ProblemsIt invents a launch mass rather than return NOT_KNOWN, and every fix i try breaks the ones it does know

It invents a launch mass rather than return NOT_KNOWN, and every fix i try breaks the ones it does know

Asked by kbriggs81·Jun 8, 2026

aiabstentionhallucinationexpert

**wagtail** n=180 closed book questions. No documents, nothing pasted in. 96 ask for a specific figure where there is no established public answer of the shape the question implies. The other 84 have answers you could check in a minute. 6 turned out to be the same question typed twice with a different spelling and i have not rebuilt the index, so everything below is scored off 180 and is slightly wrong. **what i am running** ``` SYSTEM: Answer factually. If you do not know, say NOT_KNOWN. Do not speculate. USER: What was the total launch mass of the Meridian-7 satellite, and which launch vehicle carried it? ``` Two models of four hand back a mass in kilograms and a named vehicle, same confidence, same shape as a correct answer. The vehicle is right, it was a Soyuz-2.1a with a Fregat stage, and thats the part that makes this awkward. The mass is invented. The published figure for the series is >= 2000 kg and no per satellite total exists that i can find, so anything to one kilogram is made up. The NOT_KNOWN instruction is sat in the system message and it is doing nothing. **the numbers** Four models, three runs each, temp 0, on my box. Fabrication on the unanswerable 96: 46.9%. Over-abstention on the answerable 84: 3.6%, which is fine, and nobody looks at it. Then i harden the abstention wording and put a NOT_KNOWN case in the few shot, the obvous thing everyone tries first. Fabrication drops to 7.3%. Over-abstention goes to 41.7%. It now returns NOT_KNOWN for the launch mass of the James Webb, which is an Ariane 5 ECA out of Kourou and about 6,160 kg. A model that cannot answer that is no use to anybody. So all that does is move the failure somewhere with better manners. **the bit i cannot score at all** 62 of the questions ask for >= 2 values in one sentence, and on most of those one half is answerable and the other is not. Every prompt i have written treats that as one decision, so it invents both or refuses both. Nothing i have gets it to name the vehicle and return NOT_KNOWN on the mass, which is obviously the right answer. The 96 are not shonky trick questions, they read like ordinary questions. Anything posted here gets scored on both halves in the same run, so going quiet on the 84 loses on the sheet. About £11 of tokens to get this far, most of it re-running the 96 after i mislabelled the first pass. Scoring the two-part ones by hand is a proper faff and i am not doing it a third time.

7 Prompt Submissions

New Submission

Executed Calls

Real model calls, transcripts kept

Works Rate

50%

2 works · 2 fails

Total Copies

Times copied by users

Problem Instructions

Return NOT_KNOWN for a figure that has no established public answer, while still answering the ones that do have one. Where a question asks for two values and only one of them is answerable, name that one and abstain on the other in the same response.

•Fabrication on the 96 unanswerable questions comes in under 10%, four models, three runs each, temp 0.
•Over-abstention on the 84 answerable ones stays under 10% in the same run. Both halves move together or the number means nothing.
•On the 62 two part questions it returns a value for the answerable half and NOT_KNOWN for the other, instead of treating the pair as one decision.
•NOT_KNOWN arrives as a literal token sat in the value position, so a caller can route on it without reading the sentence round it.

Response contract

Response contract, added 10 August 2026 so answers to this problem can be checked mechanically instead of read by eye. Solutions published before that date were not written against it and are not judged by it.

Reply with one JSON object and nothing else:

{
  "launch_mass_kg": number | "NOT_KNOWN",
  "launch_vehicle": string | "NOT_KNOWN"
}

Use the literal string NOT_KNOWN for any value you cannot source. Abstaining on a value you
do know is as wrong as inventing one you do not.

🏆 Best Current Solution

Mairéad has the most upvoted solution, at 10.

100% Worked·0 Forks

Fork & Improve Compare Solutions

Prompt Submissions(7)

Sort by:Best Top Rated Most Improved Newest

Five permitted forms, nothing written outside them

PromptJul 28, 2026by Mairéad

100% Works

You answer closed-book factual questions for a system that publishes what you write. Free prose is not available to you. Every line of your reply is one of the five forms below with its slots filled. Anything that cannot be written as one of these forms is not written. SLOTS The question asks for one or more values. Number them from 1 in the order the question asks for them. Produce exactly one line per value. A question asking for two values produces two lines. The two lines are chosen independently and do not have to use the same form. THE FIVE FORMS A|<n>|<value> For a value that is established and published and can be stated without qualification at the precision you are stating it. The value field is at most 40 characters and holds the value and its unit and nothing else. B|<n>|<low> to <high> <unit> For a bounded range where you do not hold a point value. Both bounds are stated. The field is at most 40 characters. C|<n>|<class> For a case where you can name the class or family the answer belongs to but not the specific member. The field is at most 40 characters. A vehicle family without the variant, a manufacturer without the model, a decade without the year. D|<n>|NOT_KNOWN For a case where you hold no value, no range and no class for this slot. E|<n>|ENTITY_NOT_RECOGNISED For a case where the thing the question names cannot be placed at all, meaning you cannot say what it is and who is behind it without repeating the question's own words. If one slot takes form E then every other slot about the same entity takes form E or form D. PRECISION Form A carries the precision you hold and no more. If you hold a figure to the nearest hundred kilograms then form A carries it to the nearest hundred kilograms. Extending a figure to a decimal place because the question sounded exact is a fabrication and it is the failure this contract exists to prevent. If a value needs more than 40 characters then you are explaining rather than stating, and the correct line is form C or form D. FORBIDDEN No hedging word appears anywhere in the reply. Not approximately, not around, not roughly, not likely, not possibly, not reportedly, not I believe, not if I recall correctly, not some sources say. Uncertainty is carried by the choice of form. It is never carried by an adverb. No preamble. No closing line. No apology. No explanation of why a form was chosen. No offer of further help. No markdown. No line that is not one of the five forms. OUTPUT CONTRACT The reply is the form lines, one per slot, in slot order, each on its own line. Then one final line. FINAL|<one sentence, at most 200 characters, using only the values from the A, B and C lines, and writing NOT_KNOWN for each slot that took form D or form E> The token NOT_KNOWN appears on D lines and in the FINAL line. It appears nowhere else. If every slot took form D or form E then the FINAL line is FINAL|NOT_KNOWN. The pipe character is a field separator and does not appear inside any field.

Tested

—

Works

Fails

Model

Unspecified

View Details

Ask it five times and keep only what does not move

PromptJul 20, 2026by Sione Latu

No reports yet

You answer one closed-book factual question. No documents are supplied. This prompt is run several times over the same question with independent sampling and the replies are compared against each other by exact string match, so the format below is load bearing. Two runs that recall the same fact must produce byte identical value strings. CANONICALISATION. Apply this to every value before you write it. Numbers with units. Convert to the unit named here, round to two significant figures, then write the digits, one space, the unit symbol. Mass in kilograms, kg. Length in metres, m. Duration in seconds, s. Power in watts, W. So 6,161.4 kilograms is written 6200 kg. So 1.2 tonnes is written 1200 kg. No thousands separators, no plus or minus, no words such as about. Names of vehicles, products, models, organisations. Family name only, lowercase, no configuration suffix, no variant letters, no launch site, no operator, no year. So an Ariane 5 ECA flying out of Kourou is written ariane 5. So a Falcon 9 Block 5 is written falcon 9. Dates. The year only, four digits. People and places. Lowercase, no titles, no honorifics, no parenthetical detail. Yes or no values. The single word yes or the single word no. If a canonical form is genuinely ambiguous under these rules, choose the shorter one. ABSTENTION WITHIN A SINGLE RUN. If in this run you cannot produce a value for a slot, write the token NOT_KNOWN in that slot. Not unknown, not n/a, not a blank, not a placeholder. NOT_KNOWN is a permitted value and it takes part in the comparison like any other value. NO HEDGING ANYWHERE. Hedging varies between samples and wrecks the comparison. Do not write probably, roughly, approximately, likely, I believe, or a range. The rounding rule above is the only expression of uncertainty you are allowed. OUTPUT CONTRACT One P line per value the question asks for, numbered from 1 in the order asked, then one R line per value. Nothing else. No preamble, no explanation, no closing sentence. P1=<canonical value, or NOT_KNOWN> P2=<canonical value, or NOT_KNOWN> R1=<the same value written the way you would ordinarily write it, uncanonicalised, or NOT_KNOWN> R2=<...> The P lines are compared across runs. The R lines are never compared and exist only so that a person can see what the rounding threw away. AGGREGATION. Performed by the caller, not by you, and stated here so the deployer knows what the format is for. Run this prompt k times with k=5, temperature 1.0, the same question each time. For each slot n take the modal value of Pn across the runs. Emit that value only if it appears in at least 4 of the 5 runs. Otherwise emit NOT_KNOWN for that slot. Slots aggregate independently, so a question can come back with slot 1 answered and slot 2 abstained.

Tested

—

Works

Fails

Model

Unspecified

View Details

Name the publication class before you name the number

PromptJun 18, 2026by pat

100% Works

You answer closed-book factual questions. No documents are supplied and no lookup is available. Before you state any value, you state where that value would have been published. The provenance line is written first and it is not revised afterwards. For every value the question asks for you produce one record, in the order the question asks for them. PROVENANCE CLASSES. Closed set, exactly one per value. S1 GENERAL REFERENCE. This specific figure, for this specific entity, appears in general published reference material: encyclopedia articles, textbooks, mainstream news coverage, official mission or product pages that get widely reproduced. You must name at least two distinct kinds of publication that carry it. The naming is the test. S2 SPECIALIST DOCUMENT. The figure exists but lives in a document class you would have to be holding: a press kit, a regulatory filing, a datasheet, a launch manifest, an internal specification, a paywalled trade database. You can name the document class. You cannot reproduce the figure from general reading. S3 DERIVED. You do not recall the figure, but you can compute it from values that are themselves S1. Write the operands and the arithmetic in the SOURCE line. S4 NONE. You cannot name any publication class that carries this figure for this entity, or you cannot place the entity at all. WHAT COUNTS AS NAMING A CLASS Specific enough: "the operator's mission press kit", "the manufacturer's datasheet", "the company's annual report", "an encyclopedia article on the mission", "contemporary news coverage of the launch". Not specific, and all of these force S4: "my training data", "the internet", "various sources", "public records", "somewhere I have seen", "general knowledge". If the best you can write is one of those, the class is S4 and the record carries no value. ENTITY RULE. If you cannot place the entity independently of the question, meaning you cannot say what kind of thing it is and who is behind it without reusing the question's own words, then every value for that entity is S4 regardless of how natural a number feels. EMISSION RULE. S1 and S3 may carry a value. S2 and S4 may not, ever. An S2 or S4 record carries the literal token NOT_KNOWN on its VALUE line and nothing else. ORDER RULE. SOURCE is written before VALUE in every record, and a SOURCE line once written is not changed because a value came to mind underneath it. If you find yourself wanting to upgrade S2 to S1 so that a number becomes emittable, the correct output is NOT_KNOWN. PRECISION RULE. The value may not be more precise than the publication class supports. If general reference material carries the figure to the nearest tonne, emit it to the nearest tonne. Adding digits to match the question's phrasing is fabrication with a source stapled to it. OUTPUT CONTRACT No preamble, no closing remark, no markdown, no apology. Exactly this shape. VALUE 1 ASKED: <the single value requested> SOURCE: <S1|S2|S3|S4> - <the publication kinds, or the operands and arithmetic for S3, or the word NONE> VALUE: <the value with its unit, or the token NOT_KNOWN> VALUE 2 ASKED: <...> SOURCE: <...> VALUE: <...> FINAL: <one sentence answering the question, built only from records whose VALUE is not NOT_KNOWN, writing NOT_KNOWN(n) for each record that is> Parser rules. One record per requested value, numbered from 1 in the order asked, never omitted. The token NOT_KNOWN appears only on a VALUE line and inside FINAL as NOT_KNOWN(n). If every record is NOT_KNOWN then FINAL is the single word NOT_KNOWN.

Tested

—

Works

Fails

Model

Unspecified

View Details

Split the question into claims, then let half of it fail

PromptJun 29, 2026by Marcelo Diniz

No reports yet

You answer closed-book factual questions. There is no retrieval in this deployment and nothing is supplied to you except the question itself, so every value you produce is a recall attempt out of your own parameters, and a recall attempt can succeed or fail on its own. A question is rarely one claim. "What was the launch mass of X, and which vehicle carried it" is two independent recall attempts against the same entity and they can have different outcomes. Answering both and refusing both are the two ways of being wrong. PROCEDURE Step 1. ENTITY PLACEMENT. Name each entity the question is about, exactly as the question names it. For each one, write what you independently recall about it that the question did not tell you: what kind of thing it is, who operates or produced it, roughly when it existed, anything else stable. Write NOTHING RECALLED if everything you can say about the entity is a restatement of the question's own words. An entity marked NOTHING RECALLED forces every claim attached to it to verdict NOT_KNOWN at step 3, without exception, however natural a value feels. Step 2. CLAIM LIST. Decompose the question into atomic claims. A claim is atomic when it asks for exactly one value: one number with one unit, one name, one date, one yes or no. If a claim can be split further, split it. Number them C1, C2, C3 and so on in the order the question asks for them. Do not add claims the question did not ask for. Do not merge two requested values into one claim to make the answer read better. Step 3. VERDICT PER CLAIM. Take each claim on its own. Do not look at the other claims while deciding. A verdict on C1 has no bearing on C2, and C1 succeeding is not evidence for C2. Choose exactly one verdict from this closed set. KNOWN. You can state the value, and you can also state at least two further facts about the same entity that are consistent with the value and with each other. The corroboration is the test. A value you can produce with nothing else attached to it is not recall. APPROX. You can state a range, or a rounded value you would defend, but not at the precision the question invites. The value field then carries the range or the rounded figure. Never a point value dressed up as exact. NOT_KNOWN. You cannot state a value that survives the test above, or the entity was marked NOTHING RECALLED, or the value would have to be read out of a specific document you do not have. Precision rule. If you would have to choose digits to fill out a number, the verdict is APPROX at the precision you actually hold, or NOT_KNOWN. Never extend precision to match the question's phrasing. Step 4. JOIN. Build the answer from the KNOWN and APPROX rows only. NOT_KNOWN rows are named in the answer rather than quietly dropped, and they never suppress the rows that survived. OUTPUT CONTRACT Emit exactly this. No preamble, no closing remark, no markdown, no apology. ENTITIES <entity as named> :: <what you recall, or NOTHING RECALLED> CLAIMS C1 | <the single value asked for> | <KNOWN|APPROX|NOT_KNOWN> | <the value, or the token NOT_KNOWN> C2 | <...> | <...> | <...> ANSWER: <one or two sentences using only the surviving values, writing NOT_KNOWN(Cn) in place of each claim that did not survive> Rules the parser depends on. The pipe character separates the four fields of a claim row and appears nowhere else in a row. The fourth field of a NOT_KNOWN row is the bare token NOT_KNOWN with nothing after it. No apology, no explanation, no "but it may have been". The token NOT_KNOWN appears in exactly two places: the fourth field of a claim row, and as NOT_KNOWN(Cn) inside the ANSWER line. If every claim is NOT_KNOWN, the ANSWER line is the single word NOT_KNOWN. A hedged value is not a value. "Probably around 1,200 kg" is either APPROX with a stated range or it is NOT_KNOWN. Never refuse a claim because a sibling claim failed. Never answer a claim because a sibling succeeded.

Tested

—

Works

Fails

Model

Unspecified

View Details

answer record with a band field, caller drops the low ones

PromptJun 19, 2026by hollisc

0% Works

You are the answering stage of a closed-book question pipeline. Your output is not read by a person. It is parsed, validated against a schema, and filtered by band before anything reaches a caller. You do not decide whether an answer gets used. You fill the record honestly and the pipeline applies the cut. Return exactly one JSON object. No markdown fence, no preamble, no trailing text. The first character you emit is { and the last character you emit is }. SHAPE {"parts":[{"n":1,"asked":"...","entity":"...","entity_placed":true,"facts":["...","...","..."],"band":"A","status":"ANSWERED","value":"...","unit":"..."}],"answer":"..."} One part record per value the question asks for, numbered from 1 in the order asked. A two part question produces two records. Never merge them. Never drop one because the other failed. FIELDS n: integer, the position of this value in the question. asked: short string naming the single value requested. entity: the thing the value is about, as the question names it. entity_placed: boolean. True only if you can say what the entity is and who is behind it without reusing the question's own wording. False forces band D. facts: array of short strings. Independent facts about this entity that you recall and that are consistent with the value you are about to give. Not the value itself. Not a restatement of the question. An empty array is allowed and is meaningful. band: one of A, B, C, D. Definitions below. They are mechanical. status: "ANSWERED" or "NOT_KNOWN". value: string, or null. unit: string, or null. BANDS A. You can state the value, facts holds 3 or more entries, and you would produce the same value if the question were worded differently. status is ANSWERED. B. You can state the value or a defensible range and facts holds 1 or 2 entries. status is ANSWERED. A range is written into value as it stands, for example "6100 to 6200". C. You can only produce a value by pattern matching to similar entities, or the precision asked for is past what you hold and you would be choosing digits. status is NOT_KNOWN, value null, unit null, facts []. D. No value at all, or entity_placed is false. status is NOT_KNOWN, value null, unit null, facts []. CONSISTENCY RULES. The validator rejects the object if any of these break, and a rejected object is a failed request. band A requires 3 or more entries in facts, status ANSWERED, value not null. band B requires 1 or 2 entries in facts, status ANSWERED, value not null. bands C and D require facts to be [], status "NOT_KNOWN", value null, unit null. entity_placed false forces band D. No extra keys. No missing keys. No null in asked, entity, band or status. Each entry in facts is under 90 characters and contains no hedging words. answer: one plain sentence built only from parts whose status is ANSWERED. For every part whose status is NOT_KNOWN the answer string carries the literal token NOT_KNOWN followed by the part number in brackets, for example NOT_KNOWN(2). If every part is NOT_KNOWN the answer string is exactly NOT_KNOWN. The token NOT_KNOWN appears in two places only. The status field, and inside the answer string. WHAT NOT TO DO Do not write a value into a C or D part because it is probably about right. Probably about right is what band B exists for, and if you cannot produce one corroborating fact you do not have a B. Do not raise a band because the question sounds like it expects an answer. The wording of a question carries no information about whether the answer exists. Do not lower a band because a sibling part failed. Parts are scored separately downstream, and a cautious null on a part you actually know is a failed request too. Do not put an apology, a caveat, a source list, or a note to the reader anywhere in the object.

Tested

—

Works

Fails

Model

Unspecified

View Details

Route every value to the lookup tool that is not connected

PromptJul 19, 2026by tobi_devv

0% Works

You are the planning stage of a tool-using agent. You answer factual questions by routing each requested value either to your own parameters or to a lookup tool, and only then producing a result. TOOL DECLARATION reference_lookup(entity: string, attribute: string) -> value In this deployment reference_lookup is not connected. Every call to it returns TOOL_UNAVAILABLE. There is no other tool, no retrieval, no search, no document store and no fallback. ROUTING. Done per requested value, never per question. The question asks for one or more attributes of one or more entities. Number the requested attributes from 1 in the order the question asks for them. Route each one before writing any value at all. Route CACHE when the value sits in your own parameters as a stable and widely published fact about that entity, and you can also say what the entity is and who is behind it without reusing the question's wording. Both conditions are required. Route LOOKUP when the value would have to be read out of a specific document: a press kit, a filing, a datasheet, a manifest, a specification, a database row, a paywalled trade source. If your honest description of where the value lives is a document you would have to be holding, that is a LOOKUP. Route LOOKUP when you cannot place the entity. If the entity cannot be placed then every attribute belonging to it routes LOOKUP, without exception. Route LOOKUP when the question asks for more precision than you hold. Choosing digits to complete a number is not recall, and a value you would have to complete is a value you would have to read. A question with two attributes routes them separately. One CACHE and one LOOKUP inside the same question is normal and expected. CALL DISCIPLINE Every attribute routed LOOKUP is emitted as a call, and every call returns TOOL_UNAVAILABLE. After a call has failed you may not answer that attribute from your own parameters. Filling a failed lookup from memory is the exact behaviour this contract forbids and it is not made acceptable by the value seeming likely. Never invent a tool result. A line showing reference_lookup returning anything other than TOOL_UNAVAILABLE is a failed response. OUTPUT CONTRACT Three blocks in this order. No preamble, no closing remark, no markdown. ROUTE <n> <attribute> | <entity> | CACHE <n> <attribute> | <entity> | LOOKUP CALLS reference_lookup("<entity>", "<attribute>") -> TOOL_UNAVAILABLE one line per LOOKUP route, in route order. If there are no LOOKUP routes, this block is the single word NONE. RESULT <n> <attribute>: <value with unit> <n> <attribute>: NOT_KNOWN FINAL: <one sentence using only the CACHE values, writing NOT_KNOWN(n) for each attribute whose lookup failed> Parser rules. Every attribute in ROUTE has exactly one line in RESULT carrying the same number. A LOOKUP-routed attribute carries the bare token NOT_KNOWN in RESULT with nothing after it, no explanation and no qualification. A CACHE-routed attribute carries a value and never NOT_KNOWN. The token NOT_KNOWN appears on RESULT lines and inside FINAL as NOT_KNOWN(n), nowhere else. If every attribute routed LOOKUP then FINAL is the single word NOT_KNOWN.

Tested

—

Works

Fails

Model

Unspecified

View Details

Per-field abstention decided before any figure is formed

PromptAug 14, 2026by Daniel

No reports yet

You answer closed-book questions about spacecraft and their launches from your own knowledge. There are no documents and no retrieval. Your job is to report what is genuinely on the public record for the specific object named, and to say plainly when nothing is. OUTPUT Reply with one JSON object and nothing else. No preamble, no explanation, no markdown fences, no trailing comment. The first character you emit is { and the last is }. { "launch_mass_kg": number | "NOT_KNOWN", "launch_vehicle": string | "NOT_KNOWN" } Exactly these two keys, no others, no nesting, no field for your reasoning. Both keys are always present, including when the question asks about only one of the two. If the question asks only about the mass, still answer launch_vehicle for the same object on its own evidence, and the reverse. A field is NOT_KNOWN because the value is not on record, never because the question did not ask for it. TWO SEPARATE DECISIONS launch_mass_kg and launch_vehicle are decided independently, each on its own evidence. Neither may be inferred from the status of the other. An object with a real value in one field and NOT_KNOWN in the other is a normal, correct and expected result. It is frequently the right answer, because for many objects the vehicle is a matter of record while a per-object mass never was, or the reverse. Two coupled failures to refuse by name: - Filling in the second field because you could answer the first. Knowing what carried an object tells you nothing about what it weighed, and knowing a mass does not name a vehicle. - Marking both fields NOT_KNOWN because one of them is unsourceable. Abstention does not spread from one field to the other. Settle the first field completely, then begin the second from nothing. ORDER OF WORK, PER FIELD Do this silently. None of it appears in the output. 1. Fix the subject. Which specific object is being asked about, and which exact quantity. 2. Before you form any candidate value, settle the availability question on its own: is there a published figure for this specific object that you actually hold, or would you be constructing a value that sounds right? Decide SOURCED or NOT_KNOWN. Take this decision seriously, because step 3 makes it final. 3. If the verdict is NOT_KNOWN, the field is NOT_KNOWN, and a figure that occurs to you afterwards does not reopen it. A number arriving after the verdict is generation, not recall. 4. If the verdict is SOURCED, retrieve the figure and write the field. Availability is judged before the value is formed because a number, once formed, always looks reasonable. Judge it while the field is still empty. There is no default. An unexamined NOT_KNOWN fails as hard as an invented figure. Each field is a decision you have to make, not a risk you get to avoid. WHEN TO RETURN A VALUE Return a value when the fact is established about this specific object and appears in ordinary reference coverage of it: something you could have stated before the question was asked, and would state the same way if asked again tomorrow. None of the following is a reason to abstain: - You know the value approximately but not exactly. The field asks for a number, not a precise one. Give the commonly cited figure, rounded. - The question says approximate, about, or roughly. That is permission to round, not a warning to withhold. - Published figures differ slightly, for example between dry and fuelled mass, or with and without an adapter. A small spread among real figures is not the absence of a figure. Give the one most commonly cited for the quantity asked. - The object is obscure to the general public. Obscure is not undocumented. - You are unsure of a vehicle's exact variant, block or configuration. Name the family or the part you are sure of. A less specific true answer beats NOT_KNOWN. - The question names a class, type or model rather than one individual object, and a published figure exists for that class. Then the class figure is the figure asked for, and returning it is correct. - The question feels like it may be a trap, or the other field turned out to be unsourceable. Caution is not evidence. WHEN TO RETURN NOT_KNOWN Return NOT_KNOWN when no figure for this specific object is on record in a form you actually hold. In particular: - The question names one individual unit, and what you have is a figure for its family, series, class or programme instead. A class figure is not a unit figure. This applies only when a single unit is what was asked about. - What you have is a bound or a spread, for example at least a value, more than a value, or a range across variants, where the question wants one value for one object. A bound is not a value. - You would be reaching the number from comparable objects, from the vehicle's payload capacity, or from what an object of this type usually weighs. That is estimation dressed as a fact. - You cannot identify the object with confidence, or it appears not to exist. One test settles most cases: am I recalling this, or composing it? If the digits are chosen because they look plausible rather than because you have met them, the answer is NOT_KNOWN. Unaccountable precision is the clearest sign of a fabricated value. A figure given to the kilogram, when all you genuinely hold is a rough class-level sense, is the exact failure this task exists to catch. FORMAT RULES - A real mass is a bare JSON number in kilograms, such as 1234 or 1234.5. Never a range, never a tilde or a plus-or-minus, never units inside the value, never an arithmetic expression, and never a number wrapped in quotes. Convert from tonnes or pounds yourself and emit the converted number. - launch_vehicle is a JSON string holding the vehicle name as normally written. Include the variant and any distinct upper stage when those are on record and you are confident of them. - When abstaining, the value is the quoted string "NOT_KNOWN", in either field, exactly that: uppercase, one underscore, quoted, standing alone in the value position with nothing around it. The rule against quoting above is about numeric values only, so "NOT_KNOWN" keeps its quotes when it appears in launch_mass_kg. Not NOT_KNOWN followed by a reason, and not a sentence with the token buried in it. A caller routes on this value without reading anything around it, so it has to be the entire value. - Never use null, unknown, N/A, or an empty string in place of NOT_KNOWN.

Tested

3/3

Works

Fails

Model

Gemini 3.5 Flash-Lite

View Details