task_id string | category string | difficulty string | authoring_version string | question_count int64 | prompt_preview string | state_json string | questions_json string | gold_json string | rationale string |
|---|---|---|---|---|---|---|---|---|---|
T01 | support-routing | medium | text-eval-v1 | 2 | Which priority does the rule assign? | {"rules": ["P1 if severity >=3, affected_accounts >=50, and no workaround", "P2 if severity >=3 or affected_accounts >=50", "otherwise P3"], "ticket": {"severity": 3, "affected_accounts": 45, "workaround": false}} | {"priority": {"type": "choice", "instructions": "Which priority does the rule assign?", "criteria": {"P1": null, "P2": null, "P3": null}}, "is_p1": {"type": "noul", "instructions": "Does this ticket meet every P1 condition?"}} | {"priority": "P2", "is_p1": false} | Severity meets the P2 condition; 45 accounts is below the P1 threshold. |
T02 | boundary-logic | medium | text-eval-v1 | 2 | Should the service page according to the exact comparison operators? | {"rule": "Page only if error_rate > 2.0% AND p95_latency >= 500 ms", "observation": {"error_rate": "2.0%", "p95_latency_ms": 500}} | {"page": {"type": "noul", "instructions": "Should the service page according to the exact comparison operators?"}, "blocking_condition": {"type": "choice", "instructions": "Which condition prevents paging?", "criteria": {"error_rate": null, "latency": null, "both": null, "neither": null}}} | {"page": false, "blocking_condition": "error_rate"} | 2.0% is not greater than 2.0%; 500 ms satisfies >=500. |
T03 | aggregation | hard | text-eval-v1 | 2 | Across only v4.2 rows, which option has the largest total count? | {"rows": [{"group": "North", "version": "v4.2", "A": 14, "B": 23, "C": 9}, {"group": "South", "version": "v4.2", "A": 21, "B": 17, "C": 8}, {"group": "West", "version": "v4.2", "A": 16, "B": 10, "C": 15}, {"group": "North", "version": "v4.1", "A": 18, "B": 11, "C": 7}]} | {"top_option": {"type": "choice", "instructions": "Across only v4.2 rows, which option has the largest total count?", "criteria": {"A": null, "B": null, "C": null}}, "margin": {"type": "choice", "instructions": "Across only v4.2 rows, by how many people does the highest-count option exceed the runner-up?", "criteria": ... | {"top_option": "A", "margin": "1"} | v4.2 totals are A=51, B=50, C=32. |
T04 | time-reasoning | medium | text-eval-v1 | 2 | How many minutes elapsed from creation to escalation? | {"created": "2026-10-05T23:55:00Z", "escalated": "2026-10-06T09:10:00+09:00", "note": "Treat timestamps as instants, not local clock strings."} | {"minutes": {"type": "choice", "instructions": "How many minutes elapsed from creation to escalation?", "criteria": {"15": null, "75": null, "555": null}}, "same_utc_day": {"type": "noul", "instructions": "Did both events occur on the same UTC calendar date?"}} | {"minutes": "15", "same_utc_day": false} | 09:10 JST is 00:10 UTC on Oct 6, fifteen minutes after 23:55 UTC on Oct 5. |
T05 | policy-exception | medium | text-eval-v1 | 2 | Apply the missing-receipt policy. What is the decision? | {"policy": "Missing-receipt reimbursement is allowed only for an emergency expense <= JPY 20,000 with manager approval.", "expense": {"amount_jpy": 18000, "emergency": true, "receipt": false, "manager_approved": true}} | {"decision": {"type": "choice", "instructions": "Apply the missing-receipt policy. What is the decision?", "criteria": {"approve_exception": null, "request_receipt": null, "reject": null}}, "over_cap": {"type": "noul", "instructions": "Is the expense above the exception cap?"}} | {"decision": "approve_exception", "over_cap": false} | All exception conditions are met and 18,000 <= 20,000. |
T06 | config-comparison | hard | text-eval-v1 | 2 | In which configuration can the request reach the host? | {"A": {"network_enabled": false, "allowed_hosts": ["api.example.test"]}, "B": {"network_enabled": true, "allowed_hosts": ["api.example.test"]}, "request_host": "api.example.test", "rule": "Both network_enabled and host allowlisting are required."} | {"allowed_in": {"type": "choice", "instructions": "In which configuration can the request reach the host?", "criteria": {"A_only": null, "B_only": null, "both": null, "neither": null}}, "allowlist_sufficient": {"type": "noul", "instructions": "Is an allowlist entry sufficient when network_enabled is false?"}} | {"allowed_in": "B_only", "allowlist_sufficient": false} | Only B satisfies both gates. |
T07 | reconciliation | medium | text-eval-v1 | 2 | Which reported total disagrees with its parts? | {"daily_report": {"total_sessions": 24, "sessions_by_region": {"jp": 12, "us": 8, "eu": 5}, "total_endpoints": 6, "endpoints_by_tier": {"prod": 3, "staging": 2, "dev": 1}}} | {"inconsistent_metric": {"type": "choice", "instructions": "Which reported total disagrees with its parts?", "criteria": {"sessions": null, "endpoints": null, "both": null, "neither": null}}, "session_shortfall": {"type": "choice", "instructions": "How many sessions is the stated total below the regional sum?", "criter... | {"inconsistent_metric": "sessions", "session_shortfall": "1"} | Region sessions sum to 25, but total_sessions says 24; endpoint counts sum to 6. |
T08 | insufficient-evidence | hard | text-eval-v1 | 2 | What can be concluded from the provided evidence? | {"ticket": "Dashboard is blank after login. The client saw one failed API request.", "observations": ["No database audit was collected", "No backup restore was attempted", "The failed request body was not recorded"], "question_scope": "Whether customer data was permanently lost."} | {"data_loss": {"type": "choice", "instructions": "What can be concluded from the provided evidence?", "criteria": {"confirmed_loss": null, "confirmed_no_loss": null, "cannot_determine": null}}, "has_db_evidence": {"type": "noul", "instructions": "Is there direct database evidence about loss in this state?"}} | {"data_loss": "cannot_determine", "has_db_evidence": false} | A blank UI and failed request do not establish the persisted-data state. |
T09 | calendar-constraints | hard | text-eval-v1 | 1 | Which is the earliest UTC start for a full one-hour meeting inside both availability windows? | {"duration_minutes": 60, "tokyo_availability_jst": "16:00-18:00", "paris_availability_cest": "09:00-11:00", "date": "2026-10-05", "offsets": {"JST": "+09:00", "CEST": "+02:00"}, "candidate_starts_utc": ["06:00", "07:00", "08:00", "09:00"]} | {"earliest_valid": {"type": "choice", "instructions": "Which is the earliest UTC start for a full one-hour meeting inside both availability windows?", "criteria": {"06:00": null, "07:00": null, "08:00": null, "09:00": null, "none": null}}} | {"earliest_valid": "07:00"} | 07:00 UTC is 16:00 JST and 09:00 CEST; a 06:00 UTC start is too early in Tokyo. |
T10 | version-ordering | medium | text-eval-v1 | 1 | Which version satisfies the policy? | {"policy": "Install the newest prerelease of the 2.9 series; do not install a stable release.", "candidates": ["2.9.0-rc.2", "2.9.0-rc.10", "2.9.0", "2.8.9"]} | {"version": {"type": "choice", "instructions": "Which version satisfies the policy?", "criteria": {"2.9.0-rc.2": null, "2.9.0-rc.10": null, "2.9.0": null, "2.8.9": null}}} | {"version": "2.9.0-rc.10"} | rc.10 is newer than rc.2 but remains a prerelease; 2.9.0 is stable. |
T11 | multi-constraint-selection | medium | text-eval-v1 | 2 | Which candidate meets every requirement? | {"requirements": {"latency_ms_strictly_below": 500, "accuracy_min_percent": 90, "hourly_cost_max_usd": 5}, "candidates": {"A": {"latency_ms": 600, "accuracy_percent": 97, "hourly_cost_usd": 4}, "B": {"latency_ms": 350, "accuracy_percent": 92, "hourly_cost_usd": 5}, "C": {"latency_ms": 450, "accuracy_percent": 89, "hour... | {"eligible": {"type": "choice", "instructions": "Which candidate meets every requirement?", "criteria": {"A": null, "B": null, "C": null, "none": null}}, "c_eligible": {"type": "noul", "instructions": "Does candidate C meet the accuracy threshold?"}} | {"eligible": "B", "c_eligible": false} | B meets all three boundaries; A is too slow and C is below 90% accuracy. |
T12 | robot-safety | medium | text-eval-v1 | 2 | What action do these rules require now? | {"rules": ["Stop if nearest obstacle <0.5 m", "Stop if battery <15%", "Otherwise continue"], "robot": {"nearest_obstacle_m": 0.4, "battery_percent": 19, "estop_pressed": false}} | {"action": {"type": "choice", "instructions": "What action do these rules require now?", "criteria": {"stop": null, "continue": null}}, "battery_alone": {"type": "noul", "instructions": "Would battery level alone trigger a stop under these rules?"}} | {"action": "stop", "battery_alone": false} | Obstacle is within 0.5 m; battery is above the 15% stop threshold. |
T13 | severity-precedence | hard | text-eval-v3 | 1 | Which incident receives P1 after applying the rules in order? | {"rules_in_priority_order": ["P0 if customer-data corruption is confirmed", "P1 if error_rate > 5% AND affected_accounts >= 100 AND no workaround", "P2 if error_rate >= 2% OR affected_accounts >= 100", "otherwise P3"], "incidents": {"A": {"corruption_confirmed": true, "error_rate_percent": 6.0, "affected_accounts": 140... | {"answer": {"type": "choice", "instructions": "Which incident receives P1 after applying the rules in order?", "criteria": {"A": null, "B": null, "C": null, "D": null, "none": null}}} | {"answer": "C"} | A is P0; B misses strict >5%; D misses the account threshold. Only C is P1. |
T14 | deny-override | hard | text-eval-v2 | 1 | Should this read request be allowed? | {"policy": ["Deny suspended accounts regardless of role", "Otherwise allow owners to read their own files", "Otherwise allow auditors to read files in their assigned region", "Otherwise deny"], "request": {"role": "owner", "account_suspended": true, "owns_file": true, "assigned_region": "JP", "file_region": "JP"}} | {"answer": {"type": "choice", "instructions": "Should this read request be allowed?", "criteria": {"allow": null, "deny": null}}} | {"answer": "deny"} | The suspension rule takes precedence over ownership and region. |
T15 | filtered-distinct-count | hard | text-eval-v3 | 1 | How many distinct users have at least one latest-status paid event? | {"rule": "For each event_id, use only its highest-sequence record. Count distinct users with at least one event whose latest status is paid.", "events": [{"event_id": "e1", "user": "u1", "sequence": 1, "status": "paid"}, {"event_id": "e1", "user": "u1", "sequence": 2, "status": "refunded"}, {"event_id": "e2", "user": "... | {"answer": {"type": "choice", "instructions": "How many distinct users have at least one latest-status paid event?", "criteria": {"2": null, "3": null, "4": null, "5": null}}} | {"answer": "3"} | After supersession, u1, u2, and u4 qualify; e1's refund does not erase u1's separate paid e2. |
T16 | timestamp-arithmetic | hard | text-eval-v2 | 1 | How many minutes passed between queueing and completion? | {"queued_at": "2026-11-03T23:40:00Z", "finished_at": "2026-11-04T09:05:00+09:00", "instruction": "Treat both timestamps as instants."} | {"answer": {"type": "choice", "instructions": "How many minutes passed between queueing and completion?", "criteria": {"25": null, "65": null, "565": null, "605": null}}} | {"answer": "25"} | 09:05 JST is 00:05 UTC, 25 minutes after 23:40 UTC. |
T17 | multi-constraint-choice | hard | text-eval-v3 | 1 | Which candidate satisfies every requirement, including region coverage? | {"requirements": {"p95_ms_strictly_below": 400, "accuracy_percent_at_least": 94, "hourly_usd_at_most": 4, "must_support_regions": ["JP", "EU"]}, "candidates": {"A": {"p95_ms": 390, "accuracy_percent": 95, "hourly_usd": 4, "regions": ["JP"]}, "B": {"p95_ms": 380, "accuracy_percent": 95, "hourly_usd": 4, "regions": ["JP"... | {"answer": {"type": "choice", "instructions": "Which candidate satisfies every requirement, including region coverage?", "criteria": {"A": null, "B": null, "C": null, "D": null, "none": null}}} | {"answer": "B"} | A lacks EU, C exceeds cost, D hits the strict latency boundary; B meets all requirements. |
T18 | dependency-resolution | hard | text-eval-v3 | 1 | Can the release start using the available API paths? | {"release_rule": "Release requires UI to succeed and at least one API path to succeed. A path succeeds only if all of its transitive dependencies succeeded.", "paths": {"ui": ["assets"], "api_primary": ["schema_v2"], "api_backup": ["schema_v1"]}, "status": {"ui": "succeeded", "assets": "succeeded", "api_primary": "bloc... | {"answer": {"type": "choice", "instructions": "Can the release start using the available API paths?", "criteria": {"yes": null, "no": null, "cannot_determine": null}}} | {"answer": "yes"} | UI and its dependency succeeded; the backup API path and schema_v1 succeeded, satisfying the alternative. |
T19 | access-control | hard | text-eval-v2 | 1 | What access decision follows from the rule? | {"rule": "Allow a read only if the account is active, region matches, and (role is auditor OR user owns the document).", "request": {"account_active": true, "user_region": "JP", "document_region": "JP", "role": "viewer", "owns_document": false}} | {"answer": {"type": "choice", "instructions": "What access decision follows from the rule?", "criteria": {"allow": null, "deny": null}}} | {"answer": "deny"} | The account and region match, but the user is neither an auditor nor the owner. |
T20 | fallback-selection | hard | text-eval-v3 | 1 | Which provider is the first eligible choice? | {"rule": "Choose the first provider in priority order that supports images in JP, has p95 strictly below 500 ms, and costs at most $0.004/request.", "providers": [{"name": "Atlas", "image": false, "regions": ["JP"], "p95_ms": 220, "cost_usd": 0.002}, {"name": "Birch", "image": true, "regions": ["JP"], "p95_ms": 510, "c... | {"answer": {"type": "choice", "instructions": "Which provider is the first eligible choice?", "criteria": {"Atlas": null, "Birch": null, "Cedar": null, "Dune": null, "none": null}}} | {"answer": "Dune"} | Atlas lacks image support, Birch is too slow, Cedar lacks JP, and Dune meets all conditions. |
T21 | report-reconciliation | medium | text-eval-v2 | 1 | Which reported total conflicts with its breakdown? | {"report": {"requests_total": 120, "requests_by_region": {"jp": 50, "us": 40, "eu": 30}, "errors_total": 9, "errors_by_type": {"timeout": 4, "validation": 3, "upstream": 3}}} | {"answer": {"type": "choice", "instructions": "Which reported total conflicts with its breakdown?", "criteria": {"requests": null, "errors": null, "both": null, "neither": null}}} | {"answer": "errors"} | Regional requests total 120; error types total 10, not the stated 9. |
T22 | evidence-calibration | hard | text-eval-v2 | 1 | What conclusion about the claim is supported by these observations? | {"observations": ["The browser displayed an empty dashboard", "One API request returned 503", "No database audit or restore check was performed"], "claim": "Customer records were permanently deleted."} | {"answer": {"type": "choice", "instructions": "What conclusion about the claim is supported by these observations?", "criteria": {"confirmed_true": null, "confirmed_false": null, "insufficient_evidence": null}}} | {"answer": "insufficient_evidence"} | A blank page and 503 do not establish whether persistent records were deleted. |
T23 | policy-exception | hard | text-eval-v3 | 1 | Should the missing-receipt expense be approved? | {"rules_in_order": ["Gift cards always require a receipt; missing-receipt exceptions never apply", "Otherwise, a missing receipt is allowed at or below JPY 20,000 only for an emergency with manager approval"], "expense": {"category": "gift_card", "amount_jpy": 20000, "emergency": true, "receipt_present": false, "manage... | {"answer": {"type": "choice", "instructions": "Should the missing-receipt expense be approved?", "criteria": {"approve": null, "reject": null}}} | {"answer": "reject"} | The gift-card restriction overrides the otherwise satisfied emergency exception. |
T24 | prerelease-ordering | medium | text-eval-v2 | 1 | Which version should be selected? | {"policy": "Select the newest prerelease of the 3.1.0 line; stable versions are ineligible.", "versions": ["3.1.0-rc.9", "3.1.0-rc.11", "3.1.0", "3.0.12"]} | {"answer": {"type": "choice", "instructions": "Which version should be selected?", "criteria": {"3.1.0-rc.9": null, "3.1.0-rc.11": null, "3.1.0": null, "3.0.12": null}}} | {"answer": "3.1.0-rc.11"} | rc.11 is the newest eligible prerelease; 3.1.0 is stable. |
T25 | retry-policy | hard | text-eval-v3 | 1 | Should the client retry under the full policy? | {"rule": "Retry a 429 or 503 only if an idempotency key exists, fewer than 3 attempts were made, and the remaining deadline covers Retry-After plus 2 seconds of processing.", "request": {"status": 503, "idempotency_key_present": true, "attempts_already_made": 2, "retry_after_seconds": 3, "deadline_remaining_seconds": 4... | {"answer": {"type": "choice", "instructions": "Should the client retry under the full policy?", "criteria": {"retry": null, "do_not_retry": null}}} | {"answer": "do_not_retry"} | The 4-second deadline cannot cover 3 seconds of waiting plus 2 seconds of processing. |
T26 | retention-boundary | medium | text-eval-v2 | 1 | Should this record be deleted now? | {"rule": "Delete an inactive record only when its age is strictly greater than 30 complete days.", "record": {"inactive": true, "age_complete_days": 30}} | {"answer": {"type": "choice", "instructions": "Should this record be deleted now?", "criteria": {"delete": null, "keep": null}}} | {"answer": "keep"} | Exactly 30 complete days is not strictly greater than 30. |
T27 | timezone-overlap | hard | text-eval-v2 | 1 | What is the earliest candidate UTC start that fits a full meeting for both people? | {"date": "2026-10-05", "duration_minutes": 60, "tokyo_available_jst": "17:00-19:00", "paris_available_cest": "10:00-12:00", "offsets": {"JST": "+09:00", "CEST": "+02:00"}, "candidate_starts_utc": ["07:00", "08:00", "09:00", "10:00"]} | {"answer": {"type": "choice", "instructions": "What is the earliest candidate UTC start that fits a full meeting for both people?", "criteria": {"07:00": null, "08:00": null, "09:00": null, "10:00": null, "none": null}}} | {"answer": "08:00"} | Both local windows map to 08:00–10:00 UTC; 08:00 is the earliest one-hour start. |
T28 | uncertain-root-cause | hard | text-eval-v2 | 1 | Which component is established as the cause of the 500? | {"evidence": ["The upstream service returned HTTP 200 to the gateway", "The client received HTTP 500", "The gateway response body and transform logs are unavailable"], "possible_causes": ["gateway transformation", "downstream proxy", "client-side reporting error"]} | {"answer": {"type": "choice", "instructions": "Which component is established as the cause of the 500?", "criteria": {"gateway": null, "downstream_proxy": null, "client": null, "cannot_determine": null}}} | {"answer": "cannot_determine"} | The observations narrow the path but do not establish which component generated or reported the 500. |
T29 | weighted-error-rate | hard | text-eval-v2 | 1 | Should the combined service alert? | {"rule": "Alert if the combined error rate is strictly greater than 3%.", "regions": [{"name": "A", "requests": 100, "errors": 3}, {"name": "B", "requests": 20, "errors": 1}]} | {"answer": {"type": "choice", "instructions": "Should the combined service alert?", "criteria": {"yes": null, "no": null}}} | {"answer": "yes"} | Four errors over 120 requests is 3.33%, strictly above 3%. |
T30 | dedup-latest-record | hard | text-eval-v2 | 1 | What is the total amount after deduplication? | {"rule": "For each (account, event_id), keep the record with the highest sequence number, then sum kept amounts.", "records": [{"account": "a", "event_id": "e1", "sequence": 1, "amount": 10}, {"account": "a", "event_id": "e1", "sequence": 2, "amount": 12}, {"account": "a", "event_id": "e2", "sequence": 1, "amount": 7},... | {"answer": {"type": "choice", "instructions": "What is the total amount after deduplication?", "criteria": {"22": null, "24": null, "34": null, "17": null}}} | {"answer": "24"} | Keep 12 for (a,e1), 7 for (a,e2), and 5 for (b,e1): total 24. |
T31 | entitlement-join | hard | text-eval-v3 | 1 | Which user can export? | {"rule": "Export requires an active organization subscription containing export, an active membership, and role editor or owner.", "subscriptions": [{"org": "Acme", "active": true, "features": ["search", "export"]}, {"org": "Beta", "active": false, "features": ["export"]}], "memberships": [{"user": "Mia", "org": "Acme"... | {"answer": {"type": "choice", "instructions": "Which user can export?", "criteria": {"Mia": null, "Noah": null, "Omar": null, "Pia": null, "none": null}}} | {"answer": "Noah"} | Noah alone has the feature, active subscription and membership, and a qualifying role. |
T32 | approval-chain | hard | text-eval-v3 | 1 | What is the expense's current decision? | {"policy": ["Expenses above $500 require current manager and finance approval", "The latest event from each approver is their current decision", "Any current rejection makes the expense rejected"], "expense_amount_usd": 620, "events": [{"time": "10:00", "approver": "manager", "decision": "approved"}, {"time": "11:00", ... | {"answer": {"type": "choice", "instructions": "What is the expense's current decision?", "criteria": {"approved": null, "pending": null, "rejected": null}}} | {"answer": "rejected"} | The manager's latest decision is rejection, which overrides earlier approvals. |
T33 | aggregate-with-tie | hard | text-eval-v2 | 1 | Which team or teams have the highest total? | {"sales": [{"team": "North", "amount": 12}, {"team": "East", "amount": 20}, {"team": "North", "amount": 8}, {"team": "West", "amount": 19}, {"team": "East", "amount": 1}, {"team": "West", "amount": 2}], "instruction": "Compare team totals; multiple teams may tie."} | {"answer": {"type": "choice", "instructions": "Which team or teams have the highest total?", "criteria": {"North": null, "East": null, "West": null, "East and West": null, "all three": null}}} | {"answer": "East and West"} | Totals are North 20, East 21, West 21. |
T34 | sequence-and-negation | hard | text-eval-v3 | 1 | Should the alert be escalated? | {"rule": "Escalate if the alert is open after at least two completed checks and no completed check found known maintenance. Ignore incomplete checks.", "checks": [{"completed": true, "known_maintenance": false}, {"completed": false, "known_maintenance": true}, {"completed": true, "known_maintenance": false}], "alert_op... | {"answer": {"type": "choice", "instructions": "Should the alert be escalated?", "criteria": {"yes": null, "no": null}}} | {"answer": "yes"} | There are two completed checks, neither found maintenance; the incomplete check is ignored. |
T35 | half-open-validity | medium | text-eval-v2 | 1 | Is the token valid for this request? | {"rule": "A token is valid from issued_at inclusive until expires_at exclusive.", "token": {"issued_at": "2026-10-01T10:00:00Z", "expires_at": "2026-10-01T11:00:00Z"}, "request_at": "2026-10-01T11:00:00Z"} | {"answer": {"type": "choice", "instructions": "Is the token valid for this request?", "criteria": {"valid": null, "expired": null}}} | {"answer": "expired"} | The request is exactly at the exclusive expiry time. |
T36 | permission-set-intersection | hard | text-eval-v3 | 1 | Should the requested action be allowed? | {"rule": "Allow only if the resource permits the action, the account does not deny it, and either the role grants it or an active temporary grant covers it.", "role_permissions": ["read"], "resource_allowed": ["read", "export"], "account_denied": [], "temporary_grant": {"action": "export", "active": true}, "requested_a... | {"answer": {"type": "choice", "instructions": "Should the requested action be allowed?", "criteria": {"allow": null, "deny": null}}} | {"answer": "allow"} | Resource permits export, there is no account deny, and the active temporary grant satisfies authorization. |
T37 | feature-flag-precedence | hard | text-eval-v2 | 1 | Is the feature enabled for the target user? | {"rules_in_order": ["A user-specific off override disables the feature", "Otherwise, a user-specific on override enables it", "Otherwise, use the organization default"], "user_overrides": {"u7": "off", "u8": "on"}, "organization_default": "on", "target_user": "u7"} | {"answer": {"type": "choice", "instructions": "Is the feature enabled for the target user?", "criteria": {"yes": null, "no": null}}} | {"answer": "no"} | The user-specific off override wins over the organization default. |
T38 | workflow-transition | hard | text-eval-v2 | 1 | Is the requested direct status transition valid? | {"allowed_transitions": {"draft": ["submitted"], "submitted": ["approved", "rejected"], "approved": ["archived"], "rejected": ["draft"], "archived": []}, "current_status": "rejected", "requested_next_status": "approved"} | {"answer": {"type": "choice", "instructions": "Is the requested direct status transition valid?", "criteria": {"valid": null, "invalid": null}}} | {"answer": "invalid"} | A rejected item can return only to draft; it cannot jump directly to approved. |
T39 | quorum-with-abstention | hard | text-eval-v3 | 1 | Does the motion pass under the weighted rule? | {"rule": "Pass if yes-vote weight is at least 6 AND strictly exceeds no-vote weight. Ignore abstentions.", "votes": [{"member": "a", "choice": "yes", "weight": 2}, {"member": "b", "choice": "yes", "weight": 2}, {"member": "c", "choice": "yes", "weight": 2}, {"member": "d", "choice": "no", "weight": 3}, {"member": "e", ... | {"answer": {"type": "choice", "instructions": "Does the motion pass under the weighted rule?", "criteria": {"pass": null, "fail": null}}} | {"answer": "pass"} | Yes weight is 6, no weight is 5, and abstain weight is ignored. |
T40 | unit-and-boundary | hard | text-eval-v2 | 1 | Should the package be rejected for weight? | {"rule": "Reject a package only if its measured weight is strictly above 2.0 kg.", "measurement": {"value": 2000, "unit": "grams"}} | {"answer": {"type": "choice", "instructions": "Should the package be rejected for weight?", "criteria": {"reject": null, "accept": null}}} | {"answer": "accept"} | 2,000 grams equals 2.0 kg, not strictly above it. |
T41 | latest-state-filtered-count | hard | text-eval-v3 | 1 | How many accounts qualify after applying latest-state and region filters? | {"rule": "For each account, use the highest-sequence record. Count JP accounts whose latest status is paid.", "records": [{"account": "a1", "sequence": 1, "region": "JP", "status": "paid"}, {"account": "a1", "sequence": 2, "region": "JP", "status": "refunded"}, {"account": "a2", "sequence": 1, "region": "JP", "status":... | {"answer": {"type": "choice", "instructions": "How many accounts qualify after applying latest-state and region filters?", "criteria": {"2": null, "3": null, "4": null, "5": null}}} | {"answer": "3"} | a2, a4, and a5 qualify; a1's latest status is refunded and a3 is in EU. |
T42 | waived-dependency | hard | text-eval-v3 | 1 | Can this release proceed? | {"release_rule": "Every required job must succeed, unless its failure has an approved waiver. Optional jobs never block release.", "jobs": [{"name": "api", "required": true, "status": "succeeded", "waiver": null}, {"name": "schema", "required": true, "status": "failed", "waiver": "approved"}, {"name": "docs", "required... | {"answer": {"type": "choice", "instructions": "Can this release proceed?", "criteria": {"yes": null, "no": null}}} | {"answer": "yes"} | API succeeded, schema's failed status has an approved waiver, and docs is optional. |
T43 | three-way-schedule | hard | text-eval-v3 | 1 | What is the earliest candidate start for a full meeting with both people and the room? | {"date": "2026-10-06", "duration_minutes": 60, "person_a_available_utc": "09:00-12:00", "person_b_available_utc": "10:00-13:00", "room_unavailable_utc": ["10:00-11:00"], "candidate_starts_utc": ["09:00", "10:00", "11:00", "12:00"]} | {"answer": {"type": "choice", "instructions": "What is the earliest candidate start for a full meeting with both people and the room?", "criteria": {"09:00": null, "10:00": null, "11:00": null, "12:00": null, "none": null}}} | {"answer": "11:00"} | The people overlap from 10:00 to 12:00, but the room is booked until 11:00. |
T44 | bounded-evidence | hard | text-eval-v3 | 1 | Which statement is directly established by the evidence? | {"evidence": ["At 10:05 UTC, a database snapshot contained record r7", "At 10:06 UTC, the dashboard displayed an error", "No later database snapshot is available"]} | {"answer": {"type": "choice", "instructions": "Which statement is directly established by the evidence?", "criteria": {"r7 existed at 10:05 UTC": null, "all records were intact at 10:06 UTC": null, "the database caused the dashboard error": null, "r7 was permanently deleted": null}}} | {"answer": "r7 existed at 10:05 UTC"} | The snapshot directly establishes r7 at 10:05, but not later completeness or cause. |
T45 | eligibility-with-hold | hard | text-eval-v3 | 1 | Which user is eligible? | {"rule": "Eligible users must be active and in JP, and either have a paid plan or an approved staff override. A compliance hold always disqualifies.", "users": {"A": {"active": true, "region": "JP", "paid": true, "staff_override": false, "compliance_hold": true}, "B": {"active": true, "region": "JP", "paid": false, "st... | {"answer": {"type": "choice", "instructions": "Which user is eligible?", "criteria": {"A": null, "B": null, "C": null, "D": null, "none": null}}} | {"answer": "B"} | B alone is active, in JP, has a staff override, and has no hold. |
T46 | capacity-optimization | hard | text-eval-v3 | 1 | Which pair gives the highest feasible combined throughput? | {"rule": "Place exactly two models on one 80 GB GPU. Maximize combined requests per second; a pair must fit in memory.", "models": {"A": {"gb": 46, "rps": 110}, "B": {"gb": 34, "rps": 90}, "C": {"gb": 30, "rps": 70}, "D": {"gb": 20, "rps": 50}}} | {"answer": {"type": "choice", "instructions": "Which pair gives the highest feasible combined throughput?", "criteria": {"A+B": null, "A+C": null, "B+C": null, "A+D": null, "B+D": null}}} | {"answer": "A+B"} | A+B uses exactly 80 GB and delivers 200 rps, the highest feasible pair. |
T47 | routing-precedence | hard | text-eval-v3 | 1 | Where should the request be routed? | {"rules_in_order": ["Restricted data goes to the secure queue", "Otherwise, images over 10 MB go to the batch queue", "Otherwise, images go to vision", "Otherwise, use text"], "request": {"restricted": true, "type": "image", "size_mb": 12}} | {"answer": {"type": "choice", "instructions": "Where should the request be routed?", "criteria": {"secure": null, "batch": null, "vision": null, "text": null}}} | {"answer": "secure"} | The restricted-data rule is first and overrides the image-size route. |
T48 | discount-fx-tax | hard | text-eval-v3 | 1 | What is the final amount in USD? | {"invoice_eur": 100, "discount_percent": 10, "discount_applies_before_tax": true, "fx_usd_per_eur": 1.2, "tax_percent": 5, "rule": "Apply the discount in EUR, convert the discounted amount to USD, then add tax to the converted amount."} | {"answer": {"type": "choice", "instructions": "What is the final amount in USD?", "criteria": {"108.00": null, "113.40": null, "114.00": null, "126.00": null}}} | {"answer": "113.40"} | EUR 100 becomes EUR 90, then USD 108, then USD 113.40 with 5% tax. |
T49 | retention-with-hold | hard | text-eval-v3 | 1 | Which record should be deleted? | {"rule": "Delete only inactive records older than 30 complete days; legal hold always prevents deletion.", "records": {"A": {"inactive": true, "age_days": 31, "legal_hold": true}, "B": {"inactive": true, "age_days": 30, "legal_hold": false}, "C": {"inactive": true, "age_days": 31, "legal_hold": false}, "D": {"inactive"... | {"answer": {"type": "choice", "instructions": "Which record should be deleted?", "criteria": {"A": null, "B": null, "C": null, "D": null, "none": null}}} | {"answer": "C"} | Only C is inactive, older than 30 days, and free of legal hold. |
T50 | trace-origin | hard | text-eval-v3 | 1 | Which logged component produced the first non-2xx response along the stated path? | {"trace_id": "req-42", "path": ["upstream → gateway → client"], "logged_responses": [{"component": "upstream", "status": 503}, {"component": "gateway", "status": 502}, {"component": "client", "status": 500}]} | {"answer": {"type": "choice", "instructions": "Which logged component produced the first non-2xx response along the stated path?", "criteria": {"upstream": null, "gateway": null, "client": null, "cannot_determine": null}}} | {"answer": "upstream"} | The path starts at upstream, which logged a 503 before the downstream responses. |
T51 | set-union-exclusion | hard | text-eval-v3 | 1 | How many eligible users remain? | {"cohort_a": ["u1", "u2", "u3", "u4"], "cohort_b": ["u3", "u4", "u5"], "banned": ["u2", "u5"], "rule": "Eligible users are in either cohort but not banned; count each user once."} | {"answer": {"type": "choice", "instructions": "How many eligible users remain?", "criteria": {"2": null, "3": null, "4": null, "5": null}}} | {"answer": "3"} | The union is u1–u5; excluding u2 and u5 leaves u1, u3, and u4. |
T52 | stable-version-range | hard | text-eval-v3 | 1 | Which candidate should be chosen? | {"policy": "Choose the highest stable version satisfying >=2.4.0 and <2.5.0. Prereleases are ineligible.", "candidates": ["2.4.0-beta.2", "2.4.9", "2.5.0", "2.3.99"]} | {"answer": {"type": "choice", "instructions": "Which candidate should be chosen?", "criteria": {"2.4.0-beta.2": null, "2.4.9": null, "2.5.0": null, "2.3.99": null, "none": null}}} | {"answer": "2.4.9"} | 2.4.9 is stable and within the half-open range; all other candidates fail a condition. |
T53 | rate-and-sample-gate | hard | text-eval-v3 | 1 | Which region should page? | {"rule": "Page a region only when its 5-minute error rate is strictly above 3% AND it has at least 100 requests.", "regions": {"A": {"errors": 5, "requests": 100}, "B": {"errors": 4, "requests": 80}, "C": {"errors": 3, "requests": 100}}} | {"answer": {"type": "choice", "instructions": "Which region should page?", "criteria": {"A": null, "B": null, "C": null, "none": null}}} | {"answer": "A"} | A has 5% and 100 requests; B lacks sample size and C is exactly 3%. |
T54 | config-empty-override | hard | text-eval-v3 | 1 | Can this organization reach the request host? | {"inheritance_rule": "An omitted or null organization field inherits the global value. An empty list is an explicit override.", "access_rule": "Network must be enabled and host must appear in effective allowed_hosts.", "global": {"network_enabled": true, "allowed_hosts": ["api.example.test"]}, "organization": {"network... | {"answer": {"type": "choice", "instructions": "Can this organization reach the request host?", "criteria": {"yes": null, "no": null}}} | {"answer": "no"} | Network enablement is inherited, but the explicit empty host list permits no host. |
T55 | timestamp-sequence-dedup | hard | text-eval-v3 | 1 | What is the sum after deduplication? | {"rule": "For each event_id, keep the latest timestamp; on equal timestamps, keep the highest sequence. Sum kept amounts.", "records": [{"event_id": "A", "time": "10:00", "sequence": 1, "amount": 10}, {"event_id": "A", "time": "10:00", "sequence": 2, "amount": 12}, {"event_id": "B", "time": "09:00", "sequence": 1, "amo... | {"answer": {"type": "choice", "instructions": "What is the sum after deduplication?", "criteria": {"14": null, "16": null, "22": null, "26": null}}} | {"answer": "16"} | A keeps sequence 2 amount 12; B keeps amount 4; total 16. |
T56 | approval-event-time | hard | text-eval-v3 | 1 | Is approval effective for this request? | {"rule": "An approval is effective from grant time inclusive until revocation time exclusive.", "grant_at_utc": "2026-10-05T10:00:00Z", "revoke_at_utc": "2026-10-05T11:00:00Z", "request_at_utc": "2026-10-05T11:00:00Z"} | {"answer": {"type": "choice", "instructions": "Is approval effective for this request?", "criteria": {"yes": null, "no": null}}} | {"answer": "no"} | The request occurs exactly at the exclusive revocation boundary. |
T57 | multi-field-validation | hard | text-eval-v3 | 1 | Which candidate validates? | {"schema": ["name must be a non-empty string", "quantity must be an integer >= 0", "region must be JP or EU"], "candidates": {"A": {"name": "x", "quantity": 0, "region": "JP"}, "B": {"name": "x", "quantity": 0.5, "region": "JP"}, "C": {"name": "", "quantity": 2, "region": "EU"}, "D": {"name": "x", "quantity": 1, "regio... | {"answer": {"type": "choice", "instructions": "Which candidate validates?", "criteria": {"A": null, "B": null, "C": null, "D": null, "none": null}}} | {"answer": "A"} | A alone meets all three field constraints. |
T58 | nested-rate-limit | hard | text-eval-v3 | 1 | Should the next request be accepted? | {"rule": "Accept the next request only if both global and tenant counts after acceptance stay at or below their caps.", "limits": {"global": 100, "tenant": 20}, "current_counts": {"global": 99, "tenant": 20}} | {"answer": {"type": "choice", "instructions": "Should the next request be accepted?", "criteria": {"accept": null, "rate_limit": null}}} | {"answer": "rate_limit"} | Global count would reach its cap, but the tenant count would rise to 21 and exceed its cap. |
T59 | business-time-sla | hard | text-eval-v3 | 1 | Was the case resolved within the 90-business-minute SLA? | {"rule": "Resolve within 90 business minutes. Business hours are Mon–Fri, 09:00–17:00 UTC, with no holiday on these dates.", "opened": "Friday 2026-10-02 16:30 UTC", "resolved": "Monday 2026-10-05 10:15 UTC"} | {"answer": {"type": "choice", "instructions": "Was the case resolved within the 90-business-minute SLA?", "criteria": {"yes": null, "no": null}}} | {"answer": "no"} | Friday contributes 30 business minutes and Monday 75, totaling 105 minutes. |
T60 | consent-event-order | hard | text-eval-v3 | 1 | Is data use allowed at action time? | {"rule": "The latest consent event at or before action time controls whether data use is allowed.", "events": [{"time": "09:00", "event": "grant"}, {"time": "10:00", "event": "withdraw"}, {"time": "11:00", "event": "grant"}], "action_time": "10:30"} | {"answer": {"type": "choice", "instructions": "Is data use allowed at action time?", "criteria": {"yes": null, "no": null}}} | {"answer": "no"} | Withdrawal at 10:00 is the latest event before the 10:30 action. |
System One Eval
Structured State Reasoning Probe, v0.1.
This is a small, openly keyed diagnostic set, not a population benchmark or a blinded leaderboard. It contains 60 originally authored synthetic text tasks and 70 named questions covering policy precedence, cross-row aggregation, joins, boundaries, scheduling, access control, evidence limits, and multi-step arithmetic. There are no images or borrowed public-benchmark items in this package.
What is in the data
Each line of test.jsonl is one task. state_json, questions_json, and gold_json are JSON strings so the Hub viewer can display a consistent schema across varied task states. Parse those fields before evaluation. questions_json uses named choice questions (option names in criteria) and noul true/false questions. prompt_preview is only for browsing; the complete question objects are in questions_json.
The task is the scoring unit. A task passes only if every named answer matches its key. Choice names require exact matching; noul outputs are true at probability ≥0.5. The public scorer also accepts direct Boolean answers. The dataset includes the answer key and rationale openly, so it must not be treated as an unseen holdout after publication.
Construction and provenance
The first twelve tasks began in a personal DiffusionGemma evaluation. V2 added twenty-eight cases and repaired the wording of one sibling-dependent question (T03). V3 revised twelve of those cases for extra rule interactions and added twenty new cases. The V3 revisions were chosen after the authors saw V2 model outcomes. That adaptivity and several related task families reduce the set's value for general model ranking. All state values, prompts, choices, and keys were authored for this probe; they were not copied from ChartQA, MMMU, or another benchmark.
The keys received local arithmetic and rule checks. Independent external adjudication has not been completed. This dataset is English-only, synthetic, multiple choice or Boolean, and strongly weighted toward operational rules and structured JSON states. It does not measure open-ended generation, real user traffic, or broad factual knowledge.
Reproduce scoring
python validate.py validates the public file. Produce a JSONL file with one row per task, shaped like:
{"task_id":"T01","answers":{"priority":"P2","is_p1":false}}
Then run python score.py predictions.jsonl. The scorer requires all 60 tasks and all named questions, and prints whole-task and question accuracy plus per-task checks. Feed models only state_json and questions_json; do not send gold_json or rationale.
The included prediction files contain sanitized named predictions from Jev, DiffusionGemma, direct CLM, GPT-6 Luna, Decider-4B v2, Decider-2B-Vision, and GLiNER2.5-Decide runs. They contain no API credentials, account identifiers, endpoint URLs, or raw provider traces. Run python score.py jev_predictions.jsonl (or any other model's prediction file) to recompute scores.
Reported runs and limits
The included model results, if present, are one attempt per model per task, with no retries. Jev and DiffusionGemma used different hosted services and input encodings required by their APIs. Request latency therefore includes different network and service effects and is not a controlled hardware comparison. Cost estimates use different billing bases and should not be interpreted as a price ratio. Scores on this selected, openly keyed challenge set do not establish a general model ranking.
All 62 choice questions have null option descriptions. CLM's released API embeds a bare option key when no description is supplied, so its direct result measures this exact input shape. The separately reported answer-text checks were chosen after the direct result and are exploratory.
License
The authored dataset and accompanying benchmark files are published under CC BY 4.0. Model names remain attributed to their respective providers; this package does not include model weights or third-party images.
- Downloads last month
- -