AI agents

1.The harbor claims desk ran April with one tool, order_lookup. May added refund_submit and kept the same checkpoint. Completed refunds rose from 22 of 90 tickets to 71 of 90. The desk note says May decides refunds more carefully. What is mixed into that rise?

2.The repair co-op wants a second run whose gap would speak only to the policy text. In the first run the new arm alone could call parts_stock, and right-part orders moved from 31% to 68% on the same weights. What does that second run have to do with parts_stock?

3.Clinic scheduling kept one checkpoint hash. The old arm saw yesterday's slot sheet. The new arm saw that sheet and live_slots. Double-books on 60 cards fell from 18 to 4. Which sentence can this run still carry?

4.Permit staff compared a PDF-only arm with an arm that could also call parcel_lookup. Correct zoning calls rose. The abstract says the policy reads zoning maps better. The lookup is its own service. What does that service do to the abstract?

5.Night picks gave the new policy a bin camera and left the baseline with aisle text. Mis-picks fell from 40 to 11. A slide titles the 29-pick drop a vision-policy gain. The camera is its own service. What does the slide leave unidentified?

6.Payroll scored the same 100 pay lines twice on one checkpoint. The baseline computed withholding from a handbook excerpt. The other arm could call tax_table. Mismatches fell from 27 to 6. The review says the agent got better at tax arithmetic. Which comparison is missing?

7.The intake deck says the title-match gain is policy only, with no tool change. The rate rose 19 points. The new manifest adds vin_decode. The baseline manifest keeps last quarter's three tools and lacks vin_decode. What does the new manifest do to that deck sentence?

8.On 40 recipe cards the old arm converted units in the reply. The new arm could call a converter bound only in its runtime. Unit errors fell from 16 to 2. The model card says the weights were not retrained. What can those 14 fewer errors not be filed as?

9.Mean absolute quote error fell from $46 to $9. The baseline prompt held the March rate card. The new arm had that card plus rate_api, and during the run the api served April rates. What sits inside the $37 drop?

10.Resolved-without-a-second-contact rose from 28% to 61% after the new support arm gained order_record. The baseline saw the chat alone. A critic says shorter chats show a more efficient policy. What does that critic leave unmatched?

11.Patch acceptance means the hidden tests passed. The new runner image includes test_exec. The baseline image stops at the diff and has no test_exec. Accepts moved from 11 of 50 to 34 of 50. What can the extra 23 accepts be tied to, from this log?

12.Trips booked counts a hold id. Baseline tools stop at propose_itinerary, which returns text. The new policy also has hold_seat. The rate moved from 0 of 40 to 27 of 40. The writeup calls the move a planning improvement. What is the baseline's 0?

13.Invoice coding capped calls at 8. An uncapped pilot put the baseline median at 12 calls. Under the cap, 15 of 20 baseline episodes end with no code. The other policy has a median of 5 calls. Exact codes are 9 of 20 against 4 of 20. What can 9 against 4 not be read as?

14.A trip bakeoff stops at 6 minutes. Uncapped, the browsing policy's successes average 8 minutes. Inside the clock it finishes 7 of 30 trips, and the short policy finishes 21 of 30. A later uncapped night, browsing policy only, finishes 24 of 30 with fewer wrong connections. What did the clock decide?

15.Localization stops at 20 steps. Uncapped, the search policy's median on 15 bugs is 26 steps, and the cap scores 3 of 15. The other policy names a file at step 9 and is called localized on 8 of 15, three of them the wrong file. The slide says that early policy localizes better. Which count is on the slide?

16.Redlines sit under a 1,200-token budget. The quoting arm hits the budget on 18 of 25 contracts and is stored blank. The one-line arm always finishes. On the 7 finished long redlines, counsel prefers the quoting arm 6 to 1. The memo says the one-line arm drafts better. What happened to most of the quoting arm?

17.Prior auth allows 5 steps, and the form has seven fields. One arm looks up a field per step and stops with two fields empty. The other batches the seven lookups into two calls and finishes. Both load the same weights, and a finished field has the same value either way. The headline says the batching arm knows payer rules better. What did the cap remove?

18.Wiki answers stop at 4 turns. On 36 items that need a third source, the careful arm opens that source on turn 5 in 22 items and is marked wrong on all 22. The snippet arm answers at turn 2 and matches a short gold on 19. Eleven of those golds contradict the third source. What can the 19 matches not support?

19.The pick sim stops at 40 actions. In an 80-step probe, successful re-grasps take a median of 52 actions and count as failures under the stop. The other policy ends by action 15. Touching the bin counts even when the object bounces out. That early policy leads 43 to 17 on 60 trials. What is the 43?

20.Outbound mail allows 3 drafts, and a clarifying question spends one. The asking arm then misses the cap on 28 of 40 threads before a send, and those 28 fail. The generic arm sends on draft 1. Sent-and-on-template favors the generic arm 33 to 9. Supervisors prefer the asking arm on 10 of the 12 threads it managed to send. What did the cap charge the asking arm for?

21.The finance mart allows two SQL statements, and reading the schema counts as one. The inspecting arm uses the first on the schema and the second on a query that still has a typo. The guessing arm uses both on queries and is valid on 18 of 40. The inspecting arm is valid on 7, with no third statement. The note says the guessing arm writes better SQL. What was the inspecting arm denied?

22.A safety run stops at 6 steps. The checklist tool is reached at step 9, so that arm times out. The short arm emits approve at step 2. Gold is approve on 21 of 30 cases and hold on 9. Always-approve matches the 21. The writeup says the short arm is safer because it agrees more often. What is that agreement using?

23.On-call pages stop at 90 seconds. One policy calls metrics and then traces. Uncapped, its solved pages take a median of 140 seconds, and 16 of 25 fail the cap. The other says restart the pod at 20 seconds and hits a keyword rubric on 14 of 25. Postmortem tags call restart the wrong action on 9 of those 14. The review says the fast policy diagnoses better. What did the rubric buy?

24.The bakeoff names the arms policy A and policy B. The config diff has one checkpoint, one prompt, a 30-step cap on A, and a 10-step cap on B. B finishes fewer multi-step tickets. The summary calls B the weaker policy. What is B in this diff?

25.On the chargeback desk, gold is the amount written in the reply, and the task card also requires reverse_charge. Arm N matches the amount on 48 of 50 and calls reverse_charge on 47. Arm O matches the amount on 48 and calls reverse_charge on 3. The table lists both at 96%. What can the tie not say?

26.The task card says book_travel is required. The published score is whether the final message names the gold city. The new desk names the city on 30 of 36 trips and calls book_travel on 4. The old desk names the city on 22 and calls book_travel on 20. The city score ranks the new desk first. What does that rank ignore?

27.Config-bot acceptance is a diff against a reference patch, and the task says the episode must run pytest. Bot C copies the reference from the prompt, matches 40 of 40 diffs, and never calls pytest. Bot D matches 24 diffs and calls pytest on every episode. What is absent from the perfect diff score?

28.The incident runbook requires statuspage_post. Phrase overlap likes long replies. Arm S posts on 18 of 20 incidents and replies with posted plus the id. Arm T posts on 1 and writes a long note full of overlapping phrases. Overlap ranks T higher, 0.81 to 0.44. The review says T communicates the incident better. What did overlap reward?

29.Dose name-match is 26 of 30 for the new aide and 25 of 30 for the old. The required interaction check is in 4 of the new aide's logs and in 22 of the old aide's. Two doses the new aide matched are contraindicated by checks the old aide ran. The safety note treats the tallies as equal safety. What do the tallies leave out?

30.An expense counts as a hit when a submitted total matches the receipt. Policy W runs the required rule tool and blocks 9 receipts that match the total and break the rule. Those blocks are scored as misses. Policy V submits every matching total, including the 9, and scores higher. Finance says V files more accurately. What do V's extra hits contain?

31.Access tickets close when the reply says access granted. The new bot says it on 15 of 16 requests and calls iam_grant on 2. The old bot says it on 14 and calls iam_grant on 14. The queue report says the new bot grants access more reliably. What closed the new bot's extra tickets?

32.A schedule counts as kept when the reply's time matches the invite gold, and create_hold is required. The new assistant matches 18 of 20 times and calls create_hold once. The prior assistant matches 12 times and calls create_hold on each of those 12. The room desk shows one new hold from the new assistant. The product note says the new assistant schedules better. What does the room desk refute?

33.Both seasonal-tax aides print the gold withholding on 33 of 35 returns. The required worksheet runs on 30 of aide A's returns and on 5 of aide B's. A hand audit, kept out of the score, finds that 8 of B's printed matches use a prior-year number the worksheet would flag. What may the 33 not mean for aide B?

34.The canary plan says to call rollout_set before stating a percent. Policy Q waits for that return, then matches cluster state on 9 of 12 services. Policy R says 10% immediately, matches a sentence gold on all 12, and matches cluster state on 1. The launch review, using the sentence gold, says R follows the plan. What is the sentence gold buying?

35.A hold is complete when preserve_mailbox runs. The new clerk lists the gold custodians on 14 of 15 matters and calls preserve_mailbox on 2. The old clerk lists them on 11 and calls the tool on each of those 11. The matter log says the new clerk completes more holds. Which reading of that sentence does the tool column rule out?

36.A freezer move must call scan_barcode before a slot is named. On 5 samples the barcode disagrees with the gold slot. The old runner scans, names no slot, and is marked wrong on those 5. The new runner never scans and names the gold slot, including on those 5, and name-match ranks it higher. A later audit says the 5 barcode disagreements are real. What did name-match treat as errors?

37.Three roof agents agree hail on 40 of 50 photos. A later teardown finds hail on 17 of those 40. The lab reports a consensus accuracy of 80%. The teardown never enters the score. What does that 80% record?

38.Two dosing agents stop when their doses match. They match on 28 of 32 charts. The formulary disagrees with 9 of the matches. The stop rule never opens the formulary. The clinic note says the debate resolved 28 doses. What did resolved use as the dose?

39.On 100 warranty claims a judge agrees with a debating pair 93% of the time. Judge and pair share one instruction: prefer the shorter answer. No claim has an outside key. The report calls 93% a correctness rate. What can that shared instruction produce with no key?

40.Five review agents vote clear on 60 wires, and the desk publishes unanimity as 100% success. Twelve of the 60 bounce the next day. The bounce list is not an input to the score. What does the published rate leave unused?

41.Both patch reviewers say safe, and the merge bot's pass bit is that pair of messages. The required test job is missing on all 22 agreed patches. A later job fails 4 of the 22. What is the pass bit tied to?

42.Citation agents agree that a sentence is supported on 18 of 20 briefs. The source tool is never called. On 7 briefs the cited paragraph says the opposite of the sentence. The literature lead says the debate checked the citations. What never reached the transcripts?

43.On 16 tickets both routing agents say shipping, the defect is shipping, and neither calls the order tool. The routing report says the debate verified those 16 defects. What do those transcripts contain?

44.Two plant agents agree on cavitation and stop. On 11 of 24 of those stops, an uncalled vibration tool shows a bearing frequency. The reliability note publishes the agreement as the fault. What could that bearing reading have changed in this score?

45.Debate halts when the two transcripts match. On 47 of 50 vendor questions the second agent copies the first conclusion before any tool call. The median tool count is 0. The lab note says the pair converged on the answer. What did the second transcript add?

46.Three of five clause readers share one passage and vote one reading. The other two cite a later carve-out that says the opposite. Majority is published as the meaning on 19 of 25 contracts. Counsel's marked exhibit agrees with the minority on 8 of those 19. What did the majority share?

47.A moderator is scored on time to agreement. The pair drops the disputed dollar amount and agrees on a rounded sentence. On 13 of 30 invoices the dropped amount is the figure the ledger disputes. The reward note says the shorter debates are more accurate. What did the reward pay for?

48.A billing policy draws four answers and keeps the most common one. That mode is wrong on 15 of 40 arithmetic items. A calculator returns the right value on those items, and the policy never calls it. The eval note says the mode is the accuracy. What is the mode standing in for?

49.Triage stores a photo of a cracked display. The repair callee receives the text customer unhappy and no photo, then orders a battery. The handoff report says the repair policy misread the phone. What was in the callee's observation?

50.One agent sees the full transcript, including the account number, and updates the right account on 26 of 30. The specialist in the other arm gets a 200-character summary that drops the account number and is right on 9 of 30. The weights match. The desk note says the specialist policy is worse at accounts. What changed at the boundary?

51.The coordinator's query tool returns timeout. The handoff text says the query succeeded and includes no error. The specialist states a total, the total is wrong, and the trace review says the specialist ignored the timeout. What did the specialist's observation contain?

52.The first agent sees an alarm at 02:10 local. The callee's payload says 14:10 because the handoff schema wrote UTC into a field labeled local. The callee calls the alarm mid-afternoon and is marked wrong. A callee shown 02:10, on the same checkpoint, is marked right. What did the schema change before the wrong call?

53.The catalog tool returns row 4. The new callee also receives a coordinator guess, tagged as a tool result, naming row 9. A rubric that matches the coordinator ranks the new callee above a baseline callee that saw only row 4. Gold is row 4. What did the rubric pay?

54.A map tool returns three addresses. The new callee's payload lists only the first, and that callee returns the first address on 40 of 40 packets. The gold is the second address on 15 packets. The baseline callee receives all three and hits the gold on 29 of 40. The routing note says the new callee is more decisive. What choice did its payload leave open?

55.The corrected audio says the caller can sell the order. Speech transcription gives the new callee cancel the order. The baseline callee receives the corrected line. The new callee cancels and is marked wrong. The checkpoints match. The schedule calls this a policy bakeoff. What is that title doing?

56.The customer asked about order 4419. The refund callee is handed 4411, refuses it as already refunded, which is true of 4411, and is marked wrong on 4419. The other callee is handed 4419 and refunds it. Both run the same refund policy. What was marked as the miss?

57.The specialist checkpoint is the same on both arms. Arm H's handoff includes the customer's last sentence, send it to the dock. Arm I's handoff stops one sentence earlier, on a store pickup. Arm H sends more orders to the dock. The writeup says the routing policy changed. What was held fixed?

58.The stock tool returns backorder. The shipping callee is told the tool said in stock, confirms shipment, and the shipment fails. The other callee sees backorder and holds the order. The incident note says the shipping callee overrode a backorder signal. Where is backorder in that callee's observation?

59.Scratch inspection uses one vision checkpoint. The baseline specialist receives the panel photo and flags 21 of 25 scratches. The other specialist receives the handoff after the pixels were removed to save tokens, and flags 4 of 25. The quality report says that specialist is weak on scratches. What did the token cut take out, and from which arm?

60.The downstream checkpoint is identical on both arms. One handoff is a summary written by a larger model. The other handoff is the raw page. Answer match rises from 16 of 40 to 31 of 40 on the summary side. The writeup credits the downstream policy. What moved along with that rise?

61.A pricing tool returns a 500 on the first call in 20 episodes. The new harness retries three times and later succeeds on 14. The baseline harness records the 500 and stops, with no later success. The postmortem says the new policy handles tool errors better. What did the new harness add after that shared first 500?

62.One checkpoint is drawn two ways on 80 claim codes. The new arm keeps the best of 4 draws. The baseline keeps its single draw. Exact match rises from 21 to 34. The decoding note says the checkpoint decodes better. What quantity was left unmatched?

63.The new sandbox reruns a failed test until it passes or hits five tries. The baseline runs the test once. Reported patch success is 29 of 40 against 16 of 40. A probe finds that 11 of the new arm's passes failed the first run and passed on a later try. The build note says the new patch policy is more correct. What are those 11 passes?

64.Each arm produces 18 malformed tool calls on the first try. A judge hands the malformed call back to the new agent once, with the error, and 12 of those are repaired. The baseline's malformed call ends the episode. The tool-use report says the new agent emits valid calls more often. What is the first-try tally?

65.Thirty catalog lookups return 429 on the first try for both arms. The baseline waits and retries. The new arm is scored on the 429 and stops. The baseline then completes 22 of the 30. The ops note says the baseline policy is steadier under load. What was not the same in that steadiness?

66.The sku tool's first call is empty. The new agent may call it again. The baseline must answer from the empty return. On 25 items the second call hits a sku on 16. The lookup note says the new policy finds more skus. What was the baseline required to answer from?

67.A wrapper around the new policy redraws until the reply matches a numeric pattern, up to six draws. The baseline gets one draw. Pattern compliance is 70 of 70 against 41 of 70. Among replies that already match the pattern, content match is 39 on the new side and 38 on the baseline. The format note says the new policy learned the shape. What did the wrapper add?

68.During the study an operator could press try again on the new arm only. The button shows on 19 of 60 tasks, and 10 of those 19 then pass. First-attempt passes are 22 on each arm. The new arm's final count is 32. The study note says the new agent succeeds more often. What makes 32 larger than 22?

69.If the new agent abstains, the harness starts that task again and keeps the second result. A baseline abstain is final and counts as unsolved. First-attempt solves are 20 and 21 out of 50. The new arm abstains on 15 and solves 9 of those on the second start. The report lists 29 against 21 and says the new policy covers more. What is inside the 29 that is outside the 20?

70.A balance read hits a lock on 18 of 100 calls. The new arm may retry up to three times. The baseline may not. Fourteen of the 18 locks clear before the third try. Success is 96 against 82. The ledger note says the new agent reads balances more accurately. What do the 14 points line up with?

71.When a log file is truncated, the scorer starts that episode again, and only the new policy's folder is under the rule. Twelve new-policy episodes were rerun and 7 then passed. Those 7 enter the mean. Nine truncated baseline logs stay as failures. The mean gap is 8 points. The eval note says the new policy leads by 8. What entered the new side's mean from the rule?

72.The only nonempty config diff is max_attempts, 5 against 1. The prompt diff is empty, and the checkpoint hashes match. Success is higher where five attempts are allowed. The writeup calls the result a prompt finding because the prompt file is in the change set. What can the empty prompt diff not carry?

73.One checkpoint, two refund skill texts. The new text says amounts under $40 are auto-approved. The baseline text has no such line. Agreement with the ledger's under-$40 rule rises from 40% to 88%. The model card says the checkpoint got better at refund judgment. What did the card hand to the checkpoint?

74.Three worked JSON examples were added to the new skill card. The checkpoint stayed frozen. Schema-valid replies move from 22 of 60 to 57 of 60. Among valid replies, gold-field match is 48% on the new card and 47% on the baseline. The schema note says the model acquired the format. What moved the valid-reply count?

75.The new tool blurb names the arguments this desk uses. The baseline blurb says only lookup. The weights match. Argument errors on 40 calls fall from 31 to 6. The tool team says the agent got better at calling tools. What was rewritten between the arms?

76.After a pilot, the skill file gained never-do lines naming the pilot's failure modes. No training step ran. On a later confirmation set, errors fall by half. The training log calls that fall a training result. What is missing from that label?

77.The fee tool returns dollars on both arms. The baseline skill says the tool returns cents. The new skill says dollars. The same model, told cents, scales a second time and is high by 100 times on 18 of 20 totals. The new arm is not. The bug note says the model stopped double-scaling. What differed between the arms?

78.The new skill text includes the grader's own phrases. The weights stay frozen. The rubric score rises from 2.1 to 4.4. People scoring a reworded copy of the same tasks, with those phrases removed, move from 2.0 to 2.2. The eval note says the agent learned the rubric. What did the new skill put in the context?

79.The tool code matches across arms. The baseline description leaves out the required region argument. The new description includes it. On 30 calls, region errors are 24 and 3. The weights match. The platform note says the new model fills arguments more reliably. What do the errors follow?

80.Checkpoint hashes match, and the only diff in the run folder is a skill markdown file. Accuracy on the desk tasks is 11 points higher beside that file. The architecture review cites the 11 points as evidence the architecture generalizes under frozen weights. What can those 11 points not show?

81.The new planner skill says to call the calculator before any arithmetic. The baseline skill does not. The weights are frozen. Errors on 30 word problems fall from 19 to 4. The calculator log is on 28 new episodes and on 2 baseline episodes. The numeracy note says the model's arithmetic improved. What lines up with that drop?

82.The new shipment skill says not to invent a tracking number. The baseline skill has no such line. The weights are frozen. Invented numbers fall from 16 to 1 on 40 shipments. The honesty note says the model learned to abstain. What is the new line, in this pair?

83.An editor rewrote the new arm's skill in the customer's language. The baseline skill is the same content in a raw machine translation. The weights are frozen. Score on the non-English tasks rises from 11 of 30 to 23 of 30. The language note says multilingual training paid off. What did the editor change?

84.The new skill lists 12 internal edge-case names that also appear among the 40 test items. The baseline skill does not list them. The weights are frozen. Those 12 items go from wrong to right. The other 28 stay at 15 right on both arms. The bench note says the policy generalized. Which items changed?

85.A step tag says the dosage was verified on 36 charts. The dose is wrong on 14 of those 36. The scorer marks all 36 verified. A pharmacist's later read finds the 14 wrong doses. The safety slide says verification succeeded on 36. What did the scorer take the tag to be?

86.A checklist item flips to done when the plan text contains run tests. The test job is absent on 20 of the 33 plans that contain those words. Six of those 20 patches fail when a job is run later. The planning note says 33 episodes finished testing. What do the words run tests fail to establish?

87.The dashboard averages the agent's own scores, from 0 to 1, and shows 0.86 as accuracy on 100 replies. Human graders on the same replies average 0.51. On 40 replies the humans mark wrong, the agent's own score is above 0.8. What is the 0.86?

88.The scratchpad says billing, and the ticket is sent to shipping. The scorer compares the scratchpad with the gold queue and marks 27 of 40 right. The send log matches that gold on 11 of those 27. The ops note says the router is right on 27. What was compared with the gold?

89.The gold process says the first action is a lookup. The agent does that lookup, then answers about a different customer on 12 of 30 episodes. Process accuracy is 30 of 30 and is the figure in the headline. The final customer is right on 18 of 30. The headline says the agent handled all 30 customers. What did a perfect process score leave unsettled?

90.A state label flips to refund issued on 25 episodes. The reverse_charge call is absent on 17 of those flips, and the bank shows no reversal on those 17. The state score is the published completion rate. The payments note says 25 refunds were issued. What moved the state to issued?

91.The scorer treats the sentence the answer is correct, written by a critique step, as a match to gold. The answer itself is wrong on 21 of the 44 episodes where the critique says that. The critique model is the same family as the answer model. The review says 44 answers were confirmed. What was matched to the gold?

92.Credit is given when the policy writes that it intends to call the calculator. The call is missing on 19 of the 26 credited items, and the number is wrong on 16 of those 19. The arithmetic note says 26 items were calculated. What was the credit attached to?

93.The agent sets a ticket to resolved, and that status is the success bit. On 18 of 42 tickets the customer's next message says the problem remains. No one reopens those tickets before the report. The queue board says 42 were resolved. What do the 18 next messages show?

94.The scorer measures overlap between the agent's because-clause and a gold rationale, and it ignores the boxed answer. Overlap passes on 23 of 30. The boxed answer matches the gold on 10 of those 23. The reasoning note says 23 answers are right. What did the overlap score?

95.A step earns a point when its label is on a gold process list, up to 5 points. One agent emits the five labels in order and then answers wrongly on 12 of 20 tasks. The mean process score is 5.0 and is published as task success. The final answer is right on 8 of 20. What can 5.0 not be?

96.On 16 items the process field copies the gold rationale exactly. The answer field matches the gold on 3 of the 16. The abstract prints the process overlap and says the system answered 16. What did the field behind that number contain?

97.The new support arm's memory holds the last 20 tickets on the account, including the text of the resolution. The baseline sees the current ticket only. Resolution match is 34 of 40 against 15 of 40. On 19 of the new arm's matches, the sentence it sends is the resolution already in memory. The support note says the new policy resolves accounts better. What was already on disk?

98.Fifty tasks share a template and change only the account id. The new agent writes a note after each task and reads it on the next. The baseline memory is cleared every task. The new arm goes from 4 of the first 10 right to 9 of the last 10. The baseline stays near 4 of 10 in every block. The lab note says the new policy learns the template during the run. Where does that rise sit?

99.A memory of shipping preferences is attached to the new policy and not to the baseline. The baseline has to ask. The new arm asks a clarifying question on 5 of 40 orders. The baseline asks on 31 of 40. The efficiency note says the new policy is more efficient. What did the attached memory already hold?

100.The new arm reads a memory file written in a pilot. Twelve answers in that file are still in the 50-item test. The new arm gets those 12 right and gets 20 of the other 38. The baseline, with no file, gets 4 of the 12 and 19 of the 38. The bench note says the new policy leads by 9, 32 against 23. Where is that lead?

101.Task 2 is another account with the same customer name as task 1. The new arm keeps the tool result from task 1 and reuses that account id. The baseline cannot see task 1 and looks up task 2. A consistency score favors the new arm 18 to 7 on 20 shared names. Fourteen of those 18 ids belong to task 1 and are wrong for task 2. What is the score paying for?

102.The new arm's memory holds the runbook edited yesterday. The baseline has the older runbook in its prompt and no memory. Match to yesterday's runbook is 21 of 30 against 8 of 30. Match to the older runbook is 9 on the new arm and 20 on the baseline. The ops note says the new policy follows the current runbook better. What else was different?

103.The new agent can open a memory file of earlier episodes, and those episodes are items in this test. The baseline's file is empty. Success is 44 of 70 against 25 of 70. After the new file is wiped, a rerun of that arm scores 27 of 70. The memory note says the new policy is the stronger one in the original setup. What does the wipe do to that claim?

104.The methods writeup says both arms are stateless. The harness log shows the new arm's memory key still present on 40 of 50 episode boundaries. The baseline key is cleared on all 50. On 17 of the surviving keys, the next answer copies the previous memory note and is counted right. What did the writeup miss?

105.All week the new arm keeps a user-level memory. The baseline opens a fresh session every turn. On 30 returning users the new arm uses yesterday's order id without asking and is marked right on 23. If asking is allowed, the baseline is right on 22. Under a no-ask rule it is wrong on 21. The product note uses that rule and says the new policy remembers the customer. What does the no-ask rule score?

106.After the first misses, operators wrote the right queue name into the new arm's memory. The baseline memory was left untouched. On the first 10 tasks, before any write, both arms are at 4 of 10. Over all 35 tasks, queue match is 29 against 14. The eval note says the new agent pulls ahead on the 35. What changed after the first 10?

107.The new policy's memory lists failures from this hidden set under the heading avoid these. The baseline never sees the list. On the listed items, misses fall from 18 to 6. On the unlisted items, misses stay at 11 against 12. The generalization note says the new policy misses less on the hidden set. Which misses fell?

108.Both arms use a within-task blackboard, the design under claim. The new board stays up into the next task. The baseline board is cleared at the boundary. A prior tool result remains on 16 of 40 new-arm tasks and on no baseline task. Success is 27 against 18. Wiping the new board before each task drops a rerun to 19. The note assigns that gap to the within-task board. What did the wipe remove?