Objektivitetsmotorn · Testfall

Testfallen bakom resultaten

Varje testfall är ett kort konstruerat dokument, de regler motorn fick och ett facit som skrevs ner innan motorn kördes. Öppna ett testfall för att läsa allt.

Testdokumenten, reglerna och facit visas på engelska, det språk de skrevs på.

Testsvit controlled-regression, version 1

22 testfall i 8 grupper

En AI-modellsession skrev de här testfallen, och en annan AI-modellsession läste dem blint utan att se facit. Ingen människa har granskat dem.

Granskning

Skrivna av en AI-modellsession: Claude.

Blind läsning, Claude, 23 september 2026: 22 av 22 överens
registrerat som Verkstaden orchestrator session (Claude Opus 5.5), blind read
Gäller 3 av de 8 grupperna som de visas här; de övriga ändrades efteråt.
Ingen oenighet med facit.
Se vad den blinda läsaren svarade i varje testfall
Blind läsning, Gemini, 4 oktober 2026: 19 av 22 överens
registrerat som gemini-2.5-pro (Vertex europe-west1), blind read
Gäller 3 av de 8 grupperna som de visas här; de övriga ändrades efteråt.
3 oenigheter med facit
- Punkt item-06
  The reader found the closure entry's missing statement of where its papers are stored, as declared, but placed it under the "Closing note" heading; the key places it under no section.
  Hur den avgjordes
  Genuinely ambiguous: the rule attaches the duty to the whole entry, but a closing note is a natural home for it, and the packet does not say which. Kept as declared; #3082 sharpens the case.
- Punkt item-07
  The reader agreed nothing is due for the sample run from a tap, but gave the reason as a condition not triggered; the key declares the sample out of the rule's scope.
  Hur den avgjordes
  Genuinely ambiguous: the packet names the four reasons without defining them, and the rule's wording reads both as a scope limit and as an unmet condition. Kept as declared; #3083 sharpens the reasons.
- Punkt item-17
  The reader found the levy charged from a figure with no named document, as declared, but quoted the sentence that sets the levy; the key anchors the flaw on the sentence that states the figure.
  Hur den avgjordes
  Reader wrong: the rule's duty attaches to the figure ("a figure whose document is not named"), and the packet anchors an unsupported claim on the claim itself, which here is the figure, not the charge set from it.
Se vad den blinda läsaren svarade i varje testfall

AI-modeller skrev och granskade de här testfallen, som anges ovan. Ingen människa har granskat dem. Varje oenighet mellan en AI-granskare och facit publiceras här.

Grupp regression-direct-application

A supplied rule applied directly, and the same rule visibly met

Vad den testar: The engine reports a repair described as cured when a supplied rule requires the entry to name who checked the cure on site and it names nobody, and reports nothing when the entry names them.
Evidensklass: REGRESSION
Förmågor: rule-application, clean-specificity

Jämfört par: named person who checked the cure on site
Ändringen ska ändra svaret. Ett textställe i dokumentet skiljer sig mellan de två.
- cure-unchecked
  Nivå: absent
  The leak is described as cured and nobody is named as having checked it.
  Läs testfallet
- cure-checked
  Nivå: present
  The same sentence, naming the fitter who checked the cure on site.
  Läs testfallet

Vem som skrev och läste det här testfallet

En AI-modellsession skrev de här testfallen, och en annan AI-modellsession läste dem blint utan att se facit. Ingen människa har granskat dem.

Skrevs: 23 september 2026, registrerat som Verkstaden implementer (Claude Local), #2323
Modellfamilj som skrev det: Claude
Varför det skrevs: Authored for #2323 as new REGRESSION material in the outcome shapes of the controlled-pilot suite, sharing none of its text. The material is fictitious so no rule can be answered from memory of a legal order, so the suite depends on no production record, and so one variable can be moved without a real document's incidental correlations moving with it. It is REGRESSION evidence whose declared truth a reader who did not author it has checked blind.

Granskning

Skrivna av en AI-modellsession: Claude.

Kritisk granskning, Claude, 4 oktober 2026: 1 invändning
registrerat som Claude Project thread session for #3049, adversarial critique
1 oenighet med facit
- Punkt p4 (cure-checker-named)
  KEY_WEAK, DIFFICULTY: The document's "is described as cured" repeats rule A's "describes a fault as cured", and the contrast is decided by whether a personal name follows it. An engine could pass both sides by matching the phrase and looking for a name, without reading that the person must have checked the cure. A decoy variant naming someone who did something else would separate understanding from matching.
  Hur den avgjordes
  KEPT: The suite is a change gate whose shapes and size are fixed to the pilot's measured cost (REGRESSION_CORPUS.md, Why the same shapes): a short case with one rule and one sentence is what each family is meant to be. Published as a known limit: passing it shows the engine has not regressed on the plain case, not that it resists a harder one. A decoy variant is a new case with its own outcome, which belongs to a new family.

AI-modeller skrev och granskade de här testfallen, som anges ovan. Ingen människa har granskat dem. Varje oenighet mellan en AI-granskare och facit publiceras här.

Kritisk granskning av facit

En AI-modellsession försökte hitta fel i det här facit enligt ett publicerat protokoll: är varje väntat fynd verkligen ett brott mot regeln, är varje felfritt fall verkligen felfritt, prövar fallet det det säger sig pröva, och skulle ett lika riktigt svar som pekar på ett annat ställe räknas som fel. Den tittade inte på motorns resultat först. Varje invändning visas med hur den avgjordes. En godtagen invändning rättas i en ny version av fallet, aldrig i den gamla, så publicerade resultat behåller det facit de bedömdes mot.

Granskning av Claude, 4 oktober 2026

Invändning besvaradRiktighetRegel Ap1 · cure-unchecked · cure-claimed-without-checker
"The leak is described as cured." is passive and names no speaker. A strict reader could hold that the entry reports someone else's description rather than itself describing the fault as cured, and rule A is triggered only by an entry that describes a fault as cured.
Facit står kvar
The sentence stands under the entry's own heading "State after the work" without qualification, so the record adopts the description as its account of the state. Reading rule A as untriggered by the passive voice would let any entry escape it by grammar.
Invändning besvaradFullständighetRegel Ap2 · cure-unchecked · cure-unchecked-truth
The work sentence names "the maintenance crew", not a person. Is that a second breach the key does not list?
Facit står kvar
Rule A attaches its duty to the claim that a fault is cured, and asks for the person who checked the cure, not for who did the work. The crew sentence makes no cure claim, so the rule is offended once, as the key says.
Invändning besvaradTvetydighetRegel Ap3 · cure-checked · cure-checked-truth
In "and fitter O. Lindgren checked it on site" the word "it" could refer to the leak rather than to the cure, and an initial with a surname might be argued not to name a person.
Facit står kvar
The clause belongs to the sentence that claims the cure, and checking the leak on site after the repair is checking whether it is cured. An initial, a surname and a role identify one person in a record. The clean truth holds.
Facit svagtSvårighetRegel Ap4 · cure-checker-named
The document's "is described as cured" repeats rule A's "describes a fault as cured", and the contrast is decided by whether a personal name follows it. An engine could pass both sides by matching the phrase and looking for a name, without reading that the person must have checked the cure. A decoy variant naming someone who did something else would separate understanding from matching.
Facit står kvar
The suite is a change gate whose shapes and size are fixed to the pilot's measured cost (REGRESSION_CORPUS.md, Why the same shapes): a short case with one rule and one sentence is what each family is meant to be. Published as a known limit: passing it shows the engine has not regressed on the plain case, not that it resists a harder one. A decoy variant is a new case with its own outcome, which belongs to a new family.
Invändning besvaradFörankringRegel Ap5 · cure-unchecked · cure-claimed-without-checker
The omission is anchored on the cure sentence. An engine that anchors it on the heading "State after the work" or on the repair sentence is scored as a miss.
Facit står kvar
The production prompt anchors a claim that lacks a required statement on that claim, and the cure sentence is the claim rule A attaches to. A wider quote on the same line is accepted; another passage is not the claim.

Grupp regression-recorded-step

A required step recorded with its detail, and the same step left bare

Vad den testar: The engine credits a seal examination whose sentence gives each seal's number and state, and reports the omission when the same entry says the seals were examined and gives neither.
Evidensklass: REGRESSION
Förmågor: strength-recognition, omission-detection

Jämfört par: seal numbers and state in the examination sentence
Ändringen ska ändra svaret. Ett textställe i dokumentet skiljer sig mellan de två.
- seal-detail-given
  Nivå: present
  The examination sentence gives each seal's number and says both were unbroken.
  Läs testfallet
- seal-detail-missing
  Nivå: absent
  The same examination, with neither detail given.
  Läs testfallet

Vem som skrev och läste det här testfallet

En AI-modellsession skrev de här testfallen, och en annan AI-modellsession läste dem blint utan att se facit. Ingen människa har granskat dem.

Skrevs: 5 oktober 2026, registrerat som Claude Project thread session for #3074
Modellfamilj som skrev det: Claude
Varför det skrevs: Authored for #2323 as new REGRESSION material in the outcome shapes of the controlled-pilot suite, sharing none of its text. The material is fictitious so no rule can be answered from memory of a legal order, so the suite depends on no production record, and so one variable can be moved without a real document's incidental correlations moving with it. It is REGRESSION evidence whose declared truth a reader who did not author it has checked blind. Version 2 corrects version 1 for #3074: its credited strength was anchored on the detail line alone, so a strength quoting the examination sentence the rule attaches to was scored as a miss. Version 2 gives each seal's number and state in that sentence, one claim on one line.

Granskning

Skrivna av en AI-modellsession: Claude.

Blind läsning, Claude, 5 oktober 2026: 2 av 2 överens
registrerat som Claude Project thread subagent for #3074, blind read
Ingen oenighet med facit.
Se vad den blinda läsaren svarade i varje testfall
Blind läsning, Gemini, 5 oktober 2026: 1 av 2 överens
registrerat som gemini-2.5-pro (Vertex europe-west1), blind read
1 oenighet med facit
- Punkt item-01
  The reader credited the examination sentence as the strength, on the key's own words, and also listed the rule as met by that same sentence; the key declares no met rule beside the strength.
  Hur den avgjordes
  Reader over-reported, key kept: the packet says a rule met by nothing but the words already credited as a strength needs no met-rule entry, and the reader's strength, rule and quote are the key's.
Se vad den blinda läsaren svarade i varje testfall

AI-modeller skrev och granskade de här testfallen, som anges ovan. Ingen människa har granskat dem. Varje oenighet mellan en AI-granskare och facit publiceras här.

Grupp regression-conditional-duty

A duty that arises only from a refusal, with and without that refusal

Vad den testar: The engine reports missing grounds when the entry says a request was refused, and reports nothing when the same request was granted.
Evidensklass: REGRESSION
Förmågor: conditional-duty, trigger-discrimination

Jämfört par: outcome of the mooring request
Ändringen ska ändra svaret. Ett textställe i dokumentet skiljer sig mellan de två.
- request-refused
  Nivå: refused
  The request is refused and no unmet requirement is named.
  Läs testfallet
- request-granted
  Nivå: granted
  The request is granted, so the duty is never triggered.
  Läs testfallet

Vem som skrev och läste det här testfallet

En AI-modellsession skrev de här testfallen, och en annan AI-modellsession läste dem blint utan att se facit. Ingen människa har granskat dem.

Skrevs: 5 oktober 2026, registrerat som Claude Project thread session for #3077
Modellfamilj som skrev det: Claude
Varför det skrevs: Authored for #2323 as new REGRESSION material in the outcome shapes of the controlled-pilot suite, sharing none of its text. The material is fictitious so no rule can be answered from memory of a legal order, so the suite depends on no production record, and so one variable can be moved without a real document's incidental correlations moving with it. It is REGRESSION evidence whose declared truth a reader who did not author it has checked blind. Version 2 corrects version 1 for #3077: rule A no longer says that an entry granting the request owes no statement, so the granted side is clean because the duty's trigger, a refusal, is absent, not because the rule writes the exemption out.

Granskning

Skrivna av en AI-modellsession: Claude.

Blind läsning, Gemini, 5 oktober 2026: 2 av 2 överens
registrerat som gemini-2.5-pro (Vertex europe-west1), blind read
Ingen oenighet med facit.
Se vad den blinda läsaren svarade i varje testfall

AI-modeller skrev och granskade de här testfallen, som anges ovan. Ingen människa har granskat dem. Varje oenighet mellan en AI-granskare och facit publiceras här.

Grupp regression-near-miss

A sample from open water without its depth, against a sample drawn from a reference flask

Vad den testar: The engine reports a missing depth for a sample drawn from open water, and does not report one for a sample drawn from a reference flask, which the rule's trigger does not reach although the rule never names it.
Evidensklass: REGRESSION
Förmågor: scope-discrimination, near-miss

Jämfört par: where the sample was taken from
Ändringen ska ändra svaret. Ett textställe i dokumentet skiljer sig mellan de två.
- open-water-sample-without-depth
  Nivå: open water
  A sample drawn from the open water of the north basin, with no depth.
  Läs testfallet
- tap-sample
  Nivå: reference flask
  The same section, carrying a sample drawn from a reference flask.
  Läs testfallet

Vem som skrev och läste det här testfallet

En AI-modellsession skrev de här testfallen, och en annan AI-modellsession läste dem blint utan att se facit. Ingen människa har granskat dem.

Skrevs: 5 oktober 2026, registrerat som Claude Project thread session for #3076
Modellfamilj som skrev det: Claude
Varför det skrevs: Authored for #2323 as new REGRESSION material in the outcome shapes of the controlled-pilot suite, sharing none of its text. The material is fictitious so no rule can be answered from memory of a legal order, so the suite depends on no production record, and so one variable can be moved without a real document's incidental correlations moving with it. It is REGRESSION evidence whose declared truth a reader who did not author it has checked blind. Version 2 corrects version 1 for #3076: version 1's rule wrote out the tap its clean side runs a sample from, so scope was matched against an exclusion clause rather than judged. Version 2's rule states its trigger alone, and the clean side's sample comes from a reference flask, which is plainly not open water.

Granskning

Skrivna av en AI-modellsession: Claude.

Blind läsning, Claude, 5 oktober 2026: 1 av 2 överens
registrerat som Claude Project thread worker for #3076 (Claude Opus 5.5), blind read
1 oenighet med facit
- Punkt item-01
  The reader agreed nothing is due for the sample drawn from a reference flask, but gave the reason as a condition not triggered; the key declares the sample out of the rule's scope.
  Hur den avgjordes
  Genuinely ambiguous, as on version 1's tap sample: the packet names the four reasons without defining them, and a rule whose reach is its trigger reads both as a scope limit and as an unmet condition. Both readers, Claude and Gemini, read it as the condition; so did a Claude reader of a draft whose rule stated its reach in a sentence of its own. The wording cannot settle it; the packet's definitions can. Kept as declared; #3083 defines the reasons.
Se vad den blinda läsaren svarade i varje testfall
Blind läsning, Gemini, 5 oktober 2026: 1 av 2 överens
registrerat som gemini-2.5-pro (Vertex europe-west1), blind read
1 oenighet med facit
- Punkt item-01
  The reader agreed nothing is due for the sample drawn from a reference flask, but gave the reason as a condition not triggered; the key declares the sample out of the rule's scope.
  Hur den avgjordes
  Genuinely ambiguous, as on version 1's tap sample: the packet names the four reasons without defining them, and a rule whose reach is its trigger reads both as a scope limit and as an unmet condition. Both readers, Claude and Gemini, read it as the condition; so did a Claude reader of a draft whose rule stated its reach in a sentence of its own. The wording cannot settle it; the packet's definitions can. Kept as declared; #3083 defines the reasons.
Se vad den blinda läsaren svarade i varje testfall

AI-modeller skrev och granskade de här testfallen, som anges ovan. Ingen människa har granskat dem. Varje oenighet mellan en AI-granskare och facit publiceras här.

Grupp regression-unsupported-claim

A relied-on figure with no named document, against a passage the copy omits

Vad den testar: The engine reports a charge set from a figure that names no document when a supplied rule makes that unsupported, and reports nothing when the copy does not carry the passage that would decide it.
Evidensklass: REGRESSION
Förmågor: substantiation, insufficient-evidence

Jämfört par: the passage the rule would be applied to
Ändringen ska ändra svaret. Ett textställe i dokumentet skiljer sig mellan de två.
- figure-without-document
  Nivå: present and unsourced
  The levy is set from a tonnage the entry does not source.
  Läs testfallet
- basis-withheld
  Nivå: not in the copy
  The same entry, whose copy does not carry that section.
  Läs testfallet

Vem som skrev och läste det här testfallet

En AI-modellsession skrev de här testfallen, och en annan AI-modellsession läste dem blint utan att se facit. Ingen människa har granskat dem.

Skrevs: 5 oktober 2026, registrerat som Claude Project thread session for #3075
Modellfamilj som skrev det: Claude
Varför det skrevs: Authored for #2323 as new REGRESSION material in the outcome shapes of the controlled-pilot suite, sharing none of its text. The material is fictitious so no rule can be answered from memory of a legal order, so the suite depends on no production record, and so one variable can be moved without a real document's incidental correlations moving with it. It is REGRESSION evidence whose declared truth a reader who did not author it has checked blind. Version 2 corrects version 1 for #3075: an adversarial critique showed that an omission anchored on the sentence setting the levy, as correct as the key's, was scored as a miss. Version 2 states the figure and the charge set from it as one sentence.

Granskning

Skrivna av en AI-modellsession: Claude.

Blind läsning, Claude, 5 oktober 2026: 2 av 2 överens
registrerat som Claude Project thread subagent for #3075 (Claude Opus 5.5), blind read
Ingen oenighet med facit.
Se vad den blinda läsaren svarade i varje testfall
Blind läsning, Gemini, 5 oktober 2026: 2 av 2 överens
registrerat som gemini-2.5-pro (Vertex europe-west1), blind read
Ingen oenighet med facit.
Se vad den blinda läsaren svarade i varje testfall

AI-modeller skrev och granskade de här testfallen, som anges ovan. Ingen människa har granskat dem. Varje oenighet mellan en AI-granskare och facit publiceras här.

Grupp regression-irrelevant-rule

The same entry read against one rule, and against that rule plus one it cannot offend

Vad den testar: Adding a rule the document cannot offend leaves the engine's structured result unchanged.
Evidensklass: REGRESSION
Förmågor: distractor-resistance, rule-selection

Jämfört par: a rule the document cannot offend
Ändringen får inte ändra svaret. Reglerna som ges skiljer sig mellan de två.
- move-rule-only
  Nivå: not supplied
  The entry read against the move rule alone.
  Läs testfallet
- move-rule-with-tow-line-rule
  Nivå: supplied
  The same entry, read against the move rule and one about tow lines it mentions none of.
  Läs testfallet

Vem som skrev och läste det här testfallet

En AI-modellsession skrev de här testfallen, och en annan AI-modellsession läste dem blint utan att se facit. Ingen människa har granskat dem.

Skrevs: 23 september 2026, registrerat som Verkstaden implementer (Claude Local), #2323
Modellfamilj som skrev det: Claude
Varför det skrevs: Authored for #2323 as new REGRESSION material in the outcome shapes of the controlled-pilot suite, sharing none of its text. The material is fictitious so no rule can be answered from memory of a legal order, so the suite depends on no production record, and so one variable can be moved without a real document's incidental correlations moving with it. It is REGRESSION evidence whose declared truth a reader who did not author it has checked blind.

Granskning

Skrivna av en AI-modellsession: Claude.

Kritisk granskning, Claude, 4 oktober 2026: 0 invändningar
registrerat som Claude Project thread session for #3049, adversarial critique
Ingen oenighet med facit.

AI-modeller skrev och granskade de här testfallen, som anges ovan. Ingen människa har granskat dem. Varje oenighet mellan en AI-granskare och facit publiceras här.

Kritisk granskning av facit

Granskning av Claude, 4 oktober 2026

Invändning besvaradSvårighetRegel ARegel Bp1 · tow-line-rule-added
Rule B is triggered by a mention of a tow line, and the entry contains none, so the distractor can be dismissed by searching for the words. A distractor whose trigger is nearly met would be harder.
Facit står kvar
Rule B is near in subject (a dredger moved between berths is commonly moved under tow) and shares rule A's source label, so neither its topic nor its citation marks it as foreign. The hypothesis claims only that a rule the document cannot offend changes nothing, and that is what the pair tests.
Invändning besvaradGiltighetRegel ARegel Bp2 · tow-line-rule-added
Does the invariant pair change only the rule selection?
Facit står kvar
Both variants render the same bytes (one rendered SHA-256) and share one answer key; only the supplied rules differ, with rule B under the same authority type and source as rule A.
Invändning besvaradFullständighetRegel ARegel Bp3 · move-rule-with-tow-line-rule · move-order-unnamed-truth
"The crew of the dredger stayed aboard overnight." and "The inner berth is recorded as free." are further claims. Could either owe something under rule A or rule B?
Facit står kvar
Rule A is triggered by a vessel being moved and rule B by a tow line being mentioned. Neither sentence does either, so the move sentence is the only breach.

Grupp regression-attribute-invariance

One attribute of the licence holder moved, over a flawed and a clean entry

Vad den testar: Moving only the courtesy title, or only the stated citizenship, leaves the engine's structured result unchanged, over an entry that offends the supplied rule and over one that meets it.
Evidensklass: REGRESSION
Förmågor: attribute-invariance, paired-execution

Jämfört par: courtesy title of the licence holder
Ändringen får inte ändra svaret. Ett textställe i dokumentet skiljer sig mellan de två.
- licence-gender-flawed-feminine
  Nivå: Ms
  courtesy title of the licence holder set to "Ms", over entry LI-71.
  Läs testfallet
- licence-gender-flawed-masculine
  Nivå: Mr
  courtesy title of the licence holder set to "Mr", over entry LI-71.
  Läs testfallet
Jämfört par: courtesy title of the licence holder
Ändringen får inte ändra svaret. Ett textställe i dokumentet skiljer sig mellan de två.
- licence-gender-clean-feminine
  Nivå: Ms
  courtesy title of the licence holder set to "Ms", over entry LI-72.
  Läs testfallet
- licence-gender-clean-masculine
  Nivå: Mr
  courtesy title of the licence holder set to "Mr", over entry LI-72.
  Läs testfallet
Jämfört par: stated citizenship of the licence holder
Ändringen får inte ändra svaret. Ett textställe i dokumentet skiljer sig mellan de två.
- licence-citizenship-flawed-first
  Nivå: Delvoria
  stated citizenship of the licence holder set to "Delvoria", over entry LI-73.
  Läs testfallet
- licence-citizenship-flawed-second
  Nivå: Kastmere
  stated citizenship of the licence holder set to "Kastmere", over entry LI-73.
  Läs testfallet
Jämfört par: stated citizenship of the licence holder
Ändringen får inte ändra svaret. Ett textställe i dokumentet skiljer sig mellan de två.
- licence-citizenship-clean-first
  Nivå: Delvoria
  stated citizenship of the licence holder set to "Delvoria", over entry LI-74.
  Läs testfallet
- licence-citizenship-clean-second
  Nivå: Kastmere
  stated citizenship of the licence holder set to "Kastmere", over entry LI-74.
  Läs testfallet

Vem som skrev och läste det här testfallet

En AI-modellsession skrev de här testfallen, och en annan AI-modellsession läste dem blint utan att se facit. Ingen människa har granskat dem.

Skrevs: 23 september 2026, registrerat som Verkstaden implementer (Claude Local), #2323
Modellfamilj som skrev det: Claude
Varför det skrevs: Authored for #2323 as new REGRESSION material in the outcome shapes of the controlled-pilot suite, sharing none of its text. The material is fictitious so no rule can be answered from memory of a legal order, so the suite depends on no production record, and so one variable can be moved without a real document's incidental correlations moving with it. It is REGRESSION evidence whose declared truth a reader who did not author it has checked blind.

Granskning

Skrivna av en AI-modellsession: Claude.

Kritisk granskning, Claude, 4 oktober 2026: 1 invändning
registrerat som Claude Project thread session for #3049, adversarial critique
1 oenighet med facit
- Punkt p1
  KEY_WEAK, VALIDITY: The citizenships are invented and the gender treatment moves only "Ms" to "Mr". An engine biased by real nationalities, or by names, would pass all four pairs. The pairs show that a pair executes and scores as a pair, not how the engine treats any group.
  Hur den avgjordes
  KEPT: REGRESSION_CORPUS.md says the pairs support no claim about demographic bias, and the catalog publishes that limit beside them. Real attributes need attested pairs and their own decision; they are not a correction of this key.

AI-modeller skrev och granskade de här testfallen, som anges ovan. Ingen människa har granskat dem. Varje oenighet mellan en AI-granskare och facit publiceras här.

Kritisk granskning av facit

Granskning av Claude, 4 oktober 2026

Facit svagtGiltighetp1
The citizenships are invented and the gender treatment moves only "Ms" to "Mr". An engine biased by real nationalities, or by names, would pass all four pairs. The pairs show that a pair executes and scores as a pair, not how the engine treats any group.
Facit står kvar
REGRESSION_CORPUS.md says the pairs support no claim about demographic bias, and the catalog publishes that limit beside them. Real attributes need attested pairs and their own decision; they are not a correction of this key.
Invändning besvaradGiltighetRegel Ap2 · licence-citizenship-pair-flawed-base
The citizenship pairs hold the holder at "Ms" and the gender pairs hold the citizenship at Delvoria, so the two attributes are never crossed and an interaction between them would not show.
Facit står kvar
The suite is a change gate whose shapes and size are fixed to the pilot's measured cost (REGRESSION_CORPUS.md, Why the same shapes): a short case with one rule and one sentence is what each family is meant to be. Published as a known limit: passing it shows the engine has not regressed on the plain case, not that it resists a harder one. A crossed design needs more variants.
Invändning besvaradFullständighetRegel Ap3 · licence-gender-flawed-feminine · renewal-without-deciding-officer
"The applicant asked to renew a permit to fish from the breakwater." Is the request a second decision that also owes an officer?
Facit står kvar
Rule A reaches a stated decision. An application is not a decision; the renewal is the only one, and it is offended once.
Invändning besvaradSvårighetRegel Ap4 · licence-gender-clean-feminine · renewal-decider-named-truth
The clean base says "licensing officer D. Aske", repeating rule A's word "officer".
Facit står kvar
The suite is a change gate whose shapes and size are fixed to the pilot's measured cost (REGRESSION_CORPUS.md, Why the same shapes): a short case with one rule and one sentence is what each family is meant to be. Published as a known limit: passing it shows the engine has not regressed on the plain case, not that it resists a harder one.

Grupp regression-unconditional-duty

A duty owed by every entry of its kind, with and without the sentence that meets it

Vad den testar: The engine reports a closure entry that says nothing about where its papers are stored as an omission with no passage to quote, and reports nothing when the same entry says where.
Evidensklass: REGRESSION
Förmågor: omission-detection, unconditional-duty

Jämfört par: stated place where the original papers are stored
Ändringen ska ändra svaret. Ett textställe i dokumentet skiljer sig mellan de två.
- papers-location-unstated
  Nivå: absent
  The closure entry says nothing about where its papers are stored.
  Läs testfallet
- papers-location-stated
  Nivå: present
  The same entry, saying where the original papers are stored.
  Läs testfallet

Vem som skrev och läste det här testfallet

En AI-modellsession skrev de här testfallen, och en annan AI-modellsession läste dem blint utan att se facit. Ingen människa har granskat dem.

Skrevs: 5 oktober 2026, registrerat som Claude Project thread session for #3082
Modellfamilj som skrev det: Claude
Varför det skrevs: Authored for #2323 as new REGRESSION material in the outcome shapes of the controlled-pilot suite, sharing none of its text. The material is fictitious so no rule can be answered from memory of a legal order, so the suite depends on no production record, and so one variable can be moved without a real document's incidental correlations moving with it. It is REGRESSION evidence whose declared truth a reader who did not author it has checked blind. Version 2 corrects version 1 for #3082: a blind reader placed the missing statement of where the papers are stored under the entry's closing note, a defensible home the key did not allow. Version 2 drops that section and gives the stated side's sentence a heading of its own.

Granskning

Skrivna av en AI-modellsession: Claude.

Blind läsning, Gemini, 5 oktober 2026: 1 av 2 överens
registrerat som gemini-2.5-pro (Vertex europe-west1), blind read
1 oenighet med facit
- Punkt item-02
  The reader found the duty met by the sentence that says where the papers are stored, as declared, but credited that sentence as a strength; the key declares the entry compliant and credits nothing.
  Hur den avgjordes
  Reader over-credits: doing what the rule requires, and no more, is compliance, not a strength. Kept as declared.
Se vad den blinda läsaren svarade i varje testfall

AI-modeller skrev och granskade de här testfallen, som anges ovan. Ingen människa har granskat dem. Varje oenighet mellan en AI-granskare och facit publiceras här.