22",
   "block": "B",
   "title": {
    "de": "Hauptmessung 1: 30 Änderungen, 5 Arme",
    "en": "Main test 1: 30 changes, 5 arms"
   },
   "kinds": [
    "api_kosten",
    "qualitaet"
   ],
   "taskType": "aenderung",
   "version": "alt",
   "build": {
    "de": "alte Architektur (21.–23.09.)",
    "en": "old architecture (21–23 Sep)"
   },
   "unit": "lauf",
   "tasks": 30,
   "runsStarted": 150,
   "runsValid": 149,
   "runsValidWithGrace": 119,
   "runsValidClaudeAlone": 30,
   "runsValidOther": 0,
   "countedInTotals": true,
   "status": "valid",
   "result": {
    "de": "Kosten −97 % (Grace schlank) – aber Claude allein war bei 24 von 30 besser. Die Qualitätsschranke war gerissen, die Ersparnis zählt deshalb nicht.",
    "en": "Cost −97 % (Grace slim) – but Claude alone was better on 24 of 30. The quality gate failed, so the saving does not count."
   },
   "invalidReason": {
    "de": "1 Lauf fehlgeschlagen; die Bewertung selbst ist nur eingeschränkt verwertbar (ein Richter fiel an einem Kontrollpaar durch).",
    "en": "1 run failed; the judging itself is of limited use (one judge failed a control pair)."
   },
   "headline": false,
   "caveat": null
  },
  {
   "id": "R12",
   "from": "2026-09-22",
   "to": "2026-09-22",
   "block": "D",
   "title": {
    "de": "Richter-Eichung an Musterpaaren",
    "en": "Judge calibration on sample pairs"
   },
   "kinds": [
    "eichung",
    "qualitaet"
   ],
   "taskType": "eichung",
   "version": "keine",
   "build": {
    "de": "ohne Grace",
    "en": "no Grace"
   },
   "unit": "analyse",
   "tasks": null,
   "runsStarted": 0,
   "runsValid": 0,
   "runsValidWithGrace": 0,
   "runsValidClaudeAlone": 0,
   "runsValidOther": 0,
   "countedInTotals": false,
   "status": "calibration",
   "result": {
    "de": "32 Urteile: Opus 15 von 16, Haiku 16 von 16 richtig.",
    "en": "32 verdicts: Opus 15 of 16, Haiku 16 of 16 correct."
   },
   "invalidReason": null,
   "headline": false,
   "caveat": null
  },
  {
   "id": "R13",
   "from": "2026-09-22",
   "to": "2026-09-24",
   "block": "D",
   "title": {
    "de": "Geplante Limit- und API-Reihen",
    "en": "Planned limit and API series"
   },
   "kinds": [
    "abo_limit",
    "api_kosten"
   ],
   "taskType": "eichung",
   "version": "alt",
   "build": null,
   "unit": "keine",
   "tasks": null,
   "runsStarted": 0,
   "runsValid": 0,
   "runsValidWithGrace": 0,
   "runsValidClaudeAlone": 0,
   "runsValidOther": 0,
   "countedInTotals": false,
   "status": "not_run",
   "result": {
    "de": "Nie gestartet.",
    "en": "Never started."
   },
   "invalidReason": null,
   "headline": false,
   "caveat": null
  },
  {
   "id": "R14",
   "from": "2026-09-22",
   "to": "2026-09-23",
   "block": "B",
   "title": {
    "de": "Hauptmessung 2: dieselben 30 Änderungen",
    "en": "Main test 2: same 30 changes"
   },
   "kinds": [
    "api_kosten",
    "qualitaet"
   ],
   "taskType": "aenderung",
   "version": "alt",
   "build": {
    "de": "alte Architektur (21.–23.09.)",
    "en": "old architecture (21–23 Sep)"
   },
   "unit": "lauf",
   "tasks": 30,
   "runsStarted": 150,
   "runsValid": 146,
   "runsValidWithGrace": 116,
   "runsValidClaudeAlone": 30,
   "runsValidOther": 0,
   "countedInTotals": true,
   "status": "valid",
   "result": {
    "de": "Kosten −96 % (Grace schlank), Claude allein besser bei 21 von 30. Arm „Grace arbeitet, Claude macht fertig“: Claude allein besser bei 12, gleich 12, Grace besser 4 (von 29).",
    "en": "Cost −96 % (Grace slim), Claude alone better on 21 of 30. Arm \"Grace works, Claude finishes\": Claude alone better on 12, equal 12, Grace better 4 (of 29)."
   },
   "invalidReason": {
    "de": "4 Läufe fehlgeschlagen; die erste Bewertung war ungültig (Urteile vom Vortag übernommen) und wurde am 25.09. neu gerichtet.",
    "en": "4 runs failed; the first judging was invalid (verdicts copied from the day before) and was redone on 25 Sep."
   },
   "headline": false,
   "caveat": null
  },
  {
   "id": "R15",
   "from": "2026-09-23",
   "to": "2026-09-23",
   "block": "B",
   "title": {
    "de": "Vorprobe Reparatur-Zeitgrenze",
    "en": "Pre-check repair time limit"
   },
   "kinds": [
    "aufbau"
   ],
   "taskType": "aenderung",
   "version": "alt",
   "build": {
    "de": "alte Architektur (21.–23.09.)",
    "en": "old architecture (21–23 Sep)"
   },
   "unit": "lauf",
   "tasks": 1,
   "runsStarted": 1,
   "runsValid": 1,
   "runsValidWithGrace": 1,
   "runsValidClaudeAlone": 0,
   "runsValidOther": 0,
   "countedInTotals": true,
   "status": "valid",
   "result": {
    "de": "Einzelprobe, kein Vergleichswert.",
    "en": "Single probe, no comparison."
   },
   "invalidReason": null,
   "headline": false,
   "caveat": null
  },
  {
   "id": "R16",
   "from": "2026-09-23",
   "to": "2026-09-24",
   "block": "B",
   "title": {
    "de": "30 neue Änderungen: schlank gegen allein",
    "en": "30 new changes: slim vs alone"
   },
   "kinds": [
    "api_kosten",
    "qualitaet"
   ],
   "taskType": "aenderung",
   "version": "alt",
   "build": {
    "de": "alte Architektur (21.–23.09.)",
    "en": "old architecture (21–23 Sep)"
   },
   "unit": "lauf",
   "tasks": 30,
   "runsStarted": 60,
   "runsValid": 60,
   "runsValidWithGrace": 30,
   "runsValidClaudeAlone": 30,
   "runsValidOther": 0,
   "countedInTotals": true,
   "status": "valid",
   "result": {
    "de": "Kosten −97 %, aber Claude allein besser bei 25 von 30 (336 Blindurteile; im zweiten Richterverfahren 23 von 30). Nicht als „gleich gut“ verwertbar.",
    "en": "Cost −97 %, but Claude alone better on 25 of 30 (336 blind verdicts; 23 of 30 with the second judging method). Not usable as \"equally good\"."
   },
   "invalidReason": null,
   "headline": false,
   "caveat": null
  },
  {
   "id": "R17",
   "from": "2026-09-24",
   "to": "2026-09-24",
   "block": "D",
   "title": {
    "de": "Limit-Eichung mit kleinen Lasten",
    "en": "Limit calibration with small loads"
   },
   "kinds": [
    "eichung",
    "abo_limit"
   ],
   "taskType": "eichung",
   "version": "keine",
   "build": {
    "de": "ohne Grace",
    "en": "no Grace"
   },
   "unit": "eichruf",
   "tasks": null,
   "runsStarted": 4,
   "runsValid": 0,
   "runsValidWithGrace": 0,
   "runsValidClaudeAlone": 0,
   "runsValidOther": 0,
   "countedInTotals": true,
   "status": "invalid",
   "result": {
    "de": "Nicht messbar.",
    "en": "Not measurable."
   },
   "invalidReason": {
    "de": "Last etwa 28× zu klein für die Anzeige-Auflösung.",
    "en": "Load about 28× too small for the display resolution."
   },
   "headline": false,
   "caveat": null
  },
  {
   "id": "R18",
   "from": "2026-09-24",
   "to": "2026-09-24",
   "block": "B",
   "title": {
    "de": "Gegenprobe nach zwei Reparaturen",
    "en": "Re-check after two repairs"
   },
   "kinds": [
    "aufbau"
   ],
   "taskType": "aenderung",
   "version": "alt",
   "build": {
    "de": "alte Architektur nach zwei Reparaturen (24.09.)",
    "en": "old architecture after two repairs (24 Sep)"
   },
   "unit": "lauf",
   "tasks": 5,
   "runsStarted": 10,
   "runsValid": 5,
   "runsValidWithGrace": 5,
   "runsValidClaudeAlone": 0,
   "runsValidOther": 0,
   "countedInTotals": true,
   "status": "partly_valid",
   "result": {
    "de": "Nach der Reparatur liefen 5 von 5 Aufgaben durch; kein Qualitätsurteil.",
    "en": "After the fix 5 of 5 tasks ran through; no quality verdict."
   },
   "invalidReason": {
    "de": "5 Fehlstarts vor der Reparatur.",
    "en": "5 false starts before the fix."
   },
   "headline": false,
   "caveat": null
  },
  {
   "id": "R19",
   "from": "2026-09-24",
   "to": "2026-09-24",
   "block": "B",
   "title": {
    "de": "Nach den Reparaturen: Übergabe gegen allein (11)",
    "en": "After the repairs: hand-over vs alone (11)"
   },
   "kinds": [
    "api_kosten",
    "qualitaet"
   ],
   "taskType": "aenderung",
   "version": "alt",
   "build": {
    "de": "alte Architektur nach zwei Reparaturen (24.09.)",
    "en": "old architecture after two repairs (24 Sep)"
   },
   "unit": "lauf",
   "tasks": 11,
   "runsStarted": 22,
   "runsValid": 22,
   "runsValidWithGrace": 11,
   "runsValidClaudeAlone": 11,
   "runsValidOther": 0,
   "countedInTotals": true,
   "status": "valid",
   "result": {
    "de": "Grace + Übergabe war 24 % TEURER als Claude allein; Qualität je nach Richterverfahren 4:4:3 oder 2:4:5 (Claude besser : gleich : Grace besser).",
    "en": "Grace + hand-over cost 24 % MORE than Claude alone; quality 4:4:3 or 2:4:5 depending on judging method (Claude better : equal : Grace better)."
   },
   "invalidReason": null,
   "headline": false,
   "caveat": null
  },
  {
   "id": "R20",
   "from": "2026-09-24",
   "to": "2026-09-25",
   "block": "B",
   "title": {
    "de": "Handmessung Website-Neubau",
    "en": "Hand-timed website build"
   },
   "kinds": [
    "api_kosten"
   ],
   "taskType": "neubau",
   "version": "alt",
   "build": null,
   "unit": "notiz",
   "tasks": null,
   "runsStarted": 1,
   "runsValid": 0,
   "runsValidWithGrace": 0,
   "runsValidClaudeAlone": 0,
   "runsValidOther": 0,
   "countedInTotals": false,
   "status": "not_counted",
   "result": {
    "de": "Nur die Claude-Seite lief (13 Mio. Token, 71 min); kein Vergleich.",
    "en": "Only the Claude side ran (13 M tokens, 71 min); no comparison."
   },
   "invalidReason": {
    "de": "Grace-Seite nie abgeschlossen; Rohdaten nicht mehr vorhanden – deshalb nirgends mitgezählt.",
    "en": "Grace side never finished; raw data no longer available – therefore not counted anywhere."
   },
   "headline": false,
   "caveat": null
  },
  {
   "id": "R21",
   "from": "2026-09-25",
   "to": "2026-09-25",
   "block": "D",
   "title": {
    "de": "Zwei Richterverfahren im Vergleich",
    "en": "Two judging methods compared"
   },
   "kinds": [
    "eichung",
    "qualitaet"
   ],
   "taskType": "eichung",
   "version": "keine",
   "build": {
    "de": "ohne Grace",
    "en": "no Grace"
   },
   "unit": "analyse",
   "tasks": null,
   "runsStarted": 0,
   "runsValid": 0,
   "runsValidWithGrace": 0,
   "runsValidClaudeAlone": 0,
   "runsValidOther": 0,
   "countedInTotals": false,
   "status": "calibration",
   "result": {
    "de": "Beide Verfahren urteilen nur bei 99 von 130 Paaren (76 %) gleich.",
    "en": "Both methods agree on only 99 of 130 pairs (76 %)."
   },
   "invalidReason": null,
   "headline": false,
   "caveat": null
  },
  {
   "id": "R22",
   "from": "2026-09-25",
   "to": "2026-09-25",
   "block": "C",
   "title": {
    "de": "Nachbesserung 1: Claude liefert Patch",
    "en": "Finisher 1: Claude returns a patch"
   },
   "kinds": [
    "api_kosten",
    "qualitaet"
   ],
   "taskType": "aenderung",
   "version": "nachbesserung",
   "build": {
    "de": "Grace-Endstände aus R16 (Stand 23.09.) + Claude als Nachbesserer",
    "en": "Grace end states from R16 (23 Sep build) + Claude as finisher"
   },
   "unit": "lauf",
   "tasks": 14,
   "runsStarted": 14,
   "runsValid": 14,
   "runsValidWithGrace": 14,
   "runsValidClaudeAlone": 0,
   "runsValidOther": 0,
   "countedInTotals": true,
   "status": "valid",
   "result": {
    "de": "3 von 14 gleich oder besser; mehrere Patches ließen sich nicht einspielen.",
    "en": "3 of 14 equal or better; several patches did not apply."
   },
   "invalidReason": null,
   "headline": false,
   "caveat": null
  },
  {
   "id": "R23",
   "from": "2026-09-25",
   "to": "2026-09-25",
   "block": "C",
   "title": {
    "de": "Nachbesserung 2: Opus bessert nach (max. 6 Runden)",
    "en": "Finisher 2: Opus fixes up (max 6 rounds)"
   },
   "kinds": [
    "api_kosten",
    "qualitaet"
   ],
   "taskType": "aenderung",
   "version": "nachbesserung",
   "build": {
    "de": "Grace-Endstände aus R16 (Stand 23.09.) + Claude als Nachbesserer",
    "en": "Grace end states from R16 (23 Sep build) + Claude as finisher"
   },
   "unit": "lauf",
   "tasks": 30,
   "runsStarted": 30,
   "runsValid": 30,
   "runsValidWithGrace": 30,
   "runsValidClaudeAlone": 0,
   "runsValidOther": 0,
   "countedInTotals": true,
   "status": "valid",
   "result": {
    "de": "18 von 30 gleich oder besser bei 44 % der Kosten von Claude allein.",
    "en": "18 of 30 equal or better at 44 % of Claude alone's cost."
   },
   "invalidReason": null,
   "headline": false,
   "caveat": null
  },
  {
   "id": "R24",
   "from": "2026-09-25",
   "to": "2026-09-25",
   "block": "C",
   "title": {
    "de": "Nachbesserung 2b: mit Grace-Selbstprüfbogen",
    "en": "Finisher 2b: with Grace self-check sheet"
   },
   "kinds": [
    "api_kosten",
    "qualitaet"
   ],
   "taskType": "aenderung",
   "version": "nachbesserung",
   "build": {
    "de": "Grace-Endstände aus R16 (Stand 23.09.) + Claude als Nachbesserer",
    "en": "Grace end states from R16 (23 Sep build) + Claude as finisher"
   },
   "unit": "lauf",
   "tasks": 30,
   "runsStarted": 30,
   "runsValid": 30,
   "runsValidWithGrace": 30,
   "runsValidClaudeAlone": 0,
   "runsValidOther": 0,
   "countedInTotals": true,
   "status": "valid",
   "result": {
    "de": "17 von 30 bei 59 % der Kosten – schlechter und teurer als 2.",
    "en": "17 of 30 at 59 % of cost – worse and dearer than 2."
   },
   "invalidReason": null,
   "headline": false,
   "caveat": null
  },
  {
   "id": "R25",
   "from": "2026-09-25",
   "to": "2026-09-25",
   "block": "C",
   "title": {
    "de": "Nachbesserung 3: knapper Nachbesserer",
    "en": "Finisher 3: terse finisher"
   },
   "kinds": [
    "api_kosten",
    "qualitaet"
   ],
   "taskType": "aenderung",
   "version": "nachbesserung",
   "build": {
    "de": "Grace-Endstände aus R16 (Stand 23.09.) + Claude als Nachbesserer",
    "en": "Grace end states from R16 (23 Sep build) + Claude as finisher"
   },
   "unit": "lauf",
   "tasks": 30,
   "runsStarted": 30,
   "runsValid": 28,
   "runsValidWithGrace": 28,
   "runsValidClaudeAlone": 0,
   "runsValidOther": 0,
   "countedInTotals": true,
   "status": "partly_valid",
   "result": {
    "de": "10 von 30 gleich oder besser.",
    "en": "10 of 30 equal or better."
   },
   "invalidReason": {
    "de": "Einzelne Läufe/Urteile ungültig (Richter-Aussetzer oder Abbruch), nicht gezählt.",
    "en": "Single runs/verdicts invalid (judge dropouts or aborts), not counted."
   },
   "headline": false,
   "caveat": null
  },
  {
   "id": "R26",
   "from": "2026-09-25",
   "to": "2026-09-25",
   "block": "C",
   "title": {
    "de": "Nachbesserer Sonnet",
    "en": "Finisher Sonnet"
   },
   "kinds": [
    "api_kosten",
    "qualitaet"
   ],
   "taskType": "aenderung",
   "version": "nachbesserung",
   "build": {
    "de": "Grace-Endstände aus R16 (Stand 23.09.) + Claude als Nachbesserer",
    "en": "Grace end states from R16 (23 Sep build) + Claude as finisher"
   },
   "unit": "lauf",
   "tasks": 30,
   "runsStarted": 30,
   "runsValid": 30,
   "runsValidWithGrace": 30,
   "runsValidClaudeAlone": 0,
   "runsValidOther": 0,
   "countedInTotals": true,
   "status": "valid",
   "result": {
    "de": "18 von 30 bei 26 % der Kosten.",
    "en": "18 of 30 at 26 % of cost."
   },
   "invalidReason": null,
   "headline": false,
   "caveat": null
  },
  {
   "id": "R27",
   "from": "2026-09-25",
   "to": "2026-09-25",
   "block": "C",
   "title": {
    "de": "Nachbesserer Haiku",
    "en": "Finisher Haiku"
   },
   "kinds": [
    "api_kosten",
    "qualitaet"
   ],
   "taskType": "aenderung",
   "version": "nachbesserung",
   "build": {
    "de": "Grace-Endstände aus R16 (Stand 23.09.) + Claude als Nachbesserer",
    "en": "Grace end states from R16 (23 Sep build) + Claude as finisher"
   },
   "unit": "lauf",
   "tasks": 30,
   "runsStarted": 30,
   "runsValid": 30,
   "runsValidWithGrace": 30,
   "runsValidClaudeAlone": 0,
   "runsValidOther": 0,
   "countedInTotals": true,
   "status": "valid",
   "result": {
    "de": "11 von 30 bei 23 % der Kosten – verworfen.",
    "en": "11 of 30 at 23 % of cost – dropped."
   },
   "invalidReason": null,
   "headline": false,
   "caveat": null
  },
  {
   "id": "R28",
   "from": "2026-09-25",
   "to": "2026-09-25",
   "block": "C",
   "title": {
    "de": "Nachbesserer Sonnet ohne Plugins",
    "en": "Finisher Sonnet without plugins"
   },
   "kinds": [
    "api_kosten",
    "qualitaet"
   ],
   "taskType": "aenderung",
   "version": "nachbesserung",
   "build": {
    "de": "Grace-Endstände aus R16 (Stand 23.09.) + Claude als Nachbesserer",
    "en": "Grace end states from R16 (23 Sep build) + Claude as finisher"
   },
   "unit": "lauf",
   "tasks": 30,
   "runsStarted": 30,
   "runsValid": 30,
   "runsValidWithGrace": 30,
   "runsValidClaudeAlone": 0,
   "runsValidOther": 0,
   "countedInTotals": true,
   "status": "valid",
   "result": {
    "de": "15 von 30 bei 25 % der Kosten.",
    "en": "15 of 30 at 25 % of cost."
   },
   "invalidReason": null,
   "headline": false,
   "caveat": null
  },
  {
   "id": "R29",
   "from": "2026-09-25",
   "to": "2026-09-25",
   "block": "C",
   "title": {
    "de": "Nachbesserer „Sonnet knapp“, 3 Wiederholungen",
    "en": "Finisher \"Sonnet terse\", 3 repeats"
   },
   "kinds": [
    "api_kosten",
    "qualitaet",
    "streuung"
   ],
   "taskType": "aenderung",
   "version": "nachbesserung",
   "build": {
    "de": "Grace-Endstände aus R16 (Stand 23.09.) + Claude als Nachbesserer",
    "en": "Grace end states from R16 (23 Sep build) + Claude as finisher"
   },
   "unit": "lauf",
   "tasks": 30,
   "runsStarted": 90,
   "runsValid": 88,
   "runsValidWithGrace": 88,
   "runsValidClaudeAlone": 0,
   "runsValidOther": 0,
   "countedInTotals": true,
   "status": "partly_valid",
   "result": {
    "de": "18 · 10 · 12 von 30 (zusammen 44 %) bei 23–28 % der Kosten; Ziel ≈ 73 % verfehlt.",
    "en": "18 · 10 · 12 of 30 (44 % overall) at 23–28 % of cost; target ≈ 73 % missed."
   },
   "invalidReason": {
    "de": "Einzelne Läufe/Urteile ungültig (Richter-Aussetzer oder Abbruch), nicht gezählt.",
    "en": "Single runs/verdicts invalid (judge dropouts or aborts), not counted."
   },
   "headline": false,
   "caveat": null
  },
  {
   "id": "R30",
   "from": "2026-09-25",
   "to": "2026-09-25",
   "block": "C",
   "title": {
    "de": "„Sonnet knapp“ mit Denkstufe low",
    "en": "\"Sonnet terse\" with low effort"
   },
   "kinds": [
    "api_kosten",
    "qualitaet"
   ],
   "taskType": "aenderung",
   "version": "nachbesserung",
   "build": {
    "de": "Grace-Endstände aus R16 (Stand 23.09.) + Claude als Nachbesserer",
    "en": "Grace end states from R16 (23 Sep build) + Claude as finisher"
   },
   "unit": "lauf",
   "tasks": 30,
   "runsStarted": 30,
   "runsValid": 30,
   "runsValidWithGrace": 30,
   "runsValidClaudeAlone": 0,
   "runsValidOther": 0,
   "countedInTotals": true,
   "status": "valid",
   "result": {
    "de": "10 von 30 bei 19 % der Kosten – verworfen.",
    "en": "10 of 30 at 19 % of cost – dropped."
   },
   "invalidReason": null,
   "headline": false,
   "caveat": null
  },
  {
   "id": "R31",
   "from": "2026-09-25",
   "to": "2026-09-25",
   "block": "C",
   "title": {
    "de": "„Sonnet knapp“ mit 5-Minuten-Cache",
    "en": "\"Sonnet terse\" with 5-minute cache"
   },
   "kinds": [
    "api_kosten",
    "qualitaet"
   ],
   "taskType": "aenderung",
   "version": "nachbesserung",
   "build": {
    "de": "Grace-Endstände aus R16 (Stand 23.09.) + Claude als Nachbesserer",
    "en": "Grace end states from R16 (23 Sep build) + Claude as finisher"
   },
   "unit": "lauf",
   "tasks": 30,
   "runsStarted": 30,
   "runsValid": 30,
   "runsValidWithGrace": 30,
   "runsValidClaudeAlone": 0,
   "runsValidOther": 0,
   "countedInTotals": true,
   "status": "valid",
   "result": {
    "de": "12 von 30 bei 23 %; der Cache sparte nur 2 %.",
    "en": "12 of 30 at 23 %; the cache saved only 2 %."
   },
   "invalidReason": null,
   "headline": false,
   "caveat": null
  },
  {
   "id": "R32",
   "from": "2026-09-25",
   "to": "2026-09-25",
   "block": "C",
   "title": {
    "de": "„Sonnet knapp“ mit Pflichtliste",
    "en": "\"Sonnet terse\" with mandatory list"
   },
   "kinds": [
    "api_kosten",
    "qualitaet"
   ],
   "taskType": "aenderung",
   "version": "nachbesserung",
   "build": {
    "de": "Grace-Endstände aus R16 (Stand 23.09.) + Claude als Nachbesserer",
    "en": "Grace end states from R16 (23 Sep build) + Claude as finisher"
   },
   "unit": "lauf",
   "tasks": 30,
   "runsStarted": 30,
   "runsValid": 30,
   "runsValidWithGrace": 30,
   "runsValidClaudeAlone": 0,
   "runsValidOther": 0,
   "countedInTotals": true,
   "status": "valid",
   "result": {
    "de": "12 von 30 bei 23 % – kein Gewinn.",
    "en": "12 of 30 at 23 % – no gain."
   },
   "invalidReason": null,
   "headline": false,
   "caveat": null
  },
  {
   "id": "R33",
   "from": "2026-09-25",
   "to": "2026-09-25",
   "block": "C",
   "title": {
    "de": "„Sonnet knapp“ mit bis zu 15 Runden",
    "en": "\"Sonnet terse\" with up to 15 rounds"
   },
   "kinds": [
    "api_kosten",
    "qualitaet"
   ],
   "taskType": "aenderung",
   "version": "nachbesserung",
   "build": {
    "de": "Grace-Endstände aus R16 (Stand 23.09.) + Claude als Nachbesserer",
    "en": "Grace end states from R16 (23 Sep build) + Claude as finisher"
   },
   "unit": "lauf",
   "tasks": 30,
   "runsStarted": 30,
   "runsValid": 30,
   "runsValidWithGrace": 30,
   "runsValidClaudeAlone": 0,
   "runsValidOther": 0,
   "countedInTotals": true,
   "status": "valid",
   "result": {
    "de": "16 von 30 bei 34 % – 50 % teurer ohne Gewinn.",
    "en": "16 of 30 at 34 % – 50 % dearer, no gain."
   },
   "invalidReason": null,
   "headline": false,
   "caveat": null
  },
  {
   "id": "R34",
   "from": "2026-09-25",
   "to": "2026-09-25",
   "block": "C",
   "title": {
    "de": "„Sonnet knapp“ korrigiert, vorab festgelegt, 3 Wiederholungen",
    "en": "\"Sonnet terse\" fixed, pre-registered, 3 repeats"
   },
   "kinds": [
    "api_kosten",
    "qualitaet",
    "streuung"
   ],
   "taskType": "aenderung",
   "version": "nachbesserung",
   "build": {
    "de": "Grace-Endstände aus R16 (Stand 23.09.) + Claude als Nachbesserer",
    "en": "Grace end states from R16 (23 Sep build) + Claude as finisher"
   },
   "unit": "lauf",
   "tasks": 30,
   "runsStarted": 90,
   "runsValid": 90,
   "runsValidWithGrace": 90,
   "runsValidClaudeAlone": 0,
   "runsValidOther": 0,
   "countedInTotals": true,
   "status": "valid",
   "result": {
    "de": "37 von 90 (41 %) bei 32–33 % der Kosten. Vorab festgelegtes Ziel (≥ 48/90 bei ≤ 27 %) verfehlt.",
    "en": "37 of 90 (41 %) at 32–33 % of cost. Pre-registered target (≥ 48/90 at ≤ 27 %) missed."
   },
   "invalidReason": null,
   "headline": false,
   "caveat": null
  },
  {
   "id": "R35",
   "from": "2026-09-25",
   "to": "2026-09-25",
   "block": "C",
   "title": {
    "de": "Staffel Sonnet → Opus",
    "en": "Relay Sonnet → Opus"
   },
   "kinds": [
    "api_kosten",
    "qualitaet"
   ],
   "taskType": "aenderung",
   "version": "nachbesserung",
   "build": {
    "de": "Grace-Endstände aus R16 (Stand 23.09.) + Claude als Nachbesserer",
    "en": "Grace end states from R16 (23 Sep build) + Claude as finisher"
   },
   "unit": "lauf",
   "tasks": 12,
   "runsStarted": 12,
   "runsValid": 12,
   "runsValidWithGrace": 12,
   "runsValidClaudeAlone": 0,
   "runsValidOther": 0,
   "countedInTotals": true,
   "status": "valid",
   "result": {
    "de": "Opus rettet 4 von 12 → 22 von 30 (73 %) bei 41 % der Kosten. Nur Obergrenze: die Auswahl traf ein Richter, den das Produkt nicht hat.",
    "en": "Opus rescues 4 of 12 → 22 of 30 (73 %) at 41 % of cost. Upper bound only: a judge picked the cases, which the product cannot do."
   },
   "invalidReason": null,
   "headline": false,
   "caveat": null
  },
  {
   "id": "R36",
   "from": "2026-09-25",
   "to": "2026-09-25",
   "block": "C",
   "title": {
    "de": "Zweiter Durchgang auf 12 offenen Aufgaben",
    "en": "Second pass on 12 open tasks"
   },
   "kinds": [
    "api_kosten",
    "qualitaet"
   ],
   "taskType": "aenderung",
   "version": "nachbesserung",
   "build": {
    "de": "Grace-Endstände aus R16 (Stand 23.09.) + Claude als Nachbesserer",
    "en": "Grace end states from R16 (23 Sep build) + Claude as finisher"
   },
   "unit": "lauf",
   "tasks": 12,
   "runsStarted": 24,
   "runsValid": 24,
   "runsValidWithGrace": 24,
   "runsValidClaudeAlone": 0,
   "runsValidOther": 0,
   "countedInTotals": true,
   "status": "valid",
   "result": {
    "de": "Opus rettet 4 von 12, Sonnet 0 von 12 (Auswahl per Richter, Obergrenze).",
    "en": "Opus rescues 4 of 12, Sonnet 0 of 12 (judge-selected, upper bound)."
   },
   "invalidReason": null,
   "headline": false,
   "caveat": null
  },
  {
   "id": "R37",
   "from": "2026-09-25",
   "to": "2026-09-25",
   "block": "D",
   "title": {
    "de": "Erkennt ein Prüfer „fertig“ gegen „offen“?",
    "en": "Can a checker tell \"done\" from \"open\"?"
   },
   "kinds": [
    "eichung"
   ],
   "taskType": "eichung",
   "version": "keine",
   "build": {
    "de": "ohne Grace",
    "en": "no Grace"
   },
   "unit": "analyse",
   "tasks": null,
   "runsStarted": 0,
   "runsValid": 0,
   "runsValidWithGrace": 0,
   "runsValidClaudeAlone": 0,
   "runsValidOther": 0,
   "countedInTotals": false,
   "status": "calibration",
   "result": {
    "de": "Kein Prüfer trennt zuverlässig (mechanisch 14 falsche Alarme bei 24).",
    "en": "No checker separates reliably (mechanical: 14 false alarms out of 24)."
   },
   "invalidReason": null,
   "headline": false,
   "caveat": null
  },
  {
   "id": "R38",
   "from": "2026-09-25",
   "to": "2026-09-25",
   "block": "D",
   "title": {
    "de": "Claude gegen Claude (30 Änderungen)",
    "en": "Claude vs Claude (30 changes)"
   },
   "kinds": [
    "eichung",
    "streuung",
    "qualitaet"
   ],
   "taskType": "eichung",
   "version": "keine",
   "build": {
    "de": "ohne Grace",
    "en": "no Grace"
   },
   "unit": "analyse",
   "tasks": null,
   "runsStarted": 0,
   "runsValid": 0,
   "runsValidWithGrace": 0,
   "runsValidClaudeAlone": 0,
   "runsValidOther": 0,
   "countedInTotals": false,
   "status": "calibration",
   "result": {
    "de": "Ein zweiter Claude-allein-Lauf ist nur bei ≈ 73 % gleich oder besser als der erste – das ist die ehrliche Messlatte.",
    "en": "A second Claude-alone run is equal or better than the first only ≈ 73 % of the time – that is the honest yardstick."
   },
   "invalidReason": null,
   "headline": false,
   "caveat": null
  },
  {
   "id": "R39",
   "from": "2026-09-25",
   "to": "2026-09-25",
   "block": "G",
   "title": {
    "de": "Claude allein mit/ohne Spar-Plugins",
    "en": "Claude alone with/without saving plugins"
   },
   "kinds": [
    "bezug",
    "api_kosten",
    "qualitaet"
   ],
   "taskType": "aenderung",
   "version": "keine",
   "build": {
    "de": "ohne Grace",
    "en": "no Grace"
   },
   "unit": "lauf",
   "tasks": 30,
   "runsStarted": 60,
   "runsValid": 60,
   "runsValidWithGrace": 0,
   "runsValidClaudeAlone": 60,
   "runsValidOther": 0,
   "countedInTotals": true,
   "status": "valid",
   "result": {
    "de": "Spar-Plugins: −11 % Kosten; Qualität ohne Plugins besser bei 11, mit besser bei 1, gleich bei 18.",
    "en": "Saving plugins: −11 % cost; quality better without plugins on 11, with on 1, equal on 18."
   },
   "invalidReason": null,
   "headline": false,
   "caveat": null
  },
  {
   "id": "R40",
   "from": "2026-09-25",
   "to": "2026-09-25",
   "block": "E",
   "title": {
    "de": "Pilot Neubau (5 Aufgaben)",
    "en": "Pilot new builds (5 tasks)"
   },
   "kinds": [
    "api_kosten"
   ],
   "taskType": "neubau",
   "version": "alt",
   "build": {
    "de": "Cockpit-Stand 25.09.",
    "en": "cockpit build 25 Sep"
   },
   "unit": "lauf",
   "tasks": 5,
   "runsStarted": 13,
   "runsValid": 5,
   "runsValidWithGrace": 0,
   "runsValidClaudeAlone": 5,
   "runsValidOther": 0,
   "countedInTotals": true,
   "status": "partly_valid",
   "result": {
    "de": "Nur Claude-Bezugswerte (0,98–26,36 $ je Aufgabe).",
    "en": "Claude reference values only ($0.98–26.36 per task)."
   },
   "invalidReason": {
    "de": "8 Grace-Versuche, alle nach genau 5 Minuten technisch abgebrochen (Zeitgrenze einer Bibliothek).",
    "en": "8 Grace attempts, all cut off after exactly 5 minutes (a library timeout)."
   },
   "headline": false,
   "caveat": null
  },
  {
   "id": "R41",
   "from": "2026-09-25",
   "to": "2026-09-25",
   "block": "D",
   "title": {
    "de": "Wie zählt der Cache am Limit?",
    "en": "How does cache count against the limit?"
   },
   "kinds": [
    "eichung",
    "abo_limit"
   ],
   "taskType": "eichung",
   "version": "keine",
   "build": {
    "de": "ohne Grace",
    "en": "no Grace"
   },
   "unit": "keine",
   "tasks": null,
   "runsStarted": 0,
   "runsValid": 0,
   "runsValidWithGrace": 0,
   "runsValidClaudeAlone": 0,
   "runsValidOther": 0,
   "countedInTotals": false,
   "status": "invalid",
   "result": {
    "de": "Kein Ergebnis.",
    "en": "No result."
   },
   "invalidReason": {
    "de": "Gestoppt: das Konto war gleichzeitig anderweitig in Benutzung.",
    "en": "Stopped: the account was in use elsewhere at the same time."
   },
   "headline": false,
   "caveat": null
  },
  {
   "id": "R42",
   "from": "2026-09-25",
   "to": "2026-09-25",
   "block": "D",
   "title": {
    "de": "Prüfregeln auf 88 Endständen",
    "en": "Check rules on 88 end states"
   },
   "kinds": [
    "eichung"
   ],
   "taskType": "eichung",
   "version": "keine",
   "build": {
    "de": "ohne Grace",
    "en": "no Grace"
   },
   "unit": "analyse",
   "tasks": null,
   "runsStarted": 0,
   "runsValid": 0,
   "runsValidWithGrace": 0,
   "runsValidClaudeAlone": 0,
   "runsValidOther": 0,
   "countedInTotals": false,
   "status": "calibration",
   "result": {
    "de": "40 von 48 Niederlagen erkannt, aber 23 von 40 Erfolge fälschlich als offen markiert.",
    "en": "40 of 48 losses flagged, but 23 of 40 successes wrongly flagged as open."
   },
   "invalidReason": null,
   "headline": false,
   "caveat": null
  },
  {
   "id": "R43",
   "from": "2026-09-25",
   "to": "2026-09-25",
   "block": "L",
   "title": {
    "de": "Lernsystem mit/ohne (Grace allein)",
    "en": "Learning system on/off (Grace alone)"
   },
   "kinds": [
    "lernsystem"
   ],
   "taskType": "aenderung",
   "version": "alt",
   "build": {
    "de": "alte Architektur mit Lernsystem (25.09.)",
    "en": "old architecture with learning system (25 Sep)"
   },
   "unit": "lauf",
   "tasks": 8,
   "runsStarted": 18,
   "runsValid": 16,
   "runsValidWithGrace": 16,
   "runsValidClaudeAlone": 0,
   "runsValidOther": 0,
   "countedInTotals": true,
   "status": "valid",
   "result": {
    "de": "E-Mail-Aufgaben: mit Lernspeicher 4 von 4 vollständig, ohne 1 von 4; Text-Aufgaben: 3 von 4 gegen 4 von 4. Sehr kleine Stichprobe.",
    "en": "E-mail tasks: with learning store 4 of 4 complete, without 1 of 4; text tasks: 3 of 4 vs 4 of 4. Very small sample."
   },
   "invalidReason": {
    "de": "2 Erstversuche wegen eines Zählfehlers im Messskript abgebrochen.",
    "en": "2 first attempts aborted because of a counting bug in the test script."
   },
   "headline": false,
   "caveat": null
  },
  {
   "id": "R44",
   "from": "2026-09-25",
   "to": "2026-09-25",
   "block": "C",
   "title": {
    "de": "Prüfregeln + Eskalation an Sonnet",
    "en": "Check rules + escalation to Sonnet"
   },
   "kinds": [
    "api_kosten",
    "qualitaet"
   ],
   "taskType": "aenderung",
   "version": "nachbesserung",
   "build": {
    "de": "Grace-Endstände aus R16 (Stand 23.09.) + Claude als Nachbesserer",
    "en": "Grace end states from R16 (23 Sep build) + Claude as finisher"
   },
   "unit": "lauf",
   "tasks": 8,
   "runsStarted": 13,
   "runsValid": 13,
   "runsValidWithGrace": 13,
   "runsValidClaudeAlone": 0,
   "runsValidOther": 0,
   "countedInTotals": true,
   "status": "valid",
   "result": {
    "de": "43 von 90 (48 %) gleich oder besser bei 32 % der Kosten.",
    "en": "43 of 90 (48 %) equal or better at 32 % of cost."
   },
   "invalidReason": null,
   "headline": false,
   "caveat": null
  },
  {
   "id": "R45",
   "from": "2026-09-25",
   "to": "2026-09-25",
   "block": "L",
   "title": {
    "de": "Lernsystem 2 (eine Aufgabe)",
    "en": "Learning system 2 (one task)"
   },
   "kinds": [
    "lernsystem",
    "api_kosten"
   ],
   "taskType": "aenderung",
   "version": "alt",
   "build": {
    "de": "alte Architektur mit Lernsystem (25.09.)",
    "en": "old architecture with learning system (25 Sep)"
   },
   "unit": "lauf",
   "tasks": 1,
   "runsStarted": 3,
   "runsValid": 3,
   "runsValidWithGrace": 2,
   "runsValidClaudeAlone": 1,
   "runsValidOther": 0,
   "countedInTotals": true,
   "status": "valid",
   "result": {
    "de": "Alle drei vollständig: mit Lernspeicher 0,029 $, ohne 0,036 $, Claude allein 0,159 $ (n = 1).",
    "en": "All three complete: with learning store $0.029, without $0.036, Claude alone $0.159 (n = 1)."
   },
   "invalidReason": null,
   "headline": false,
   "caveat": null
  },
  {
   "id": "R46",
   "from": "2026-09-25",
   "to": "2026-09-25",
   "block": "G",
   "title": {
    "de": "Sonnet allein gegen Opus allein",
    "en": "Sonnet alone vs Opus alone"
   },
   "kinds": [
    "bezug",
    "api_kosten",
    "qualitaet"
   ],
   "taskType": "aenderung",
   "version": "keine",
   "build": {
    "de": "ohne Grace",
    "en": "no Grace"
   },
   "unit": "lauf",
   "tasks": 30,
   "runsStarted": 30,
   "runsValid": 30,
   "runsValidWithGrace": 0,
   "runsValidClaudeAlone": 30,
   "runsValidOther": 0,
   "countedInTotals": true,
   "status": "valid",
   "result": {
    "de": "Sonnet 16 von 30 gleich oder besser bei −32,5 % Kosten.",
    "en": "Sonnet 16 of 30 equal or better at −32.5 % cost."
   },
   "invalidReason": null,
   "headline": false,
   "caveat": null
  },
  {
   "id": "R47",
   "from": "2026-09-25",
   "to": "2026-09-25",
   "block": "D",
   "title": {
    "de": "Einordnung von 937 Fundstellen",
    "en": "Classifying 937 code locations"
   },
   "kinds": [
    "eichung"
   ],
   "taskType": "eichung",
   "version": "keine",
   "build": {
    "de": "ohne Grace",
    "en": "no Grace"
   },
   "unit": "analyse",
   "tasks": null,
   "runsStarted": 0,
   "runsValid": 0,
   "runsValidWithGrace": 0,
   "runsValidClaudeAlone": 0,
   "runsValidOther": 0,
   "countedInTotals": false,
   "status": "calibration",
   "result": {
    "de": "Regeln: 82 % Fehlalarme; Sonnet: 13 %, erkennt 26 von 48 Niederlagen.",
    "en": "Rules: 82 % false alarms; Sonnet: 13 %, catches 26 of 48 losses."
   },
   "invalidReason": null,
   "headline": false,
   "caveat": null
  },
  {
   "id": "R48",
   "from": "2026-09-25",
   "to": "2026-09-25",
   "block": "C",
   "title": {
    "de": "Nachbesserung mit Fundstellenregister",
    "en": "Finisher with location register"
   },
   "kinds": [
    "api_kosten",
    "qualitaet"
   ],
   "taskType": "aenderung",
   "version": "nachbesserung",
   "build": {
    "de": "Grace-Endstände aus R16 (Stand 23.09.) + Claude als Nachbesserer",
    "en": "Grace end states from R16 (23 Sep build) + Claude as finisher"
   },
   "unit": "lauf",
   "tasks": 13,
   "runsStarted": 32,
   "runsValid": 30,
   "runsValidWithGrace": 30,
   "runsValidClaudeAlone": 0,
   "runsValidOther": 0,
   "countedInTotals": true,
   "status": "partly_valid",
   "result": {
    "de": "40 von 90 (44 %) bei 43 % der Kosten; der erwartete Gewinn blieb aus.",
    "en": "40 of 90 (44 %) at 43 % of cost; the expected gain did not appear."
   },
   "invalidReason": {
    "de": "Einzelne Läufe/Urteile ungültig (Richter-Aussetzer oder Abbruch), nicht gezählt.",
    "en": "Single runs/verdicts invalid (judge dropouts or aborts), not counted."
   },
   "headline": false,
   "caveat": null
  },
  {
   "id": "R49",
   "from": "2026-09-25",
   "to": "2026-09-25",
   "block": "L",
   "title": {
    "de": "Grace-allein-Basis auf 30 Änderungen",
    "en": "Grace-alone baseline on 30 changes"
   },
   "kinds": [
    "qualitaet"
   ],
   "taskType": "aenderung",
   "version": "alt",
   "build": {
    "de": "alte Architektur mit Lernsystem (25.09.)",
    "en": "old architecture with learning system (25 Sep)"
   },
   "unit": "lauf",
   "tasks": 2,
   "runsStarted": 3,
   "runsValid": 2,
   "runsValidWithGrace": 2,
   "runsValidClaudeAlone": 0,
   "runsValidOther": 0,
   "countedInTotals": true,
   "status": "incomplete",
   "result": {
    "de": "Kein Ergebniswert.",
    "en": "No result value."
   },
   "invalidReason": {
    "de": "Nach 2 fertigen von 30 Aufgaben pausiert; der dritte Lauf blieb unvollständig.",
    "en": "Paused after 2 finished tasks of 30; the third run stayed incomplete."
   },
   "headline": false,
   "caveat": null
  },
  {
   "id": "R50",
   "from": "2026-09-25",
   "to": "2026-09-26",
   "block": "E",
   "title": {
    "de": "Zurückgelegte Prüfaufgaben: nur Claude-Seite",
    "en": "Held-out tasks: Claude side only"
   },
   "kinds": [
    "bezug"
   ],
   "taskType": "neubau",
   "version": "keine",
   "build": {
    "de": "ohne Grace",
    "en": "no Grace"
   },
   "unit": "lauf",
   "tasks": 29,
   "runsStarted": 38,
   "runsValid": 29,
   "runsValidWithGrace": 0,
   "runsValidClaudeAlone": 29,
   "runsValidOther": 0,
   "countedInTotals": true,
   "status": "partly_valid",
   "result": {
    "de": "29 Claude-allein-Läufe als Bezug; nie gegen Grace gefahren. 6 dieser Aufgaben kamen später im Groß- und Fülltest vor (dort erstmals mit Grace).",
    "en": "29 Claude-alone runs as reference; never run against Grace. 6 of these tasks were later used in the large and fill tests (first time with Grace there)."
   },
   "invalidReason": {
    "de": "Grace-Teil zu früh gestartet und verworfen (8 Läufe); 1 Claude-Lauf abgebrochen.",
    "en": "Grace part started too early and discarded (8 runs); 1 Claude run aborted."
   },
   "headline": false,
   "caveat": null
  },
  {
   "id": "R51",
   "from": "2026-09-25",
   "to": "2026-09-25",
   "block": "E",
   "title": {
    "de": "Übernahme ohne Schreibrecht",
    "en": "Hand-over without write permission"
   },
   "kinds": [
    "qualitaet"
   ],
   "taskType": "neubau",
   "version": "uebernahme",
   "build": {
    "de": "Übernahme-Prototyp (25.–26.09.)",
    "en": "hand-over prototype (25–26 Sep)"
   },
   "unit": "lauf",
   "tasks": 0,
   "runsStarted": 6,
   "runsValid": 0,
   "runsValidWithGrace": 0,
   "runsValidClaudeAlone": 0,
   "runsValidOther": 0,
   "countedInTotals": true,
   "status": "invalid",
   "result": {
    "de": "Kein Ergebnis.",
    "en": "No result."
   },
   "invalidReason": {
    "de": "Claude durfte bei der Übernahme nicht schreiben (Einstellungsfehler); dazu 3 doppelt gestartete Läufe.",
    "en": "Claude was not allowed to write during hand-over (config error); plus 3 runs started twice."
   },
   "headline": false,
   "caveat": null
  },
  {
   "id": "R52",
   "from": "2026-09-25",
   "to": "2026-09-25",
   "block": "E",
   "title": {
    "de": "Übernahme durch Sonnet (4 Aufgaben)",
    "en": "Hand-over to Sonnet (4 tasks)"
   },
   "kinds": [
    "api_kosten",
    "qualitaet"
   ],
   "taskType": "neubau",
   "version": "uebernahme",
   "build": {
    "de": "Übernahme-Prototyp (25.–26.09.)",
    "en": "hand-over prototype (25–26 Sep)"
   },
   "unit": "lauf",
   "tasks": 4,
   "runsStarted": 8,
   "runsValid": 8,
   "runsValidWithGrace": 8,
   "runsValidClaudeAlone": 0,
   "runsValidOther": 0,
   "countedInTotals": true,
   "status": "valid",
   "result": {
    "de": "Claude allein besser bei 4 von 4; Grace-Weg −57 % Kosten (Claude-allein-Läufe aus R53 als Bezug).",
    "en": "Claude alone better on 4 of 4; Grace path −57 % cost (Claude-alone runs from R53 as reference)."
   },
   "invalidReason": null,
   "headline": false,
   "caveat": null
  },
  {
   "id": "R53",
   "from": "2026-09-25",
   "to": "2026-09-26",
   "block": "E",
   "title": {
    "de": "Übernahme: drei Varianten (10 Aufgaben)",
    "en": "Hand-over: three variants (10 tasks)"
   },
   "kinds": [
    "api_kosten",
    "qualitaet"
   ],
   "taskType": "neubau",
   "version": "uebernahme",
   "build": {
    "de": "Übernahme-Prototyp (25.–26.09.)",
    "en": "hand-over prototype (25–26 Sep)"
   },
   "unit": "lauf",
   "tasks": 10,
   "runsStarted": 50,
   "runsValid": 50,
   "runsValidWithGrace": 40,
   "runsValidClaudeAlone": 10,
   "runsValidOther": 0,
   "countedInTotals": true,
   "status": "valid",
   "result": {
    "de": "Abnahmequote: Grace + Sonnet 69 % (−60 % Kosten), Grace + Opus 90 % (−2 %) gegen Opus allein 90 %.",
    "en": "Acceptance rate: Grace + Sonnet 69 % (−60 % cost), Grace + Opus 90 % (−2 %) vs Opus alone 90 %."
   },
   "invalidReason": null,
   "headline": false,
   "caveat": null
  },
  {
   "id": "R54",
   "from": "2026-09-25",
   "to": "2026-09-26",
   "block": "D",
   "title": {
    "de": "Opus gegen Opus (10 Neubauten)",
    "en": "Opus vs Opus (10 new builds)"
   },
   "kinds": [
    "eichung",
    "streuung",
    "qualitaet"
   ],
   "taskType": "neubau",
   "version": "keine",
   "build": {
    "de": "ohne Grace",
    "en": "no Grace"
   },
   "unit": "lauf",
   "tasks": 10,
   "runsStarted": 10,
   "runsValid": 10,
   "runsValidWithGrace": 0,
   "runsValidClaudeAlone": 10,
   "runsValidOther": 0,
   "countedInTotals": true,
   "status": "valid",
   "result": {
    "de": "Zwei gleiche Läufe: 89,6 % gegen 90,0 % Abnahmequote; der Paarrichter findet trotzdem fast immer einen „Sieger“.",
    "en": "Two identical runs: 89.6 % vs 90.0 % acceptance rate; the pair judge still almost always finds a \"winner\"."
   },
   "invalidReason": null,
   "headline": false,
   "caveat": null
  },
  {
   "id": "R55",
   "from": "2026-09-26",
   "to": "2026-09-26",
   "block": "E",
   "title": {
    "de": "Grace mit vorab bekannter Abnahmeliste",
    "en": "Grace with acceptance list known in advance"
   },
   "kinds": [
    "api_kosten",
    "qualitaet"
   ],
   "taskType": "neubau",
   "version": "uebernahme",
   "build": {
    "de": "Übernahme-Prototyp (25.–26.09.)",
    "en": "hand-over prototype (25–26 Sep)"
   },
   "unit": "lauf",
   "tasks": 6,
   "runsStarted": 20,
   "runsValid": 12,
   "runsValidWithGrace": 12,
   "runsValidClaudeAlone": 0,
   "runsValidOther": 0,
   "countedInTotals": true,
   "status": "partly_valid",
   "result": {
    "de": "Nur Obergrenze (die Liste war bekannt): 97–99 % bei −36 bis −38 % Kosten auf 6 Aufgaben.",
    "en": "Upper bound only (list was known): 97–99 % at −36 to −38 % cost on 6 tasks."
   },
   "invalidReason": {
    "de": "8 Läufe unbewertet.",
    "en": "8 runs unscored."
   },
   "headline": false,
   "caveat": null
  },
  {
   "id": "R56",
   "from": "2026-09-26",
   "to": "2026-09-26",
   "block": "G",
   "title": {
    "de": "Claude allein in 7 Einstellungen",
    "en": "Claude alone in 7 settings"
   },
   "kinds": [
    "bezug",
    "api_kosten",
    "qualitaet"
   ],
   "taskType": "neubau",
   "version": "keine",
   "build": {
    "de": "ohne Grace",
    "en": "no Grace"
   },
   "unit": "lauf",
   "tasks": 10,
   "runsStarted": 61,
   "runsValid": 59,
   "runsValidWithGrace": 0,
   "runsValidClaudeAlone": 59,
   "runsValidOther": 0,
   "countedInTotals": true,
   "status": "valid",
   "result": {
    "de": "Abnahmequote 77–99 % je nach Modell und Denkstufe; Opus 5.5 low 94 % bei −58 % Kosten gegen Standard.",
    "en": "Acceptance rate 77–99 % depending on model and effort; Opus 5.5 low 94 % at −58 % cost vs default."
   },
   "invalidReason": {
    "de": "2 Läufe ungültig.",
    "en": "2 runs invalid."
   },
   "headline": false,
   "caveat": null
  },
  {
   "id": "R57",
   "from": "2026-09-26",
   "to": "2026-09-26",
   "block": "E",
   "title": {
    "de": "Strang: Arbeitsliste → Grace → Lückenfüller",
    "en": "Chain: work list → Grace → gap filler"
   },
   "kinds": [
    "api_kosten",
    "qualitaet"
   ],
   "taskType": "neubau",
   "version": "uebernahme",
   "build": {
    "de": "Strang-/Paket-Prototyp (26.09.)",
    "en": "chain/package prototype (26 Sep)"
   },
   "unit": "lauf",
   "tasks": 10,
   "runsStarted": 61,
   "runsValid": 59,
   "runsValidWithGrace": 59,
   "runsValidClaudeAlone": 0,
   "runsValidOther": 0,
   "countedInTotals": true,
   "status": "valid",
   "result": {
    "de": "Bestwert 94 % bei −46 % gegen Opus Standard – die Ersparnis kam von der niedrigeren Denkstufe, nicht von Grace (die allein: −58 %).",
    "en": "Best 94 % at −46 % vs Opus default – the saving came from the lower effort setting, not from Grace (that alone: −58 %)."
   },
   "invalidReason": {
    "de": "2 Läufe abgebrochen.",
    "en": "2 runs aborted."
   },
   "headline": false,
   "caveat": null
  },
  {
   "id": "R58",
   "from": "2026-09-26",
   "to": "2026-09-26",
   "block": "E",
   "title": {
    "de": "Graces Paket → Übernehmer",
    "en": "Grace package → finisher"
   },
   "kinds": [
    "api_kosten",
    "qualitaet"
   ],
   "taskType": "neubau",
   "version": "uebernahme",
   "build": {
    "de": "Strang-/Paket-Prototyp (26.09.)",
    "en": "chain/package prototype (26 Sep)"
   },
   "unit": "lauf",
   "tasks": 10,
   "runsStarted": 37,
   "runsValid": 36,
   "runsValidWithGrace": 36,
   "runsValidClaudeAlone": 0,
   "runsValidOther": 0,
   "countedInTotals": true,
   "status": "valid",
   "result": {
    "de": "95 % bei −65 % gegen Opus Standard; gegen dieselbe Denkstufe allein kaum Ersparnis.",
    "en": "95 % at −65 % vs Opus default; hardly any saving vs the same effort setting alone."
   },
   "invalidReason": {
    "de": "1 Lauf unbewertet.",
    "en": "1 run unscored."
   },
   "headline": false,
   "caveat": null
  },
  {
   "id": "R59",
   "from": "2026-09-26",
   "to": "2026-09-26",
   "block": "E",
   "title": {
    "de": "Grace schreibt Liste, Übernehmer füllt Lücken",
    "en": "Grace writes list, finisher fills gaps"
   },
   "kinds": [
    "api_kosten",
    "qualitaet"
   ],
   "taskType": "neubau",
   "version": "uebernahme",
   "build": {
    "de": "Strang-/Paket-Prototyp (26.09.)",
    "en": "chain/package prototype (26 Sep)"
   },
   "unit": "lauf",
   "tasks": 3,
   "runsStarted": 11,
   "runsValid": 10,
   "runsValidWithGrace": 10,
   "runsValidClaudeAlone": 0,
   "runsValidOther": 0,
   "countedInTotals": true,
   "status": "valid",
   "result": {
    "de": "Rund doppelt so teuer wie Claude allein – verworfen.",
    "en": "About twice the cost of Claude alone – dropped."
   },
   "invalidReason": {
    "de": "1 Lauf abgebrochen.",
    "en": "1 run aborted."
   },
   "headline": false,
   "caveat": null
  },
  {
   "id": "R60",
   "from": "2026-09-26",
   "to": "2026-09-26",
   "block": "D",
   "title": {
    "de": "Arbeitslisten je Modell",
    "en": "Work lists per model"
   },
   "kinds": [
    "eichung"
   ],
   "taskType": "neubau",
   "version": "keine",
   "build": {
    "de": "ohne Grace",
    "en": "no Grace"
   },
   "unit": "eichruf",
   "tasks": null,
   "runsStarted": 91,
   "runsValid": 91,
   "runsValidWithGrace": 0,
   "runsValidClaudeAlone": 0,
   "runsValidOther": 91,
   "countedInTotals": true,
   "status": "calibration",
   "result": {
    "de": "Abdeckung 80–96 % je nach Modell.",
    "en": "Coverage 80–96 % depending on model."
   },
   "invalidReason": null,
   "headline": false,
   "caveat": null
  },
  {
   "id": "R61",
   "from": "2026-09-26",
   "to": "2026-09-26",
   "block": "D",
   "title": {
    "de": "Prüft die Abnahmeliste, was sie soll?",
    "en": "Does the acceptance list measure what it should?"
   },
   "kinds": [
    "eichung",
    "qualitaet"
   ],
   "taskType": "eichung",
   "version": "keine",
   "build": {
    "de": "ohne Grace",
    "en": "no Grace"
   },
   "unit": "analyse",
   "tasks": null,
   "runsStarted": 0,
   "runsValid": 0,
   "runsValidWithGrace": 0,
   "runsValidClaudeAlone": 0,
   "runsValidOther": 0,
   "countedInTotals": false,
   "status": "calibration",
   "result": {
    "de": "192 Endstände bewertet; 11 von 114 Kriterien kippen schon zwischen zwei gleichen Läufen.",
    "en": "192 end states judged; 11 of 114 criteria flip even between two identical runs."
   },
   "invalidReason": null,
   "headline": false,
   "caveat": null
  },
  {
   "id": "R62",
   "from": "2026-09-26",
   "to": "2026-09-26",
   "block": "F",
   "title": {
    "de": "Werkzeug-Probe",
    "en": "Tool probe"
   },
   "kinds": [
    "api_kosten"
   ],
   "taskType": "neubau",
   "version": "zwischenstand",
   "build": {
    "de": "Zwischenstand als Werkzeug (26.09.)",
    "en": "interim version as a tool (26 Sep)"
   },
   "unit": "lauf",
   "tasks": 3,
   "runsStarted": 4,
   "runsValid": 3,
   "runsValidWithGrace": 3,
   "runsValidClaudeAlone": 0,
   "runsValidOther": 0,
   "countedInTotals": true,
   "status": "valid",
   "result": {
    "de": "Kein Unterschied: Claude gab keinen Code an Grace ab.",
    "en": "No difference: Claude handed no code to Grace."
   },
   "invalidReason": {
    "de": "1 verworfener Erstlauf.",
    "en": "1 discarded first run."
   },
   "headline": false,
   "caveat": null
  },
  {
   "id": "R63",
   "from": "2026-09-26",
   "to": "2026-09-26",
   "block": "F",
   "title": {
    "de": "Schwelle je Teilaufgabe gegen Opus",
    "en": "Threshold per sub-task vs Opus"
   },
   "kinds": [
    "qualitaet"
   ],
   "taskType": "teilaufgabe",
   "version": "zwischenstand",
   "build": {
    "de": "Zwischenstand (26.09.)",
    "en": "interim version (26 Sep)"
   },
   "unit": "lauf",
   "tasks": 15,
   "runsStarted": 30,
   "runsValid": 30,
   "runsValidWithGrace": 15,
   "runsValidClaudeAlone": 15,
   "runsValidOther": 0,
   "countedInTotals": true,
   "status": "valid",
   "result": {
    "de": "Bis Schwierigkeitsstufe 3 fast gleich gut (97–99 % gegen 99–100 %), auf Stufe 5 deutlich schwächer (61 % gegen 100 %).",
    "en": "Up to difficulty level 3 almost as good (97–99 % vs 99–100 %), at level 5 clearly weaker (61 % vs 100 %)."
   },
   "invalidReason": null,
   "headline": false,
   "caveat": null
  },
  {
   "id": "R64",
   "from": "2026-09-26",
   "to": "2026-09-26",
   "block": "F",
   "title": {
    "de": "Ganze Änderungen: allein / Variante / Filter / zuerst",
    "en": "Whole changes: alone / variant / filter / first"
   },
   "kinds": [
    "api_kosten",
    "qualitaet"
   ],
   "taskType": "aenderung",
   "version": "zwischenstand",
   "build": {
    "de": "Zwischenstand, Schwelle 3 (26.09.)",
    "en": "interim version, threshold 3 (26 Sep)"
   },
   "unit": "lauf",
   "tasks": 5,
   "runsStarted": 20,
   "runsValid": 20,
   "runsValidWithGrace": 15,
   "runsValidClaudeAlone": 5,
   "runsValidOther": 0,
   "countedInTotals": true,
   "status": "valid",
   "result": {
    "de": "Keine Ersparnis (5,85–6,31 $ gegen 6,03 $ allein).",
    "en": "No saving ($5.85–6.31 vs $6.03 alone)."
   },
   "invalidReason": null,
   "headline": false,
   "caveat": null
  },
  {
   "id": "R65",
   "from": "2026-09-26",
   "to": "2026-09-26",
   "block": "F",
   "title": {
    "de": "Änderungs-Zwischenstand v1 und v2",
    "en": "Change version v1 and v2"
   },
   "kinds": [
    "api_kosten",
    "qualitaet"
   ],
   "taskType": "aenderung",
   "version": "zwischenstand",
   "build": {
    "de": "Änderungs-Zwischenstand v1 und v2 (26.09.)",
    "en": "change version v1 and v2 (26 Sep)"
   },
   "unit": "lauf",
   "tasks": 5,
   "runsStarted": 11,
   "runsValid": 11,
   "runsValidWithGrace": 11,
   "runsValidClaudeAlone": 0,
   "runsValidOther": 0,
   "countedInTotals": true,
   "status": "valid",
   "result": {
    "de": "v1 +3 % Kosten, v2 −20 % bei 97,2 % gegen 97,5 % Qualität.",
    "en": "v1 +3 % cost, v2 −20 % at 97.2 % vs 97.5 % quality."
   },
   "invalidReason": null,
   "headline": false,
   "caveat": null
  },
  {
   "id": "R66",
   "from": "2026-09-26",
   "to": "2026-09-26",
   "block": "F",
   "title": {
    "de": "Neubau-Zwischenstand, erste Probe",
    "en": "New-build version, first probe"
   },
   "kinds": [
    "api_kosten",
    "qualitaet"
   ],
   "taskType": "neubau",
   "version": "zwischenstand",
   "build": {
    "de": "Neubau-Zwischenstand v1 (26.09.)",
    "en": "new-build version v1 (26 Sep)"
   },
   "unit": "lauf",
   "tasks": 6,
   "runsStarted": 12,
   "runsValid": 10,
   "runsValidWithGrace": 4,
   "runsValidClaudeAlone": 6,
   "runsValidOther": 0,
   "countedInTotals": true,
   "status": "partly_valid",
   "result": {
    "de": "−38 % Kosten, aber Qualität 84 % gegen 93 %.",
    "en": "−38 % cost, but quality 84 % vs 93 %."
   },
   "invalidReason": {
    "de": "2 Grace-Läufe ohne Ergebnis.",
    "en": "2 Grace runs without result."
   },
   "headline": false,
   "caveat": null
  },
  {
   "id": "R67",
   "from": "2026-09-26",
   "to": "2026-09-27",
   "block": "F",
   "title": {
    "de": "Zwei Konten parallel (4 Aufgaben)",
    "en": "Two accounts in parallel (4 tasks)"
   },
   "kinds": [
    "api_kosten",
    "qualitaet",
    "abo_limit"
   ],
   "taskType": "neubau",
   "version": "zwischenstand",
   "build": {
    "de": "Neubau-Zwischenstand v1 (26.09.)",
    "en": "new-build version v1 (26 Sep)"
   },
   "unit": "lauf",
   "tasks": 4,
   "runsStarted": 9,
   "runsValid": 9,
   "runsValidWithGrace": 4,
   "runsValidClaudeAlone": 5,
   "runsValidOther": 0,
   "countedInTotals": true,
   "status": "valid",
   "result": {
    "de": "−19 % Kosten, Qualität 99 % gegen 98 %; 5-Stunden-Limit: Grace +5, allein +2 Punkte – hier verbrauchte Grace MEHR (Ablesung grob).",
    "en": "−19 % cost, quality 99 % vs 98 %; 5-hour limit: Grace +5, alone +2 points – Grace used MORE here (coarse reading)."
   },
   "invalidReason": {
    "de": "1 zusätzlicher Claude-Lauf gemessen, aber ohne Qualitätsbewertung.",
    "en": "1 extra Claude run measured but not quality-judged."
   },
   "headline": false,
   "caveat": null
  },
  {
   "id": "R68",
   "from": "2026-09-27",
   "to": "2026-09-27",
   "block": "F",
   "title": {
    "de": "Limit-Test v3 (15 Aufgaben)",
    "en": "Limit test v3 (15 tasks)"
   },
   "kinds": [
    "abo_limit",
    "api_kosten",
    "qualitaet"
   ],
   "taskType": "neubau",
   "version": "zwischenstand",
   "build": {
    "de": "Zwischenstand v3 (27.09.)",
    "en": "interim version v3 (27 Sep)"
   },
   "unit": "lauf",
   "tasks": 15,
   "runsStarted": 30,
   "runsValid": 28,
   "runsValidWithGrace": 13,
   "runsValidClaudeAlone": 15,
   "runsValidOther": 0,
   "countedInTotals": true,
   "status": "partly_valid",
   "result": {
    "de": "−4 % Kosten, aber Claude schrieb mehr als allein; 5-Stunden-Limit: Grace +16, allein +12 Punkte – Grace verbrauchte MEHR. Qualität auf 13 sauberen Paaren 85 % gegen 84 %; zählt man den Fehllauf als Niederlage, 79 % gegen 85 %.",
    "en": "−4 % cost, but Claude wrote more than alone; 5-hour limit: Grace +16, alone +12 points – Grace used MORE. Quality on 13 clean pairs 85 % vs 84 %; counting the failed run as a loss, 79 % vs 85 %."
   },
   "invalidReason": {
    "de": "1 Grace-Lauf lieferte nach 3 Minuten nichts (Fehllauf), 1 Grace-Lauf ohne Bewertung.",
    "en": "1 Grace run delivered nothing after 3 minutes (failed run), 1 Grace run not judged."
   },
   "headline": false,
   "caveat": null
  },
  {
   "id": "R69",
   "from": "2026-09-27",
   "to": "2026-09-27",
   "block": "F",
   "title": {
    "de": "Limit-Test v4 (7 Aufgaben)",
    "en": "Limit test v4 (7 tasks)"
   },
   "kinds": [
    "abo_limit",
    "api_kosten",
    "qualitaet"
   ],
   "taskType": "neubau",
   "version": "zwischenstand",
   "build": {
    "de": "Zwischenstand v4 (27.09.)",
    "en": "interim version v4 (27 Sep)"
   },
   "unit": "lauf",
   "tasks": 7,
   "runsStarted": 7,
   "runsValid": 7,
   "runsValidWithGrace": 7,
   "runsValidClaudeAlone": 0,
   "runsValidOther": 0,
   "countedInTotals": true,
   "status": "valid",
   "result": {
    "de": "−16 % Kosten, Claude schrieb etwas mehr als allein, Qualität 92,7 % gegen 92,6 % (Claude-allein-Bezug aus R68).",
    "en": "−16 % cost, Claude wrote a little more than alone, quality 92.7 % vs 92.6 % (Claude-alone reference from R68)."
   },
   "invalidReason": null,
   "headline": false,
   "caveat": null
  },
  {
   "id": "R70",
   "from": "2026-09-27",
   "to": "2026-09-27",
   "block": "F",
   "title": {
    "de": "Limit-Test v5 mit Lernsystem",
    "en": "Limit test v5 with learning system"
   },
   "kinds": [
    "abo_limit",
    "api_kosten",
    "qualitaet",
    "lernsystem"
   ],
   "taskType": "neubau",
   "version": "zwischenstand",
   "build": {
    "de": "Zwischenstand v5: Schritt-Modus + Lernsystem (27.09.)",
    "en": "interim version v5: step mode + learning system (27 Sep)"
   },
   "unit": "lauf",
   "tasks": 7,
   "runsStarted": 7,
   "runsValid": 7,
   "runsValidWithGrace": 7,
   "runsValidClaudeAlone": 0,
   "runsValidOther": 0,
   "countedInTotals": true,
   "status": "valid",
   "result": {
    "de": "−29 % Kosten, Qualität 90,4 % gegen 92,6 % (3 schlechter, 3 gleich, 1 besser).",
    "en": "−29 % cost, quality 90.4 % vs 92.6 % (3 worse, 3 equal, 1 better)."
   },
   "invalidReason": null,
   "headline": false,
   "caveat": null
  },
  {
   "id": "R71",
   "from": "2026-09-27",
   "to": "2026-09-27",
   "block": "F",
   "title": {
    "de": "Limit-Test v6",
    "en": "Limit test v6"
   },
   "kinds": [
    "abo_limit",
    "api_kosten"
   ],
   "taskType": "neubau",
   "version": "zwischenstand",
   "build": {
    "de": "Zwischenstand v5.1 (27.09.)",
    "en": "interim version v5.1 (27 Sep)"
   },
   "unit": "lauf",
   "tasks": 10,
   "runsStarted": 13,
   "runsValid": 13,
   "runsValidWithGrace": 3,
   "runsValidClaudeAlone": 10,
   "runsValidOther": 0,
   "countedInTotals": true,
   "status": "incomplete",
   "result": {
    "de": "Gestoppt nach 3 fertigen Paaren; nicht bewertet.",
    "en": "Stopped after 3 finished pairs; not judged."
   },
   "invalidReason": {
    "de": "Reihe abgebrochen.",
    "en": "Series stopped."
   },
   "headline": false,
   "caveat": null
  },
  {
   "id": "R72",
   "from": "2026-09-27",
   "to": "2026-09-27",
   "block": "F",
   "title": {
    "de": "Anderer Coding-Agent als Basis",
    "en": "Different coding agent as base"
   },
   "kinds": [
    "api_kosten"
   ],
   "taskType": "neubau",
   "version": "zwischenstand",
   "build": {
    "de": "Zwischenstand v5.1 mit anderem Agenten (27.09.)",
    "en": "interim version v5.1 with a different agent (27 Sep)"
   },
   "unit": "keine",
   "tasks": null,
   "runsStarted": 0,
   "runsValid": 0,
   "runsValidWithGrace": 0,
   "runsValidClaudeAlone": 0,
   "runsValidOther": 0,
   "countedInTotals": false,
   "status": "not_run",
   "result": {
    "de": "Nicht durchgeführt; eine Mini-Probe ohne Rohdatei wird nicht gewertet.",
    "en": "Not run; a mini probe without raw file is not counted."
   },
   "invalidReason": {
    "de": "Hauptlauf nie gestartet.",
    "en": "Main run never started."
   },
   "headline": false,
   "caveat": null
  },
  {
   "id": "R73",
   "from": "2026-09-27",
   "to": "2026-09-27",
   "block": "G",
   "title": {
    "de": "Kommerzielles Konkurrenz-Plugin: Neubau-Probe",
    "en": "Commercial competitor plugin: new-build probe"
   },
   "kinds": [
    "konkurrenz"
   ],
   "taskType": "neubau",
   "version": "keine",
   "build": {
    "de": "ohne Grace",
    "en": "no Grace"
   },
   "unit": "lauf",
   "tasks": 5,
   "runsStarted": 14,
   "runsValid": 6,
   "runsValidWithGrace": 0,
   "runsValidClaudeAlone": 6,
   "runsValidOther": 0,
   "countedInTotals": true,
   "status": "invalid",
   "result": {
    "de": "Kein verwertbarer Vergleich.",
    "en": "No usable comparison."
   },
   "invalidReason": {
    "de": "Konkurrenz-Arm durch eine Einstellung blockiert; nur 6 Claude-Läufe gültig.",
    "en": "Competitor arm blocked by a setting; only 6 Claude runs valid."
   },
   "headline": false,
   "caveat": null
  },
  {
   "id": "R74",
   "from": "2026-09-27",
   "to": "2026-09-27",
   "block": "G",
   "title": {
    "de": "Kommerzielles Konkurrenz-Plugin: 15 Änderungen",
    "en": "Commercial competitor plugin: 15 changes"
   },
   "kinds": [
    "konkurrenz",
    "api_kosten",
    "qualitaet"
   ],
   "taskType": "aenderung",
   "version": "keine",
   "build": {
    "de": "ohne Grace",
    "en": "no Grace"
   },
   "unit": "lauf",
   "tasks": 15,
   "runsStarted": 42,
   "runsValid": 30,
   "runsValidWithGrace": 0,
   "runsValidClaudeAlone": 15,
   "runsValidOther": 15,
   "countedInTotals": true,
   "status": "partly_valid",
   "result": {
    "de": "Fremdes Plugin gegen Claude + Spar-Plugins: −17 % Kosten, gleiche Qualität. Die Claude-Läufe sind der Bezug für R79–R82.",
    "en": "Third-party plugin vs Claude + saving plugins: −17 % cost, equal quality. The Claude runs are the reference for R79–R82."
   },
   "invalidReason": {
    "de": "12 Läufe mit gesperrten Werkzeugen aussortiert.",
    "en": "12 runs with blocked tools discarded."
   },
   "headline": false,
   "caveat": null
  },
  {
   "id": "R75",
   "from": "2026-09-27",
   "to": "2026-09-28",
   "block": "G",
   "title": {
    "de": "Kommerzielles Konkurrenz-Plugin: 28 Neubauten",
    "en": "Commercial competitor plugin: 28 new builds"
   },
   "kinds": [
    "konkurrenz",
    "bezug"
   ],
   "taskType": "neubau",
   "version": "keine",
   "build": {
    "de": "ohne Grace",
    "en": "no Grace"
   },
   "unit": "lauf",
   "tasks": 28,
   "runsStarted": 53,
   "runsValid": 32,
   "runsValidWithGrace": 0,
   "runsValidClaudeAlone": 28,
   "runsValidOther": 4,
   "countedInTotals": true,
   "status": "partly_valid",
   "result": {
    "de": "Nur 4 Paare gültig (gemischt); 28 Claude-Läufe dienen als Bezug. Nebenbefund: 5-Stunden-Limit ≈ 1,1 Punkte je 100 000 Ausgabe-Token.",
    "en": "Only 4 pairs valid (mixed); 28 Claude runs serve as reference. Side finding: 5-hour limit ≈ 1.1 points per 100,000 output tokens."
   },
   "invalidReason": {
    "de": "Gratis-Kontingent des Fremdwerkzeugs mitten im Lauf erschöpft.",
    "en": "Free quota of the third-party tool ran out mid-run."
   },
   "headline": false,
   "caveat": null
  },
  {
   "id": "R76",
   "from": "2026-09-28",
   "to": "2026-09-28",
   "block": "H",
   "title": {
    "de": "Abo-Vortest (Plugin und Neubau)",
    "en": "Subscription pre-test (plugin and new builds)"
   },
   "kinds": [
    "api_kosten",
    "qualitaet",
    "tempo"
   ],
   "taskType": "gemischt",
   "version": "umbau",
   "build": {
    "de": "Abo-Modus, erster Stand (28.09.)",
    "en": "subscription mode, first build (28 Sep)"
   },
   "unit": "lauf",
   "tasks": 5,
   "runsStarted": 10,
   "runsValid": 10,
   "runsValidWithGrace": 10,
   "runsValidClaudeAlone": 0,
   "runsValidOther": 0,
   "countedInTotals": true,
   "status": "valid",
   "result": {
    "de": "Änderungen (3): −30 % Kosten, Tests grün. Neubau (2): −68 % Kosten, Qualität 100/91 % gegen 100/100 %, aber eine Aufgabe 66 statt 5 Minuten.",
    "en": "Changes (3): −30 % cost, tests green. New builds (2): −68 % cost, quality 100/91 % vs 100/100 %, but one task 66 instead of 5 minutes."
   },
   "invalidReason": null,
   "headline": false,
   "caveat": null
  },
  {
   "id": "R77",
   "from": "2026-09-28",
   "to": "2026-09-28",
   "block": "H",
   "title": {
    "de": "Tempotest einer Entwicklungsvariante (ohne Claude)",
    "en": "Speed test of a development variant (no Claude)"
   },
   "kinds": [
    "tempo",
    "qualitaet"
   ],
   "taskType": "neubau",
   "version": "umbau",
   "build": {
    "de": "Tempo-Umbau 1 und 2 (28.09.)",
    "en": "speed rebuild 1 and 2 (28 Sep)"
   },
   "unit": "lauf",
   "tasks": 2,
   "runsStarted": 11,
   "runsValid": 11,
   "runsValidWithGrace": 11,
   "runsValidClaudeAlone": 0,
   "runsValidOther": 0,
   "countedInTotals": true,
   "status": "valid",
   "result": {
    "de": "~18 statt 63 Minuten; schnellere Varianten machten deutlich mehr Fehler.",
    "en": "~18 instead of 63 minutes; faster variants made clearly more errors."
   },
   "invalidReason": null,
   "headline": false,
   "caveat": null
  },
  {
   "id": "R78",
   "from": "2026-09-28",
   "to": "2026-09-28",
   "block": "H",
   "title": {
    "de": "Abo-Kurztest nach Tempo-Umbau",
    "en": "Subscription short test after speed rebuild"
   },
   "kinds": [
    "api_kosten",
    "qualitaet",
    "tempo"
   ],
   "taskType": "neubau",
   "version": "umbau",
   "build": {
    "de": "Tempo-Umbau 3 (28.09.)",
    "en": "speed rebuild 3 (28 Sep)"
   },
   "unit": "lauf",
   "tasks": 2,
   "runsStarted": 2,
   "runsValid": 2,
   "runsValidWithGrace": 2,
   "runsValidClaudeAlone": 0,
   "runsValidOther": 0,
   "countedInTotals": true,
   "status": "valid",
   "result": {
    "de": "Kosten −42 % / −72 %, Qualität 96 % / 91 % gegen 100 % / 100 %, Dauer 18 / 24 gegen 5 / 6 Minuten.",
    "en": "Cost −42 % / −72 %, quality 96 % / 91 % vs 100 % / 100 %, time 18 / 24 vs 5 / 6 minutes."
   },
   "invalidReason": null,
   "headline": false,
   "caveat": null
  },
  {
   "id": "R79",
   "from": "2026-09-28",
   "to": "2026-09-28",
   "block": "H",
   "title": {
    "de": "Plugin auf Änderungsaufgaben (Abo)",
    "en": "Plugin on change tasks (subscription)"
   },
   "kinds": [
    "api_kosten",
    "qualitaet",
    "tempo"
   ],
   "taskType": "aenderung",
   "version": "umbau",
   "build": {
    "de": "Abo-Stand 2 (28.09.)",
    "en": "subscription build 2 (28 Sep)"
   },
   "unit": "lauf",
   "tasks": 8,
   "runsStarted": 8,
   "runsValid": 8,
   "runsValidWithGrace": 8,
   "runsValidClaudeAlone": 0,
   "runsValidOther": 0,
   "countedInTotals": true,
   "status": "partly_valid",
   "result": {
    "de": "Auf den 6 bewerteten Aufgaben: −25 % Kosten, Claude schrieb weniger als allein, gleiche Qualität (6 von 6), 2,4× schneller. Über alle 8 Läufe −31 % Kosten.",
    "en": "On the 6 judged tasks: −25 % cost, Claude wrote less than alone, equal quality (6 of 6), 2.4× faster. Across all 8 runs −31 % cost."
   },
   "invalidReason": {
    "de": "Stufe „mittel“ nach 2 Aufgaben abgebrochen; diese 2 Läufe sind gemessen, aber nicht bewertet.",
    "en": "\"Medium\" level stopped after 2 tasks; those 2 runs are measured but not judged."
   },
   "headline": false,
   "caveat": null
  },
  {
   "id": "R80",
   "from": "2026-09-28",
   "to": "2026-09-28",
   "block": "H",
   "title": {
    "de": "API: Änderungs-Zwischenstand voll",
    "en": "API: change version, full"
   },
   "kinds": [
    "api_kosten",
    "qualitaet"
   ],
   "taskType": "aenderung",
   "version": "umbau",
   "build": {
    "de": "API-Stand 1 (28.09.)",
    "en": "API build 1 (28 Sep)"
   },
   "unit": "lauf",
   "tasks": 6,
   "runsStarted": 6,
   "runsValid": 6,
   "runsValidWithGrace": 6,
   "runsValidClaudeAlone": 0,
   "runsValidOther": 0,
   "countedInTotals": true,
   "status": "valid",
   "result": {
    "de": "−12 % (1-Stunden-Cache) / −19 % (5-Minuten-Cache), Qualität 97–100 %.",
    "en": "−12 % (1-hour cache) / −19 % (5-minute cache), quality 97–100 %."
   },
   "invalidReason": null,
   "headline": false,
   "caveat": null
  },
  {
   "id": "R81",
   "from": "2026-09-28",
   "to": "2026-09-28",
   "block": "H",
   "title": {
    "de": "API: schlanker Weg v1",
    "en": "API: slim path v1"
   },
   "kinds": [
    "api_kosten",
    "qualitaet"
   ],
   "taskType": "aenderung",
   "version": "umbau",
   "build": {
    "de": "API-Stand 2 (28.09.)",
    "en": "API build 2 (28 Sep)"
   },
   "unit": "lauf",
   "tasks": 6,
   "runsStarted": 6,
   "runsValid": 6,
   "runsValidWithGrace": 6,
   "runsValidClaudeAlone": 0,
   "runsValidOther": 0,
   "countedInTotals": true,
   "status": "valid",
   "result": {
    "de": "−45 % Kosten, aber eine Aufgabe nur 33 % Qualität (Claude gab bei einem Syntaxfehler auf) → in v2 behoben.",
    "en": "−45 % cost, but one task only 33 % quality (Claude gave up on a syntax error) → fixed in v2."
   },
   "invalidReason": null,
   "headline": false,
   "caveat": null
  },
  {
   "id": "R82",
   "from": "2026-09-28",
   "to": "2026-09-28",
   "block": "H",
   "title": {
    "de": "API: schlanker Weg v2 (Zahl auf der Website)",
    "en": "API: slim path v2 (figure on the website)"
   },
   "kinds": [
    "api_kosten",
    "qualitaet",
    "tempo"
   ],
   "taskType": "aenderung",
   "version": "umbau",
   "build": {
    "de": "API-Stand 3, Zwischenstand vor dem Release (28.09.)",
    "en": "API build 3, interim build before release (28 Sep)"
   },
   "unit": "lauf",
   "tasks": 6,
   "runsStarted": 6,
   "runsValid": 6,
   "runsValidWithGrace": 6,
   "runsValidClaudeAlone": 0,
   "runsValidOther": 0,
   "countedInTotals": true,
   "status": "valid",
   "result": {
    "de": "−48 % Kosten (5-Minuten-Cache; −46 % mit 1-Stunden-Cache), Qualität 95–100 % gegen 97–100 %, 2,7× schneller.",
    "en": "−48 % cost (5-minute cache; −46 % with 1-hour cache), quality 95–100 % vs 97–100 %, 2.7× faster."
   },
   "invalidReason": null,
   "headline": true,
   "caveat": {
    "de": "Zwischenstand vom 28.09., nicht die Release-Fassung. Dieselben 6 Aufgaben, an denen der Fehler aus R81 gefunden und behoben wurde. Claude-allein-Bezug vom Vortag (R74).",
    "en": "Interim build of 28 Sep, not the release version. The same 6 tasks on which the bug from R81 was found and fixed. Claude-alone reference from the day before (R74)."
   }
  },
  {
   "id": "R83",
   "from": "2026-09-28",
   "to": "2026-09-29",
   "block": "R",
   "title": {
    "de": "Großtest, Stufe einfach (7 Neubauten)",
    "en": "Large test, easy level (7 new builds)"
   },
   "kinds": [
    "api_kosten",
    "qualitaet",
    "abo_limit",
    "tempo"
   ],
   "taskType": "neubau",
   "version": "release",
   "build": {
    "de": "Release-Fassung (28.09.2026)",
    "en": "release version (28 Sep 2026)"
   },
   "unit": "lauf",
   "tasks": 7,
   "runsStarted": 21,
   "runsValid": 21,
   "runsValidWithGrace": 14,
   "runsValidClaudeAlone": 7,
   "runsValidOther": 0,
   "countedInTotals": true,
   "status": "valid",
   "result": {
    "de": "API-Kosten −48 % (5-Minuten-Cache; −51 % mit 1-Stunden-Cache); Qualität auf allen 7: 85 % gegen 85 % (ohne die eine Aufgabe, an der beide scheiterten: 94 % gegen 97 %); 5-Stunden-Limit beide +2 Punkte (zu grob); 2,1× länger.",
    "en": "API cost −48 % (5-minute cache; −51 % with 1-hour cache); quality on all 7: 85 % vs 85 % (without the one task both failed: 94 % vs 97 %); 5-hour limit both +2 points (too coarse); 2.1× longer."
   },
   "invalidReason": {
    "de": "Nach Stufe 1 von 5 beendet und durch den Fülltest ersetzt.",
    "en": "Ended after level 1 of 5 and replaced by the fill test."
   },
   "headline": true,
   "caveat": {
    "de": "Die API-Kosten sind aus Abo-Läufen zu Anthropic-Listenpreisen gerechnet, nicht in einem eigenen API-Lauf gemessen. Qualität auf allen 7 Aufgaben 85 % gegen 85 %; „94 % gegen 97 %“ gilt nur für 6 von 7 (ohne die Aufgabe, an der beide scheiterten). Keine der 7 Aufgaben wurde vorher in einer gültigen Grace-Messung benutzt.",
    "en": "API cost is computed from subscription runs at Anthropic list prices, not measured in a separate API run. Quality on all 7 tasks 85 % vs 85 %; \"94 % vs 97 %\" holds only for 6 of 7 (without the task both failed). None of the 7 tasks was used before in a valid Grace measurement."
   }
  },
  {
   "id": "R84",
   "from": "2026-09-29",
   "to": "2026-09-29",
   "block": "R",
   "title": {
    "de": "Fülltest: 34 Neubauten, zwei Konten",
    "en": "Fill test: 34 new builds, two accounts"
   },
   "kinds": [
    "abo_limit",
    "api_kosten",
    "qualitaet",
    "tempo"
   ],
   "taskType": "neubau",
   "version": "release",
   "build": {
    "de": "Release-Fassung (28.09.2026)",
    "en": "release version (28 Sep 2026)"
   },
   "unit": "lauf",
   "tasks": 34,
   "runsStarted": 68,
   "runsValid": 65,
   "runsValidWithGrace": 32,
   "runsValidClaudeAlone": 33,
   "runsValidOther": 0,
   "countedInTotals": true,
   "status": "valid",
   "result": {
    "de": "5-Stunden-Limit 25 % gegen 44 % (−43 % abgelesen), Wochenlimit 4 gegen 7 Punkte; auf 31 sauberen Paaren API-Kosten −47 %, Qualität 91 % gegen 94 % (12 schlechter, 15 gleich, 4 besser), Dauer etwa gleich.",
    "en": "5-hour limit 25 % vs 44 % (−43 % as read), weekly limit 4 vs 7 points; on 31 clean pairs API cost −47 %, quality 91 % vs 94 % (12 worse, 15 equal, 4 better), time about equal."
   },
   "invalidReason": {
    "de": "3 Läufe nicht gewertet: 2 Grace-Abstürze beim Start (Datenbanksperre, behoben), 1 Claude-Lauf vom Inhaltsfilter gestoppt.",
    "en": "3 runs not scored: 2 Grace crashes at start (database lock, fixed), 1 Claude run stopped by a content filter."
   },
   "headline": true,
   "caveat": {
    "de": "Das 5-Stunden-Limit zeigt nur ganze Prozentpunkte: abgelesen 25 % gegen 44 % = −43 %; wegen der Rundung liegt der wahre Wert etwa zwischen −41 % und −45 %. Gerechnete Bereinigungen: ohne andere Nutzung auf dem Grace-Konto ≈ −45 %, ohne die 2 abgestürzten Grace-Läufe ≈ −37 %, beide zusammen ≈ −39 %. Im Test luden Graces innere Aufrufe die Projekteinstellungen; das Produkt lädt die Nutzereinstellungen, mit allen Hooks ausgeschaltet (seit Plugin 0.3.2). Keine der 34 Aufgaben wurde vorher in einer gültigen Grace-Messung benutzt.",
    "en": "The 5-hour limit shows whole percentage points only: read 25 % vs 44 % = −43 %; because of rounding the true value lies roughly between −41 % and −45 %. Calculated adjustments: without other usage on the Grace account ≈ −45 %, without the 2 crashed Grace runs ≈ −37 %, both together ≈ −39 %. In the test Grace's inner calls loaded project settings; the product loads user settings with all hooks switched off (since plugin 0.3.2). None of the 34 tasks was used before in a valid Grace measurement."
   }
  },
  {
   "id": "R85",
   "from": "2026-09-29",
   "to": "2026-09-29",
   "block": "G",
   "title": {
    "de": "Drei kostenlose Delegations-Werkzeuge",
    "en": "Three free delegation tools"
   },
   "kinds": [
    "konkurrenz",
    "api_kosten",
    "tempo"
   ],
   "taskType": "aenderung",
   "version": "keine",
   "build": {
    "de": "ohne Grace",
    "en": "no Grace"
   },
   "unit": "lauf",
   "tasks": 1,
   "runsStarted": 12,
   "runsValid": 8,
   "runsValidWithGrace": 0,
   "runsValidClaudeAlone": 2,
   "runsValidOther": 6,
   "countedInTotals": true,
   "status": "partly_valid",
   "result": {
    "de": "Alle drei teurer als Claude allein (1,3–2,9× Kosten); je Arm 2 Läufe auf derselben Kleinaufgabe, ohne Qualitätsbewertung.",
    "en": "All three cost more than Claude alone (1.3–2.9×); 2 runs per arm on the same small task, no quality judging."
   },
   "invalidReason": {
    "de": "4 Läufe ungültig: mit neutralem Hinweis delegierte Claude nie.",
    "en": "4 runs invalid: with a neutral hint Claude never delegated."
   },
   "headline": false,
   "caveat": null
  },
  {
   "id": "R86",
   "from": "2026-09-29",
   "to": "2026-09-29",
   "block": "G",
   "title": {
    "de": "Vergleich mit kostenlosem Werkzeug (Hauptlauf)",
    "en": "Comparison with free tool (main run)"
   },
   "kinds": [
    "konkurrenz"
   ],
   "taskType": "neubau",
   "version": "keine",
   "build": {
    "de": "ohne Grace",
    "en": "no Grace"
   },
   "unit": "keine",
   "tasks": null,
   "runsStarted": 0,
   "runsValid": 0,
   "runsValidWithGrace": 0,
   "runsValidClaudeAlone": 0,
   "runsValidOther": 0,
   "countedInTotals": false,
   "status": "not_run",
   "result": {
    "de": "Geplant, noch nicht gestartet.",
    "en": "Planned, not started yet."
   },
   "invalidReason": null,
   "headline": false,
   "caveat": null
  },
  {
   "id": "R87",
   "from": "2026-10-01",
   "to": "2026-10-01",
   "block": "R",
   "title": {
    "de": "Grace 0.3.6 auf 30 Aufgaben des Fülltests",
    "en": "Grace 0.3.6 on 30 fill-test tasks"
   },
   "kinds": [
    "api_kosten",
    "qualitaet",
    "tempo"
   ],
   "taskType": "neubau",
   "version": "v036",
   "build": {
    "de": "Grace 0.3.6 (01.10.2026)",
    "en": "Grace 0.3.6 (1 Oct 2026)"
   },
   "unit": "lauf",
   "tasks": 30,
   "runsStarted": 30,
   "runsValid": 30,
   "runsValidWithGrace": 30,
   "runsValidClaudeAlone": 0,
   "runsValidOther": 0,
   "countedInTotals": true,
   "status": "valid",
   "result": {
    "de": "API-Kosten 65,00 $ gegen 107,54 $ für Claude allein auf denselben 30 Aufgaben (Claude-allein-Läufe vom 29.09., R84): −40 %; Zeit −27 % (bei 25 von 30 schneller); Qualität 93,6 % gegen 92,1 %; 0 Abstürze. Das 5-Stunden-Limit war nicht sauber ablesbar (das Konto wurde parallel genutzt).",
    "en": "API cost $65.00 vs $107.54 for Claude alone on the same 30 tasks (Claude-alone runs of 29 Sep, R84): −40 %; time −27 % (faster on 25 of 30); quality 93.6 % vs 92.1 %; 0 crashes. The 5-hour limit could not be read cleanly (the account was in use at the same time)."
   },
   "invalidReason": null,
   "headline": false,
   "caveat": null
  },
  {
   "id": "R88",
   "from": "2026-10-01",
   "to": "2026-10-01",
   "block": "R",
   "title": {
    "de": "Grace 0.3.7: Kontrollmessung (6 Aufgaben)",
    "en": "Grace 0.3.7: control run (6 tasks)"
   },
   "kinds": [
    "api_kosten",
    "qualitaet"
   ],
   "taskType": "neubau",
   "version": "v037",
   "build": {
    "de": "Grace 0.3.7 (01.10.2026)",
    "en": "Grace 0.3.7 (1 Oct 2026)"
   },
   "unit": "lauf",
   "tasks": 6,
   "runsStarted": 7,
   "runsValid": 7,
   "runsValidWithGrace": 7,
   "runsValidClaudeAlone": 0,
   "runsValidOther": 0,
   "countedInTotals": true,
   "status": "valid",
   "result": {
    "de": "API-Kosten 7,66 $ gegen 7,93 $ mit 0.3.6 (R87) und 9,56 $ für Claude allein (R84) auf denselben 6 Aufgaben; Qualität 96,8 % gegen 95,1 %; 0 Abstürze; dazu 1 Lauf über den Weg des Produkts.",
    "en": "API cost $7.66 vs $7.93 with 0.3.6 (R87) and $9.56 for Claude alone (R84) on the same 6 tasks; quality 96.8 % vs 95.1 %; 0 crashes; plus 1 run through the product path."
   },
   "invalidReason": null,
   "headline": false,
   "caveat": null
  },
  {
   "id": "R89",
   "from": "2026-10-02",
   "to": "2026-10-02",
   "block": "R",
   "title": {
    "de": "Gleichzeitiger Vergleich: 30 neue Aufgaben, zwei Konten",
    "en": "Simultaneous comparison: 30 new tasks, two accounts"
   },
   "kinds": [
    "abo_limit",
    "api_kosten",
    "qualitaet",
    "tempo"
   ],
   "taskType": "neubau",
   "version": "v037",
   "build": {
    "de": "Grace 0.3.7 (01.10.2026)",
    "en": "Grace 0.3.7 (1 Oct 2026)"
   },
   "unit": "lauf",
   "tasks": 30,
   "runsStarted": 60,
   "runsValid": 59,
   "runsValidWithGrace": 30,
   "runsValidClaudeAlone": 29,
   "runsValidOther": 0,
   "countedInTotals": true,
   "status": "valid",
   "result": {
    "de": "5-Stunden-Limit 23 % gegen 32 % abgelesen; ohne ≈ 1,0 Punkte andere Nutzung auf dem Grace-Konto ≈ −31 %; Wochenlimit 4 gegen 5 Punkte (zu grob); API-Kosten 57,30 $ gegen 96,77 $ (−41 %); Zeit −29 % (bei 28 von 30 schneller); Qualität 94,9 % gegen 93,7 % (E031 deutlich schlechter).",
    "en": "5-hour limit 23 % vs 32 % as read; without ≈ 1.0 points of other usage on the Grace account ≈ −31 %; weekly limit 4 vs 5 points (too coarse); API cost $57.30 vs $96.77 (−41 %); time −29 % (faster on 28 of 30); quality 94.9 % vs 93.7 % (E031 clearly worse)."
   },
   "invalidReason": {
    "de": "1 Claude-allein-Lauf (G028) endete ohne Ergebnis und ist nicht gewertet.",
    "en": "1 Claude-alone run (G028) ended without a result and is not scored."
   },
   "headline": true,
   "caveat": {
    "de": "Gleichzeitig gemessen, beide Konten Claude Max 20x. Das 5-Stunden-Limit zeigt nur ganze Prozentpunkte: abgelesen 23 % gegen 32 %; die ≈ 1,0 Punkte andere Nutzung auf dem Grace-Konto sind aus dessen Sitzungsprotokollen gerechnet. 5 der 30 Aufgaben hatten früher Claude-allein-Läufe (R40, R50), keine einen Grace-Lauf.",
    "en": "Measured at the same time, both accounts Claude Max 20x. The 5-hour limit shows whole percentage points only: read 23 % vs 32 %; the ≈ 1.0 points of other usage are calculated from the session logs of the Grace account. 5 of the 30 tasks had earlier Claude-alone runs (R40, R50), none a Grace run."
   }
  },
  {
   "id": "R90",
   "from": "2026-10-02",
   "to": "2026-10-02",
   "block": "R",
   "title": {
    "de": "Grace 0.3.8 (Freigabekandidat), getrennt gemessen",
    "en": "Grace 0.3.8 (release candidate), measured separately"
   },
   "kinds": [
    "abo_limit",
    "api_kosten",
    "qualitaet"
   ],
   "taskType": "neubau",
   "version": "v038",
   "build": {
    "de": "Grace 0.3.8, Freigabekandidat (02.10.2026)",
    "en": "Grace 0.3.8, release candidate (2 Oct 2026)"
   },
   "unit": "lauf",
   "tasks": 30,
   "runsStarted": 30,
   "runsValid": 30,
   "runsValidWithGrace": 30,
   "runsValidClaudeAlone": 0,
   "runsValidOther": 0,
   "countedInTotals": true,
   "status": "valid",
   "result": {
    "de": "5-Stunden-Limit 18 % abgelesen (Fenster bei 0 % gestartet) gegen 32 % des Claude-allein-Laufs aus R89 (dieselben Aufgaben, derselbe Tag, getrennt gemessen): ≈ −44 %; Wochenlimit 2 Punkte (zu grob); API-Kosten 34,41 $ gegen 96,77 $; Qualität 92,2 % gegen 93,7 % auf 29 Paaren (G020 deutlich schlechter).",
    "en": "5-hour limit 18 % as read (window started at 0 %) vs 32 % for the Claude-alone run of R89 (same tasks, same day, measured separately): ≈ −44 %; weekly limit 2 points (too coarse); API cost $34.41 vs $96.77; quality 92.2 % vs 93.7 % on 29 pairs (G020 clearly worse)."
   },
   "invalidReason": null,
   "headline": true,
   "caveat": {
    "de": "Getrennt gemessen: Grace 0.3.8 lief allein auf einem Konto mit Claude Max 20x; Vergleich ist der Claude-allein-Lauf aus R89 (dieselben 30 Aufgaben, derselbe Tag, nicht gleichzeitig) — deshalb kein Zeitvergleich. Abgelesen in ganzen Prozentpunkten: 18 % gegen 32 %. Freigabekandidat von 0.3.8 mit demselben Verhalten wie die Freigabe.",
    "en": "Measured separately: Grace 0.3.8 ran alone on a Claude Max 20x account; the comparison is the Claude-alone run of R89 (same 30 tasks, same day, not at the same time) — so no time comparison. Read in whole percentage points: 18 % vs 32 %. Release candidate of 0.3.8 with the same behaviour as the release."
   }
  },
  {
   "id": "R91",
   "from": "2026-10-03",
   "to": "2026-10-03",
   "block": "R",
   "title": {
    "de": "Bestätigung: Grace 0.3.8 gleichzeitig, 30 Aufgaben, zwei Konten",
    "en": "Confirmation: Grace 0.3.8 at the same time, 30 tasks, two accounts"
   },
   "kinds": [
    "abo_limit",
    "api_kosten",
    "qualitaet",
    "tempo"
   ],
   "taskType": "neubau",
   "version": "v038",
   "build": {
    "de": "Grace 0.3.8 (03.10.2026)",
    "en": "Grace 0.3.8 (3 Oct 2026)"
   },
   "unit": "lauf",
   "tasks": 30,
   "runsStarted": 60,
   "runsValid": 59,
   "runsValidWithGrace": 30,
   "runsValidClaudeAlone": 29,
   "runsValidOther": 0,
   "countedInTotals": true,
   "status": "valid",
   "result": {
    "de": "5-Stunden-Limit 14 % gegen 31 % abgelesen (beide Fenster bei 0 % gestartet, keine andere Nutzung): ≈ −55 %; zusammen mit R90 (60 Aufgaben) 32 gegen 63 Punkte: ≈ −49 %; Wochenlimit 2 gegen 5 Punkte (zu grob); API-Kosten 39,76 $ gegen 95,10 $ (−58 %); Zeit −49 % (bei 28 von 30 schneller); Qualität 92,8 % gegen 90,2 % auf 29 Paaren (E004 deutlich schlechter).",
    "en": "5-hour limit 14 % vs 31 % as read (both windows started at 0 %, no other usage): ≈ −55 %; together with R90 (60 tasks) 32 vs 63 points: ≈ −49 %; weekly limit 2 vs 5 points (too coarse); API cost $39.76 vs $95.10 (−58 %); time −49 % (faster on 28 of 30); quality 92.8 % vs 90.2 % on 29 pairs (E004 clearly worse)."
   },
   "invalidReason": {
    "de": "1 Claude-allein-Lauf (X003) wurde vom Inhaltsfilter vor dem Ende gestoppt und ist nicht gewertet.",
    "en": "1 Claude-alone run (X003) was stopped by a content filter before the end and is not scored."
   },
   "headline": true,
   "caveat": {
    "de": "Gleichzeitig gemessen, beide Konten Claude Max 20x, abgelesen in ganzen Prozentpunkten; die Regeln der Messung standen vor dem Start fest. Eine davon ist formal verfehlt: Auf dem Vergleichskonto fehlen an einem 5-Stunden-Neubeginn 54 Sekunden Ablesung, der Verbrauch dieser Sekunden fehlt in dessen 31 % (etwa 0,2 Punkte beim mittleren Verbrauch). Das kann nur den Vergleich kleiner und Grace damit schlechter aussehen lassen, nie besser. Alle Regeln zum Ergebnis selbst (Ersparnis, Qualität, Abstürze, Zeit) sind erfüllt. Die 30 Aufgaben stammen aus dem Fülltest (R84) und liefen mit Grace 0.3.6 schon in R87; keine davon ist in R89/R90.",
    "en": "Measured at the same time, both accounts Claude Max 20x, read in whole percentage points; the rules of the test were fixed before the start. One of them was formally missed: on the baseline account, 54 seconds of readings are missing at a 5-hour reset, so the usage in those seconds is missing from its 31 % (about 0.2 points at its average rate). That can only make the baseline smaller and Grace look worse, never better. Every rule on the result itself (saving, quality, crashes, time) was met. The 30 tasks come from the fill test (R84) and ran with Grace 0.3.6 before in R87; none of them is in R89/R90."
   }
  }
 ],
 "notCounted": [
  {
   "de": "Vorstufe von Grace als eigenständiger Agent (14.–17.09.2026): rund 210 Läufe auf 18 Website-Aufgaben. Nicht in den Summen, weil ein anderes Produkt ohne Vergleich mit Claude Code und nicht einzeln auf Gültigkeit geprüft.",
   "en": "Grace's predecessor as a standalone agent (14–17 Sep 2026): about 210 runs on 18 website tasks. Not in the totals: a different product, not compared with Claude Code and not checked run by run for validity."
  },
  {
   "de": "Richter- und Auswertungsaufrufe (mehrere tausend Einzelurteile). Sie bewerten Läufe, sind aber selbst keine Messläufe.",
   "en": "Judge and analysis calls (several thousand single verdicts). They score runs but are not runs themselves."
  },
  {
   "de": "Handläufe ohne erhaltene Rohdaten (u. a. R20 und eine Mini-Probe zu R72).",
   "en": "Manual runs without surviving raw data (including R20 and a mini probe for R72)."
  },
  {
   "de": "Die Entwicklung vor dem 13.09.2026 (andere Produktstufe).",
   "en": "Development before 13 Sep 2026 (different product stage)."
  }
 ],
 "notPublished": "Acceptance criteria, full task texts, project and customer names, file paths and the names of third-party tools are withheld. Raw data per task is published for the fill test of 29 Sep 2026, the two tests of 2 Oct 2026 and the confirmation test of 3 Oct 2026."
}