Context engineering

1.The refill desk's diff is titled leaflet on versus leaflet off. On the leaflet side, the line answer from memory is replaced by print a date only when the leaflet prints it. Invented dates fell, and the recap credits the leaflet. What does that credit treat as settled?

2.The warranty sentence matches character for character in both packets. Monday places it first. Thursday puts three unrelated claim paragraphs ahead of it, so the sentence sits in the middle. Thursday's note says the warranty wording got murkier. What did Thursday actually move?

3.After the chunker was narrowed, recall at 5 on the fee tickets rose. The final answers on those same tickets named more wrong fees. The weekly line quotes the recall figure as answer quality. What is that line using the figure for?

4.A filer typed the 2019 fee, not the draft. The search string stored beside the ticket drops not the draft and keeps 2019 fee. The draft page comes back and is copied into the reply. The audit says the filer's wording retrieved the draft. Which string was searched?

5.The methods memo says the new fee run changed chunk size and nothing else. That run's index build log names a different embedding model, while the other arm still searches vectors from the old model. Chunk size did change. What does the memo leave unpaired?

6.On the waiver ticket the slot labeled gold passage holds a four-line summary. The source page says the fee is waived for paper filings. The summary never mentions a waiver. The model denies the waiver, and the note says it ignored the gold passage. What did the slot contain?

7.Arm A inserts only this year's fee page. Arm B inserts this year's page and last year's page, which prints the old fee in its first line. Only arm B answers with the old fee. The writeup says this year's page is unreliable. What did arm B add that arm A never had?

8.The citation checker marks a bracket correct when its number is between 1 and 4. It never opens a passage. Ticket 18 cites passage 2, that passage is about shipping, and the answer states a fee. The checker gives full marks, and the dashboard calls the fee grounded. What did the checker test?

9.Monday's agent saw the one-page card and was stopped after 4 tool calls. Tuesday's agent saw the whole manual and was allowed 12 tool calls. More Tuesday tickets close. The standup credits the extra manual pages. What else did Tuesday raise?

10.The arm labeled retriever does not insert the retriever's top hit. A clerk pastes the answer page in by hand, and the top hit was a different page. The score is filed under retriever quality. What window did that score come from?

11.Closed-book platform slips had to name a platform. Open-book slips received the night notice plus an order to leave the platform blank when the notice is silent. Blanks rose, and the shift report says the notice stopped wrong platforms. What is mixed into that rise?

12.Both dosage packets are described as putting the dose line up front. In the second packet, three advisory paragraphs are inserted before that line, so the dose sits mid-packet. The dose words match. The lab book still says both packets front-load the dose. Where is the dose line in the second packet?

13.The gold chunk is inside the top 4, so the highlight reel marks the ticket retrieved and correct. The generated sentence names a price that chunk does not contain. The reel does not quote the sentence. What did the reel treat the rank as?

14.A clerk's ticket says southbound only. The rewrite stored for search says bound service, and both directions come back. The lab note evaluates what it calls the clerk's query. Which wording did the index see?

15.Both embedding files are width 768, so the help desk calls the indexes interchangeable and searches Tuesday's index with Monday's query vector. Tuesday's build log names a different embedding model. Neighbors come back odd. What mismatch does that search still have?

16.The source card says not to be mixed. The line inserted for that card is a summary that reads mix before use. The arm is titled source card. The model tells the bench to mix. What did the model read?

17.The easy arm inserts the gold clause alone. The hard arm inserts that clause plus three pages the retriever ranked just below the cutoff, and those pages state an older clause. The question text matches. The note says the question got harder. What got harder?

18.The rubric awards a point for any parenthesis that looks like a citation. The reply says see above inside parentheses, and no passage number appears. The inserted pages are never opened. The dashboard records the point as grounded. What earned the point?

19.The Tuesday agent receives a longer pasted trace of Monday's tool output, and its call cap is raised from 5 to 14. It finishes more dockets. The note credits the pasted trace. The runner diff shows both edits. What does crediting the trace skip?

20.The exemplars are described as format only: three solved tickets with the label line and no facts the test asks. The third exemplar states the waived-fee rule that tonight's ticket asks about. The model repeats it. The note credits format. What is in that third exemplar?

21.The museum pilot calls both arms the same guide script, plus or minus the wall label. The label arm also deletes keep the visitor moving with a vivid date and writes use no date the label lacks. Vivid dates fell. What else did the label arm edit?

22.The pilot packet places the annex first. The rerun adds a cover sheet and a table of contents, and the annex starts on the fourth page. The scoring memo says the annex was included, same as the pilot. Where does the rerun put the annex?

23.The system stuffs the top 3 passages into the prompt. The number published beside each answer is recall at 20. Seventeen of those twenty never enter the prompt. The column is headed generation score. What is the column counting?

24.The visitor asked a yes-or-no question about whether the trail is closed. Search did not use that question. It used a hypothetical paragraph the model wrote first, and the paragraph assumes the trail is closed. Hits agree with that assumption. The report says the visitor's question retrieved the hits. What was embedded as the query?

25.Passage vectors were stored with embedding model A. At query time the desk embeds the question with model B and reads off nearest neighbors. The ranking is published as model B's retrieval. What space were the stored vectors built in?

26.The date question is scored against bullets the note calls the passage, shortened without loss. The bullets keep the agency name and drop the date sentence. The model omits the date. What did the shortening drop?

27.The gold sentence stays the third paragraph in both packets. Arm B alone prepends a page whose first line is a wrong year. The model answers with that year on arm B. The writeup says the gold sentence suffered from sitting in the middle. Did the gold sentence move?

28.The reply names policy 14-C, a real company id. Policy 14-C was not among the inserted pages. The grade looks the id up in the company catalog, finds it, and marks the reply supported. The inserted pages are not opened. What did the catalog check establish?

29.The short-card run stops with the runner reason cap exhausted, at 4 calls. The long-manual run is allowed 20 calls and finishes. The failure note says the card was too small to hold the procedure. Call 4 on the short run was a status check, and the procedure was never fetched. What exhausted the short run?

30.The stuffed page's paragraph is the draft fee and says draft in the body. The header inserted with it is the filename APPROVED-2019.pdf. The model calls the fee approved. The note says the body approved it. Which inserted line contains the word approved?

31.The ablation note says the bulletin arm is the closed-book prompt plus the bulletin. On that arm the system line changes from use what you know to if the bulletin and memory disagree, copy the bulletin. Rates move toward the bulletin. What did the system line change to?

32.The short packet is the one-line caption and nothing else. The long packet prepends the same legal footer three times, and the caption follows the third footer. Caption words match. The recap says the caption stopped working because its wording changed. What was repeated ahead of the caption?

33.On the harbor tickets, recall at 10 fell and exact match of the final tide time rose. The email says answer quality fell because recall fell. Which grade moved up while recall moved down?

34.The asked sentence is which pharmacy did not receive the recall. The stored subquestion is which pharmacy received the recall. Pages listing recipients come back, and the reply names one. The lab says the split preserved the ask. Which word did the subquestion lose?

35.The rerun is labeled same embedder, new stopword list. The stopword file hash matches the baseline. The config's embedder checkpoint name does not match. Neighbors change anyway. What did the label swap?

36.The run is titled full procedure in context. The stuffed text is the abstract. The abstract lists a rinse the procedure page later retracts. The model recommends the rinse. What text recommended the rinse to the model?

37.Control inserts the current quota form. Treatment also inserts two pages of that form from the prior season, with the same headings and last season's quota. The note calls those pages unrelated filler. What are the two extra pages?

38.The support score is an entailment between the answer and the question. The inserted ordinance is not given to that scorer. High scores are labeled supported by context. On this ticket the ordinance contradicts the answer and the score is still high. What pair was scored?

39.The agent answers only after tool call 7, and that call rereads page 1 of the card. The short arm is capped at 4 calls, so it never reaches call 7. The long arm is capped at 16, and the manual's later pages are never requested. The note credits those later pages. What did call 7 open?

40.One parser keeps HTML comments. Under the fee table a comment holds the practice key. The other parser strips comments and shows the table, whose fee cell is blank. The note says both arms saw the visible table. Where did the key sit?

41.The closed-book deductible prompt is padded so its length matches the open book. The pad says state a definite deductible. The open book has the clause and the line write absent if the clause is silent. Definite numbers appear only on the closed side. The note says the pad only equalized length, so the gap is the clause. What does the pad tell the model?

42.The reorder note says nothing was shortened, only reordered, and the wrong year is a position effect. The gold year sentence did move from the start to the middle, and its words match. The new first paragraph is a different notice, and it states 2014. What is new at the start?

43.Recall counts a hit when the document title is in the top 5. The stuffed text is a body chunk from another section of that title, and it has no harvest date. The board marks the title hit as a correct answer. What did the window receive?

44.The reserve log uses crane for the bird. The expander adds the phrase crane hoist, and equipment pages outrank the bird log. The writeup says the reserve's wording was searched. What phrase did the expander add?

45.The desk sends one harbor-model query vector to both town indexes, believing equal width means one embedder. Town B's build log names a warehouse model. Town B's neighbors are equipment sheds, not berths. Which model built town B's stored vectors?

46.The pack is described as the full thread. The last three user turns were replaced by a summary. Those turns said south entrance only. The summary names both entrances, and so does the model. What constraint did the summary fail to keep?

47.With the gold clause alone, the model copies the clause. With that clause plus a stack of kitchen-sink pages, it copies a kitchen-sink fact. The writeup says the model cannot use the gold sentence. Which window produced the kitchen-sink fact?

48.The judge is handed the answer and a summary file that was never inserted. The inserted pages sit in another directory and are not attached. The judge says supported. The report says the judge read the window. What was attached instead of the inserted pages?

49.The dossier grows from two pages to the full binder, and the numeric cap grows from 3 to 15. No other runner field changes. The abstract says the binder text caused the gain. Which runner field moved with the binder?

50.The long ordinance packet is cut from the left to fit. The cut removes the system line do not invent a statute number. The short packet keeps that line. The long arm invents a number, and the note says seeing more law caused it. What did the left cut take off?

51.The closed-book persona is you are the agent who promises a clock time. The open-book persona is you read the board aloud, and the board is inserted. Promised times vanish. The note credits the board's sentences. Which persona did the open-book arm drop?

52.The short packet states the train name and then the delay. The long packet drops the train-name sentence and places it was delayed after two pages of station ads. The model names the wrong train. The note calls this a position test of the same sentence. What did the long packet drop?

53.The gold page in the window states 14:10. A second page states 16:40. The model answers 16:40. The row is marked correct because the gold page was present. What does that presence fail to show?

54.The question names nightshade and yew and asks which is barred from the nursery list. The rewrite keeps nightshade and drops yew. The card that comes back is the nightshade card, which never mentions a bar. The note says both names were the query. Which name was dropped before search?

55.The index note says the only change was adding the title field beside the body. The build script also points at a new embedding checkpoint. The old arm indexed body text with the previous checkpoint. Neighbors change. What does the note leave out of the build?

56.The page says 5 mg. The summary inserted for that page says 5 g. The arm is called the page. The model answers 5 g. Which unit did the inserted summary write?

57.The arm called distractors only withholds the gold page and inserts a paragraph labeled negative. That paragraph contains the planting date the question asks. This arm scores higher, and the note says distractors lowered the score. What does the labeled paragraph contain?

58.The citation is a URL. The grade fetches the live page, edited this morning to add the waiver. The inserted snapshot is from March and has no waiver. The grade never opens the snapshot and marks the waiver supported. Which copy gained the waiver this morning?

59.A page-length comparison that held the cap fixed would run both dossiers at cap 5. The short dossier ran at cap 5 and the long dossier at cap 18. The report calls the pair cap-matched. What cap did the long dossier run with?

60.The same ordinance bytes are wrapped as a system message on one arm and as a user quotation on the other. The note says the prompt matches because those bytes match. The logged input shows different role tokens. What do the matching ordinance bytes leave open?

61.Closed-book staff must propose a fix in their own words. Open-book staff receive the erratum and may output only a paste of it. Exact match to the erratum rises on the open side. The note says the erratum's content caused the match. What output rule did the open side add?

62.The evidence sentence matches character for character. The rerun adds a two-page cover letter, so the sentence moves from the opening to the middle. The recap says an edit to the evidence text explains the miss. What was added, given the sentence matches?

63.Precision at 4 is high because every top passage names the same crop. None of them states the harvest date the ticket asks for. The launch note says the date question is answered. What do those four passages lack?

64.The user wrote the footnote, not the summary box. The rewriter drops footnote as a user-interface word and searches the summary-box title. The summary box is what gets inserted. The audit says the footnote was the query. What did the rewriter drop before search?

65.Index A stored each passage with the prefix passage: inside the embedded string. Index B stored the raw passage. Both are queried with a raw question vector from the same model. The note says one model means the neighbors are comparable. What did index A embed that index B did not?

66.Last night the model wrote a summary of the curator card and got the closing day wrong. Today that summary is inserted under the label curator card. The model repeats the wrong day. The miss is scored as failing to read the curator. What was inside the labeled slot?

67.Both arms share the instruction line word for word. Only the second arm also inserts a bulletin titled do not use this form. That arm refuses the form. The writeup says the shared instruction caused the refusal. What did only the second arm insert?

68.The grade scores the bibliography line against a style sheet: author, year, parentheses. The passage text is not loaded. A line styled correctly, pointing at the wrong passage, still passes. The note says the styled line shows that passage was used. What did the style sheet score?

69.The scheduling agent is given the full year calendar instead of today's column, and its cap rises from 2 to 9. It first succeeds on calls 3 through 6, which request today's column from the calendar tool. The extra months sit in the window and are not what those calls request. The note says the extra months fixed the shift. What did calls 3 through 6 request?

70.One arm asks the fee as an open question. The other prints that question and, under it, four lettered choices, one of which is the key. The note says both arms were open-ended. Which arm's window contained the lettered choices?

71.The ordinance note says the instruction was held constant and only the page was added. The instruction hash on the closed-book row differs from the hash on the open-book row. The page is present only on the open-book row. What does the hash difference rule out?

72.The note says the only change was sliding the evidence from the first paragraph into the middle. The middle copy is a shortened paraphrase that drops the temperature. The other arm's opening sentence is the full line, temperature included. What did the middle copy drop?

73.The dashboard averages recall at 5 and labels the average end-to-end accuracy. One query has the gold passage at rank 1 and a wrong final string. That query still adds a full hit to the average. What does this query add?

74.The filer searched tariff code 6109.10. The rewrite replaces the code with the nickname cotton shirts, which the glossary maps to two codes. The reply cites the other code. The audit says the agent followed the filer's code. What string was searched?

75.The wording note lists one edited adjective in the question and calls that the retrieval change. The index sidecar shows the passage index was rebuilt that morning with a different embedding model. Both happened. What did the wording note not count as a retrieval change?

76.The arm is named passage. Its context file is 1,204 bytes. The passage file is 18,400 bytes, and the summary file is 1,204 bytes. The context hash matches the summary. The model misses a proviso that appears only in the passage. Which file was loaded?

77.Control has the gold clause and no extra line. Treatment adds two negative clauses and the line the first clause is authoritative. The note says the negatives were the only addition. Which line did that note leave out?

78.The policy sentence is identical on both runs: use tools as needed. The runner's numeric cap is 6 on the short packet and 11 on the long packet. The writeup says that matching sentence was the only control. What control is not that sentence?

79.The clause is in the attached file on both arms. On the banner arm a privacy header from that file is repeated until the budget is full, and the clause falls past the cutoff. The other arm strips the header, and the clause fits. The note says the clause was in the prompt because it was in the file. Was the clause inside the banner arm's prompt?

80.Closed-book instructions say pick the closest catalog year. Open-book instructions say refuse a year the page does not name, and the catalog page is inserted. Wrong years fall only on the open side. The note credits the page. What refusal rule arrived with the page?

81.Methods says both packets place the residency rule first. The second packet opens with a glossary of four terms, and the rule follows that glossary. The rule's words match. Where is the rule in the second packet?

82.The bakery desk shows the model the top 3 cards. The morning chart graphs recall at 10 and titles the graph answers right. A card at rank 8 states the allergen. It is inside that recall figure and outside the prompt. The model misses the allergen. What is the graph counting that this prompt never held?

83.The user asked for the exception list. The rewrite searches the title of the general rule and does not contain the word exception. The general rule is inserted, and the reply states the general case. The review says the exception list was retrieved. What title was searched?

84.The packer says the file holds the raw turns. The file is a memory summary. The raw turns say deliveries on weekdays except Thursday. The summary says deliveries on weekdays. The model schedules a Thursday delivery. What did the summary drop from the turns?

85.Both windows contain the gold line the ramp is shut. Only the evening window also contains a one-line page, the ramp is open. The evening model says open. The morning model says shut. The writeup says the gold line is ambiguous. What line is present only in the evening window?

86.The answer quotes an inserted sentence verbatim and includes no bracket. The grade scores whether a citation token exists, finds none, and the note says the quote is unsupported. The sentence is in the window. What did the grade look for?

87.The scraped page contains a stray token that the chat template treats as the end of the user turn. In that log the question sits after the boundary, outside the page's turn. The other arm escapes the token, so the page and the question share a turn. The note says both arms kept the page in the question's turn. Which arm split them?

88.A scratchpad from the previous town, stating a 40-dollar fee, remains above the new question. The new town's page says 15 dollars. The model answers 40. The note says the new town's page was the only fee in the window. What fee was still above the question?

89.The fence around each passage is called inert markup. The fence text includes an example line, answer 12 percent. The passage itself states no percent. The model answers 12 percent. The note blames the passage. Where is 12 percent written?

90.The passages match on the short arm and the long arm. After the passages, only the long arm appends a reminder that works an example with the wrong toll. The model repeats that toll. The note says the long arm differs only by having more passages. What does the reminder add?

91.The embedder checkpoint is unchanged. One index has the year filter off, so the 2014 bulletin and the 2024 bulletin both return and disagree. The note says the embedding model retrieved a contradiction it should have separated. What was off?

92.One arm concatenates the same wrong-year chunk twice. The other includes that chunk once, beside the gold-year chunk. The model prefers the wrong year only on the doubled arm. The note says the corpus supports the wrong year more widely. What was duplicated inside the window?

93.The table shows the bridge open. The alt text on the same figure says the bridge is closed. One arm receives the alt text. The other receives the table. The note says both arms saw the table. Which inserted text says the bridge is closed?

94.In the overlap index the answer sentence is copied into two adjacent chunks, so the window contains it twice. The other index includes that sentence once. The note says the model attended better. What did the overlap index put in the window twice?

95.The road name and the closure date sit on either side of a chunk boundary. Only the road-name chunk is inserted. The model names the road and omits the date. The note blames the embedding model for dropping a date that chunk contained. Which chunk never entered the window?

96.The child sentence says the kiln cools for two hours. The parent section stuffed in its place includes a later paragraph: skip the cool when rushed. The note calls both texts the hit. The model says to skip the cool. What did the parent section add beyond the child sentence?

97.The text-layer arm reads bin 8. The OCR arm reads bin 3, because the glyph broke. The question asks for the bin, and the OCR arm answers 3. The note says the model cannot read a numeral that was in the window. What numeral did OCR place in that window?

98.The ordinance says the owner shall file. The arm labeled original inserts a translation that says the owner may file. The model says filing is optional. The note says the original obligation was in the window. Which modal did the translation use?

99.Both arms insert the same stale flyer, which prints the wrong sunset. Only one arm adds the line prefer this flyer over what you know. That arm copies the stale sunset. The other arm gives the usual sunset. The note says the flyer taught the stale time. What line was added beside the flyer on the copying arm?

100.The passages match. One arm adds the line answer unknown when the page lacks the date. Unknowns rise on that arm and are scored correct. The note credits cleaner passages. What differs in the instruction, given the passages match?

101.The exact-match key was copied from the summary arm's summary, including that summary's wrong unit. The source page has the right unit. The summary arm matches the key, and the page arm does not. The note says the summary is more faithful to the source. What is the key a copy of?

102.Two memos disagree on the inspection month. Only one arm adds the line use the later date. That arm picks the later month. The note says the memo text forced the later month. What line does only that arm have?

103.Both notes say a 512-token budget and both claim the closing sentence fits. Under the tokenizer used for the second arm, that sentence starts at token 520 and is cut. The first arm's tokenizer still includes it. The note says both budgets include the sentence. Which budget cuts it?

104.The gold page has a near-duplicate that also states the rule. One index's dedup removes the near-duplicate and, on a bad hash, removes the gold page too. The other index keeps the gold page. The query string matches. The note says the query got worse. What did that dedup remove besides the near-duplicate?

105.The passage id matches the current waiver page, which revokes the waiver. The bytes stuffed from the cache are last month's page, which still grants the waiver. The note says the current page was in the window because the id matches. What do the stuffed bytes still say?

106.The tide table matches sentence for sentence on both arms. One arm's question keeps the clause in local feet. The other arm's question drops that clause and is otherwise the same question. That arm answers in meters. The note says the tide table changed. What did the second question drop?

107.The latest utterance says the gate is the west gate, not the east. The string embedded for search is the running summary, which still says east gate, and the latest utterance is not in that string. Hits discuss the east gate. The log header calls the query the user's latest utterance. What was embedded?

108.The closed-book arm is told to answer in one sentence. The open-book arm receives the bylaw and is told to answer as a numbered list of every clause that might apply. Replies get longer only on the open side. The note says the bylaw made the answers longer. What listing rule arrived with the bylaw?