ShiftDesk v3: 47 tasks, 27 families (source github.com/nice-bills/shiftdesk v3.0.0)

README.md

@@ -2,49 +2,66 @@
 license: apache-2.0
 pretty_name: ShiftDesk
 tags: [openenv, reinforcement-learning, agents, tool-use]
+configs:
+  - config_name: default
+    data_files:
+      - split: train
+        path: tasks.jsonl
 ---
-# ShiftDesk
-
-An OpenEnv environment for OpenEnv Arena. 24 tasks: 8 families (one per arena domain) x 3 levels.
-Every reset generates a fresh instance (files, question, answer) from a seeded generator, so there are no
-fixed answer keys and this dataset holds none.
-
-| Domain | Family | What the agent does |
-|---|---|---|
-| Software engineering | module-repair | Fix 1 to 3 seeded bugs in utils.py; submit the whole file as a JSON string; hidden randomized tests against the docstring spec |
-| Industrial & physical | line-oee | Availability and OEE from a machine state-change log, production counts and breaks |
-| Natural science | kinetics-lab | Rate constant, half-life and reaction order from noisy concentration data with QC-failed rows and unit changes |
-| Office & white-collar | expense-policy | Reimbursable total under a written policy: caps, receipts, currencies, duplicates |
-| Finance & economics | loan-desk | Annuity payment, balance after k payments, interest with extra payments |
-| Math & formal reasoning | exact-count | Inclusion-exclusion counts, lattice paths on a blocked grid, digit DP modulo 1e9+7 |
-| Cybersecurity | auth-hunt | Find the brute-force that succeeded in an sshd log; decoys; /24 rotation at level 3 |
-| Media & content | subtitle-qc | SRT on-screen time, edit notes with shifts and cuts, reading-speed violations |
-
-## What it teaches
-
-- **Read the tool list, don't memorize it.** Tool names and argument names are drawn per instance
-  (for example `exec_code` with `code`, or `sandbox` with `program`; `final_answer` with `value`). The reset text
-  lists them with one example call each. An unknown tool costs a call and returns the valid names.
-- **One valid JSON object per turn, including code.** The Python tool takes source as a normal JSON string, and the
-  software task needs a whole file as the answer, so the policy has to escape newlines and quotes correctly.
-- **Use tools for data, not eyes.** Reads are capped at 2,500 characters; logs run to ~10,000, so the policy has to
-  compute in Python.
+# ShiftDesk v3
+
+An OpenEnv environment for OpenEnv Arena. 47 tasks from 27 procedural families covering all eight arena
+domains (software 5, industrial 6, natural science 5, office 5, finance 7, math and formal reasoning 9,
+cybersecurity 5, media 5). Every reset generates a fresh instance (files, question, answer) from a seeded
+generator, so there are no fixed answer keys and this dataset holds none: `tasks.jsonl` lists each task id
+with a seed-0 example question, the answer format and the file names.
+
+## What the agent does
+
+Each episode is a small workspace of text files and four plain tools, listed with one example call each in the
+first observation:
+
+- `list_files` `{}`
+- `read_file` `{"path": ..., "offset": optional first line}` (2,500 characters per call)
+- `run_python` `{"code": ...}` (Python 3 standard library, task files in the working directory, runs as `nobody`)
+- `submit` `{"answer": ...}` ends the episode
+
+Answers are small JSON objects with named fields (for example `{"best_project": "P41", "npv": 18234.55}`).
+
+| Domain | Families (levels) |
+|---|---|
+| Software engineering | change-impact (1,2): rebuild set and deterministic build order with dev-only edges; version-resolve (1,2): semver ranges, plugin constraints, yanked releases; module-repair (3) |
+| Industrial & physical | bom-explosion (1,2): BOM explosion with scrap and stock netting; tank-balance (1,2): level and alarm time with overflow and a tripped pump; line-oee (2,3) |
+| Natural science | beer-lambert (1,2): least-squares calibration with blank subtraction, an excluded standard, dilution; stoichiometry (1,2): limiting reagent, purity, process yield; kinetics-lab (3) |
+| Office & white-collar | meeting-slot (1,2): earliest common slot across time zones with a late calendar change; invoice-match (1,2): three-way match with tolerance and credit notes; expense-policy (3) |
+| Finance & economics | npv-choice (1,2): NPV by risk class with errata; bond-desk (1,2): price and modified duration; fifo-basis (1,2): FIFO realized gain with fees and a split; loan-desk (3) |
+| Math & formal reasoning | congruences (1,2): CRT with non-coprime moduli; logic-grid (1,2): unique-solution seating puzzles; shortest-route (1,2): closures and a waypoint; exact-probability (1,2): reduced fractions; exact-count (3) |
+| Cybersecurity | firewall-audit (1,2): first-match rules with an approved change; vuln-triage (1,2): numeric semver ranges and vendor backports; auth-hunt (3) |
+| Media & content | edit-timeline (1,2): EDL frames and timecode after review notes; headline-style (1,2): a two-page style guide; subtitle-qc (3) |
+
+## What it is meant to teach
+
+General desk habits rather than one tool's quirks: read every file (level-2 tasks put a correction, errata,
+change log or second page in a separate file), compute with Python instead of by eye, and submit exactly the
+requested JSON fields. Tool names are conventional and fixed; there is no ritual end action.
 
 ## Reward
 
-1.0 for a correct final answer (numeric tolerance stated per family, exact match for ids, sets and timestamps),
-otherwise 0.0. No answer, running out of calls, or an invalid reply scores 0.0.
+All or nothing: 1.0 only when every requested field is right (numeric tolerances are stated in the answer
+format; ids, lists, versions and timecodes match exactly, lists in any order unless an order is asked for),
+otherwise 0.0. No answer, an empty answer, running out of calls, or an invalid reply scores 0.0. The finish
+action `{"tool": "submit", "args": {"answer": ""}}` scores 0.0.
 
 ## Same instance for a group
 
 Rollouts that reset the same task within the same 5-minute window get the same instance (seeded from a
-secret in the image), so the five parallel rollouts of an optimizer step are compared on one problem.
+secret in the image), so the parallel rollouts of an optimizer step are compared on one problem.
 
 ## Checks
 
-Reference answers score 1 and plausible wrong answers score 0 on hundreds of seeds per task; independent
-solvers that read only the files reproduce the labels for the log, OEE and subtitle families; every bug
-mutation is detectable by the hidden tests; `openenv validate` passes; every task replays to a terminal
-reward over WebSocket. Policy code runs as `nobody` with CPU, memory and process limits.
+For every task over 100+ seeds: the reference answer scores 1 as an object and as a JSON string, plausible
+wrong answers (the answer you get by skipping the caveat file) and empty answers score 0, and generation is
+deterministic per seed. For all 19 new families an independent solver that reads only the question and files
+reproduces every label. `openenv validate` passes and every task replays to a terminal reward over WebSocket.
 
 Source: https://github.com/nice-bills/shiftdesk

tasks.jsonl

@@ -1,24 +1,47 @@
-{"task_id": "module-repair-L1", "domain": "software-engineering", "family": "module-repair", "level": 1, "budget_calls": 8, "example_question": "utils.py has 1 bug (one per function at most). test_utils.py fails. Fix utils.py so every function matches its docstring. Hidden tests check every function against its docstring. Submit the complete corrected contents of utils.py as the answer string.", "answer_format": "a string holding the whole fixed file, with newlines escaped as \\n inside the JSON string", "files": ["test_utils.py", "utils.py"]}
-{"task_id": "module-repair-L2", "domain": "software-engineering", "family": "module-repair", "level": 2, "budget_calls": 10, "example_question": "utils.py has 2 bugs (one per function at most). test_utils.py fails. Fix utils.py so every function matches its docstring. Hidden tests check every function against its docstring. Submit the complete corrected contents of utils.py as the answer string.", "answer_format": "a string holding the whole fixed file, with newlines escaped as \\n inside the JSON string", "files": ["test_utils.py", "utils.py"]}
+{"task_id": "change-impact-L1", "domain": "software-engineering", "family": "change-impact", "level": 1, "budget_calls": 8, "example_question": "Which modules must be rebuilt, and in what order?", "answer_format": "a JSON object {\"rebuild\": [\"core\", \"web\"], \"build_order\": [\"core\", \"web\"]}", "files": ["BUILD.md", "change.txt", "modules.json"]}
+{"task_id": "change-impact-L2", "domain": "software-engineering", "family": "change-impact", "level": 2, "budget_calls": 8, "example_question": "Which modules must be rebuilt, and in what order?", "answer_format": "a JSON object {\"rebuild\": [\"core\", \"web\"], \"build_order\": [\"core\", \"web\"]}", "files": ["BUILD.md", "change.txt", "modules.json"]}
+{"task_id": "version-resolve-L1", "domain": "software-engineering", "family": "version-resolve", "level": 1, "budget_calls": 8, "example_question": "Resolve one version for each package in requirements.txt using registry.json.", "answer_format": "a JSON object mapping every package to a version string, e.g. {\"flask\": \"2.3.1\", \"axios\": \"none\"}", "files": ["RESOLVER.md", "registry.json", "requirements.txt"]}
+{"task_id": "version-resolve-L2", "domain": "software-engineering", "family": "version-resolve", "level": 2, "budget_calls": 8, "example_question": "Resolve one version for each package in requirements.txt using registry.json.", "answer_format": "a JSON object mapping every package to a version string, e.g. {\"flask\": \"2.3.1\", \"axios\": \"none\"}", "files": ["RESOLVER.md", "advisory.txt", "registry.json", "requirements.txt"]}
 {"task_id": "module-repair-L3", "domain": "software-engineering", "family": "module-repair", "level": 3, "budget_calls": 12, "example_question": "utils.py has 3 bugs (one per function at most). test_utils.py fails. Fix utils.py so every function matches its docstring. Hidden tests check every function against its docstring. The visible tests do not cover every bug. Submit the complete corrected contents of utils.py as the answer string.", "answer_format": "a string holding the whole fixed file, with newlines escaped as \\n inside the JSON string", "files": ["test_utils.py", "utils.py"]}
-{"task_id": "line-oee-L1", "domain": "industrial-physical-systems", "family": "line-oee", "level": 1, "budget_calls": 8, "example_question": "From states.csv and notes.txt: what was the availability of machine M36 this shift, as a percent of the 480 planned minutes spent in RUN? Round to 1 decimal.", "answer_format": "a number such as 73.5", "files": ["notes.txt", "states.csv"]}
+{"task_id": "bom-explosion-L1", "domain": "industrial-physical-systems", "family": "bom-explosion", "level": 1, "budget_calls": 8, "example_question": "Explode the work order through bom.csv. How many units of each raw material are required?", "answer_format": "a JSON object mapping each raw material id to an integer, e.g. {\"RM-12\": 340, \"RM-40\": 95}", "files": ["bom.csv", "work_order.txt"]}
+{"task_id": "bom-explosion-L2", "domain": "industrial-physical-systems", "family": "bom-explosion", "level": 2, "budget_calls": 8, "example_question": "Explode the work order through bom.csv. How many units of each raw material are required to buy?", "answer_format": "a JSON object mapping each raw material id to an integer, e.g. {\"RM-12\": 340, \"RM-40\": 95}", "files": ["bom.csv", "stock.csv", "work_order.txt"]}
+{"task_id": "tank-balance-L1", "domain": "industrial-physical-systems", "family": "tank-balance", "level": 1, "budget_calls": 8, "example_question": "What is the tank level in litres at 11:35, and at what time (HH:MM, rounded up to the whole minute) does the level first reach the alarm level?", "answer_format": "a JSON object {\"final_level_l\": 7350, \"first_alarm\": \"08:47\"}", "files": ["flows.csv", "tank.txt"]}
+{"task_id": "tank-balance-L2", "domain": "industrial-physical-systems", "family": "tank-balance", "level": 2, "budget_calls": 8, "example_question": "What is the tank level in litres at 10:40, and at what time (HH:MM, rounded up to the whole minute) does the level first reach the alarm level?", "answer_format": "a JSON object {\"final_level_l\": 7350, \"first_alarm\": \"08:47\"}", "files": ["flows.csv", "shift_log.txt", "tank.txt"]}
 {"task_id": "line-oee-L2", "domain": "industrial-physical-systems", "family": "line-oee", "level": 2, "budget_calls": 10, "example_question": "What was the OEE of machine M58 this shift, in percent? Round to 1 decimal.", "answer_format": "a number such as 61.2", "files": ["notes.txt", "production.csv", "states.csv"]}
 {"task_id": "line-oee-L3", "domain": "industrial-physical-systems", "family": "line-oee", "level": 3, "budget_calls": 12, "example_question": "Which machine had the lowest OEE this shift, and what was it? Answer as the machine id, a space, and the OEE percent rounded to 1 decimal.", "answer_format": "a string such as \"M17 54.3\"", "files": ["notes.txt", "production.csv", "states.csv"]}
-{"task_id": "kinetics-lab-L1", "domain": "natural-science", "family": "kinetics-lab", "level": 1, "budget_calls": 8, "example_question": "From run.csv and protocol.txt, what is the rate constant k in 1/min? Give 4 significant figures.", "answer_format": "a number such as 0.04213", "files": ["protocol.txt", "run.csv"]}
-{"task_id": "kinetics-lab-L2", "domain": "natural-science", "family": "kinetics-lab", "level": 2, "budget_calls": 10, "example_question": "From run.csv and protocol.txt, what is the half-life in minutes? Give 4 significant figures.", "answer_format": "a number such as 17.42", "files": ["protocol.txt", "run.csv"]}
+{"task_id": "beer-lambert-L1", "domain": "natural-science", "family": "beer-lambert", "level": 1, "budget_calls": 10, "example_question": "Using the calibration standards, find the fitted slope (AU per uM) and the original concentration of sample U3 in uM.", "answer_format": "a JSON object {\"slope\": 0.012345, \"concentration_uM\": 41.27} (within 0.5%)", "files": ["notebook.txt", "samples.csv", "standards.csv"]}
+{"task_id": "beer-lambert-L2", "domain": "natural-science", "family": "beer-lambert", "level": 2, "budget_calls": 10, "example_question": "Using the calibration standards, find the fitted slope (AU per uM) and the original concentration of sample U5 in uM.", "answer_format": "a JSON object {\"slope\": 0.012345, \"concentration_uM\": 41.27} (within 0.5%)", "files": ["notebook.txt", "samples.csv", "standards.csv"]}
+{"task_id": "stoichiometry-L1", "domain": "natural-science", "family": "stoichiometry", "level": 1, "budget_calls": 8, "example_question": "Which reactant limits the batch, and what mass of Na2SO4 do you expect in theory? Use atomic_masses.txt for molar masses.", "answer_format": "a JSON object {\"limiting\": \"<formula>\", \"theoretical_g\": 12.34} (mass in grams, within 0.5%)", "files": ["atomic_masses.txt", "batch.txt"]}
+{"task_id": "stoichiometry-L2", "domain": "natural-science", "family": "stoichiometry", "level": 2, "budget_calls": 8, "example_question": "Which reactant limits the batch, and what mass of Na2SO4 do you expect to isolate? Use atomic_masses.txt for molar masses.", "answer_format": "a JSON object {\"limiting\": \"<formula>\", \"expected_g\": 12.34} (mass in grams, within 0.5%)", "files": ["atomic_masses.txt", "batch.txt"]}
 {"task_id": "kinetics-lab-L3", "domain": "natural-science", "family": "kinetics-lab", "level": 3, "budget_calls": 12, "example_question": "From run.csv and protocol.txt: what is the reaction order and the half-life in minutes? Answer as the order, a space, and the half-life to 4 significant figures.", "answer_format": "a string such as \"1 23.45\"", "files": ["protocol.txt", "run.csv"]}
-{"task_id": "expense-policy-L1", "domain": "office-white-collar", "family": "expense-policy", "level": 1, "budget_calls": 8, "example_question": "Using expenses.csv and policy.md, how much should Lena Weiss be reimbursed in USD for 2026-04? Round to cents.", "answer_format": "a number such as 812.40", "files": ["expenses.csv", "policy.md"]}
-{"task_id": "expense-policy-L2", "domain": "office-white-collar", "family": "expense-policy", "level": 2, "budget_calls": 10, "example_question": "Using expenses.csv and policy.md, how much should Lena Weiss be reimbursed in USD for 2026-04? Round to cents.", "answer_format": "a number such as 812.40", "files": ["expenses.csv", "policy.md", "rates.txt"]}
+{"task_id": "meeting-slot-L1", "domain": "office-white-collar", "family": "meeting-slot", "level": 1, "budget_calls": 8, "example_question": "Find the earliest start time for the call, in UTC (HH:MM).", "answer_format": "a JSON object {\"start_utc\": \"13:15\"}", "files": ["calendars.csv", "request.txt"]}
+{"task_id": "meeting-slot-L2", "domain": "office-white-collar", "family": "meeting-slot", "level": 2, "budget_calls": 8, "example_question": "Find the earliest start time for the call, in UTC (HH:MM).", "answer_format": "a JSON object {\"start_utc\": \"13:15\"}", "files": ["calendars.csv", "inbox.txt", "request.txt"]}
+{"task_id": "invoice-match-L1", "domain": "office-white-collar", "family": "invoice-match", "level": 1, "budget_calls": 8, "example_question": "Which invoices can be approved under ap_policy.txt, and what is the total payable for them?", "answer_format": "a JSON object {\"approved\": [\"INV-12345\"], \"total\": 1234.56}", "files": ["ap_policy.txt", "invoices.csv", "purchase_orders.csv", "receipts.csv"]}
+{"task_id": "invoice-match-L2", "domain": "office-white-collar", "family": "invoice-match", "level": 2, "budget_calls": 8, "example_question": "Which invoices can be approved under ap_policy.txt, and what is the total payable for them?", "answer_format": "a JSON object {\"approved\": [\"INV-12345\"], \"total\": 1234.56}", "files": ["ap_policy.txt", "credit_notes.csv", "invoices.csv", "purchase_orders.csv", "receipts.csv"]}
 {"task_id": "expense-policy-L3", "domain": "office-white-collar", "family": "expense-policy", "level": 3, "budget_calls": 12, "example_question": "Using expenses.csv and policy.md, how much should Lena Weiss be reimbursed in USD for 2026-04? Round to cents.", "answer_format": "a number such as 812.40", "files": ["expenses.csv", "policy.md", "rates.txt"]}
-{"task_id": "loan-desk-L1", "domain": "finance-economics", "family": "loan-desk", "level": 1, "budget_calls": 8, "example_question": "Read quote.txt. What is the fixed monthly payment in dollars? Round to cents.", "answer_format": "a number such as 1234.56", "files": ["quote.txt"]}
-{"task_id": "loan-desk-L2", "domain": "finance-economics", "family": "loan-desk", "level": 2, "budget_calls": 10, "example_question": "Use loans.csv and rules.txt. What is the outstanding balance of loan LN-5969 right after its 51th scheduled payment? Dollars, rounded to cents.", "answer_format": "a number such as 98765.43", "files": ["loans.csv", "rules.txt"]}
+{"task_id": "npv-choice-L1", "domain": "finance-economics", "family": "npv-choice", "level": 1, "budget_calls": 8, "example_question": "Which project has the highest net present value, and what is that NPV in USD?", "answer_format": "a JSON object {\"best_project\": \"P00\", \"npv\": 123456.78} (NPV within 1 USD)", "files": ["memo.txt", "projects.csv"]}
+{"task_id": "npv-choice-L2", "domain": "finance-economics", "family": "npv-choice", "level": 2, "budget_calls": 8, "example_question": "Which project has the highest net present value, and what is that NPV in USD?", "answer_format": "a JSON object {\"best_project\": \"P00\", \"npv\": 123456.78} (NPV within 1 USD)", "files": ["memo.txt", "projects.csv"]}
+{"task_id": "bond-desk-L1", "domain": "finance-economics", "family": "bond-desk", "level": 1, "budget_calls": 8, "example_question": "What is the price of bond BD323 per its full face value?", "answer_format": "a JSON object {\"price\": 101.2345} (within 0.01 per 100 face)", "files": ["bonds.csv", "conventions.txt"]}
+{"task_id": "bond-desk-L2", "domain": "finance-economics", "family": "bond-desk", "level": 2, "budget_calls": 8, "example_question": "For bond BD323, give its price per full face value and its modified duration in years.", "answer_format": "a JSON object {\"price\": 101.2345, \"modified_duration\": 4.5678} (price within 0.01 per 100 face, duration within 0.005)", "files": ["bonds.csv", "conventions.txt"]}
+{"task_id": "fifo-basis-L1", "domain": "finance-economics", "family": "fifo-basis", "level": 1, "budget_calls": 8, "example_question": "For NOVA only: what is the total realized gain (USD) and how many shares are still held?", "answer_format": "a JSON object {\"realized_gain\": -123.45, \"shares_held\": 40} (gain within 0.02)", "files": ["policy.txt", "trades.csv"]}
+{"task_id": "fifo-basis-L2", "domain": "finance-economics", "family": "fifo-basis", "level": 2, "budget_calls": 8, "example_question": "For KORA only: what is the total realized gain (USD) and how many shares are still held?", "answer_format": "a JSON object {\"realized_gain\": -123.45, \"shares_held\": 40} (gain within 0.02)", "files": ["policy.txt", "trades.csv"]}
 {"task_id": "loan-desk-L3", "domain": "finance-economics", "family": "loan-desk", "level": 3, "budget_calls": 12, "example_question": "Use loans.csv and rules.txt. With the extra payments, how much total interest is paid on loan LN-4578 until it is fully repaid? Dollars, rounded to cents.", "answer_format": "a number such as 12345.67", "files": ["loans.csv", "rules.txt"]}
-{"task_id": "exact-count-L1", "domain": "mathematics-or-formal-reasoning", "family": "exact-count", "level": 1, "budget_calls": 6, "example_question": "How many integers x with 1 <= x <= 1285975 are divisible by 11 or by 15, but not divisible by 3?", "answer_format": "an integer", "files": []}
-{"task_id": "exact-count-L2", "domain": "mathematics-or-formal-reasoning", "family": "exact-count", "level": 2, "budget_calls": 8, "example_question": "grid.txt is a map: '.' is open, '#' is blocked. Moving only right or down one cell at a time, how many paths go from the top-left cell to the bottom-right cell without entering a blocked cell? Give the exact integer (0 if none).", "answer_format": "an integer", "files": ["grid.txt"]}
+{"task_id": "congruences-L1", "domain": "mathematics-or-formal-reasoning", "family": "congruences", "level": 1, "budget_calls": 6, "example_question": "system.txt lists a system of congruences. Give the smallest non-negative solution x and the modulus M such that the solutions are exactly x + k*M.", "answer_format": "a JSON object {\"x\": 1234, \"modulus\": 5040} with integers", "files": ["system.txt"]}
+{"task_id": "congruences-L2", "domain": "mathematics-or-formal-reasoning", "family": "congruences", "level": 2, "budget_calls": 6, "example_question": "system.txt lists a system of congruences. Give the smallest non-negative solution x and the modulus M such that the solutions are exactly x + k*M, and the count it asks for.", "answer_format": "a JSON object {\"x\": 1234, \"modulus\": 5040, \"count_in_range\": 7} with integers", "files": ["system.txt"]}
+{"task_id": "logic-grid-L1", "domain": "mathematics-or-formal-reasoning", "family": "logic-grid", "level": 1, "budget_calls": 8, "example_question": "Solve the puzzle in puzzle.txt. Give each person's city.", "answer_format": "a JSON object mapping every person to a value, e.g. {\"Goran\": \"Quito\", ...}", "files": ["puzzle.txt"]}
+{"task_id": "logic-grid-L2", "domain": "mathematics-or-formal-reasoning", "family": "logic-grid", "level": 2, "budget_calls": 8, "example_question": "Solve the puzzle in puzzle.txt. Give each person's pet.", "answer_format": "a JSON object mapping every person to a value, e.g. {\"Goran\": \"cat\", ...}", "files": ["puzzle.txt"]}
+{"task_id": "shortest-route-L1", "domain": "mathematics-or-formal-reasoning", "family": "shortest-route", "level": 1, "budget_calls": 8, "example_question": "Roads in roads.csv are two-way. Find the fastest route from J to G. Give the total minutes and the route as a list of nodes, start and end included.", "answer_format": "a JSON object {\"minutes\": 37, \"route\": [\"A\", \"D\", \"C\"]}", "files": ["roads.csv"]}
+{"task_id": "shortest-route-L2", "domain": "mathematics-or-formal-reasoning", "family": "shortest-route", "level": 2, "budget_calls": 8, "example_question": "Roads in roads.csv are two-way. Find the fastest route from D to E that passes through I, avoiding closed roads (see notice.txt). Give the total minutes and the route as a list of nodes, start and end included.", "answer_format": "a JSON object {\"minutes\": 37, \"route\": [\"A\", \"D\", \"C\"]}", "files": ["notice.txt", "roads.csv"]}
+{"task_id": "exact-probability-L1", "domain": "mathematics-or-formal-reasoning", "family": "exact-probability", "level": 1, "budget_calls": 6, "example_question": "Answer the question in problem.txt as an exact reduced fraction.", "answer_format": "a JSON object {\"probability\": \"a/b\"} with the fraction in lowest terms", "files": ["problem.txt"]}
+{"task_id": "exact-probability-L2", "domain": "mathematics-or-formal-reasoning", "family": "exact-probability", "level": 2, "budget_calls": 6, "example_question": "Answer the question in problem.txt as an exact reduced fraction.", "answer_format": "a JSON object {\"probability\": \"a/b\"} with the fraction in lowest terms", "files": ["problem.txt"]}
 {"task_id": "exact-count-L3", "domain": "mathematics-or-formal-reasoning", "family": "exact-count", "level": 3, "budget_calls": 8, "example_question": "Count the 49-digit positive integers (no leading zero) whose digits sum to 292 and that never use the digit 6. Give the count modulo 1000000007.", "answer_format": "an integer", "files": []}
-{"task_id": "auth-hunt-L1", "domain": "cybersecurity", "family": "auth-hunt", "level": 1, "budget_calls": 8, "example_question": "Which source IP has the most failed password attempts in auth.log? Answer with the IP address. The log is auth.log.", "answer_format": "a string", "files": ["auth.log"]}
-{"task_id": "auth-hunt-L2", "domain": "cybersecurity", "family": "auth-hunt", "level": 2, "budget_calls": 10, "example_question": "Find the compromise: a source IP with at least 5 failed passwords for one user within any 10-minute span, followed by an accepted login from that same IP for that same user no more than 10 minutes after its last failure. Exactly one IP matches. Answer with that IP. The log is auth.log.", "answer_format": "a string", "files": ["auth.log"]}
+{"task_id": "firewall-audit-L1", "domain": "cybersecurity", "family": "firewall-audit", "level": 1, "budget_calls": 8, "example_question": "Which packets in packets.csv does the live firewall allow?", "answer_format": "a JSON object {\"allowed\": [\"P01\", \"P07\"]} listing every allowed packet id", "files": ["packets.csv", "rules.txt"]}
+{"task_id": "firewall-audit-L2", "domain": "cybersecurity", "family": "firewall-audit", "level": 2, "budget_calls": 8, "example_question": "Which packets in packets.csv does the live firewall allow? Apply change_log.txt to rules.txt first.", "answer_format": "a JSON object {\"allowed\": [\"P01\", \"P07\"]} listing every allowed packet id", "files": ["change_log.txt", "packets.csv", "rules.txt"]}
+{"task_id": "vuln-triage-L1", "domain": "cybersecurity", "family": "vuln-triage", "level": 1, "budget_calls": 8, "example_question": "Versions are semantic (major.minor.patch, compared numerically). Which hosts run at least one package version covered by an advisory? Also give how many hosts that is.", "answer_format": "a JSON object {\"hosts\": [\"app-03\", \"web-01\"], \"count\": 2}", "files": ["advisories.json", "inventory.csv"]}
+{"task_id": "vuln-triage-L2", "domain": "cybersecurity", "family": "vuln-triage", "level": 2, "budget_calls": 8, "example_question": "Versions are semantic (major.minor.patch, compared numerically). Which hosts run at least one package version covered by an advisory? Also give how many hosts that is.", "answer_format": "a JSON object {\"hosts\": [\"app-03\", \"web-01\"], \"count\": 2}", "files": ["advisories.json", "inventory.csv", "vendor_notes.txt"]}
 {"task_id": "auth-hunt-L3", "domain": "cybersecurity", "family": "auth-hunt", "level": 3, "budget_calls": 12, "example_question": "An attacker rotated source addresses inside one /24 network. Treat all IPs in the same /24 as one source. Find the /24 with at least 5 failed passwords for one user within any 10-minute span, followed within 10 minutes of its last failure by an accepted login for that same user from the same /24. Answer in CIDR form, like 198.51.100.0/24. The log is auth.log.", "answer_format": "a string", "files": ["auth.log"]}
-{"task_id": "subtitle-qc-L1", "domain": "media-content-production", "family": "subtitle-qc", "level": 1, "budget_calls": 8, "example_question": "episode.srt is a subtitle file; each line starts with the speaker's name. What is the total time, in seconds, that ANA's subtitles are on screen? Give 3 decimals.", "answer_format": "a number such as 41.250", "files": ["episode.srt"]}
-{"task_id": "subtitle-qc-L2", "domain": "media-content-production", "family": "subtitle-qc", "level": 2, "budget_calls": 10, "example_question": "Apply notes.txt to episode.srt. What is the end timestamp of the last remaining cue after the edit? Answer in SRT form HH:MM:SS,mmm.", "answer_format": "a string such as \"00:03:12,480\"", "files": ["episode.srt", "notes.txt"]}
+{"task_id": "edit-timeline-L1", "domain": "media-content-production", "family": "edit-timeline", "level": 1, "budget_calls": 8, "example_question": "Clips play back to back in the listed order. What is the total program length in frames, and as a timecode HH:MM:SS:FF at the timeline frame rate?", "answer_format": "a JSON object {\"frames\": 4321, \"timecode\": \"00:02:52:21\"}", "files": ["timeline.edl"]}
+{"task_id": "edit-timeline-L2", "domain": "media-content-production", "family": "edit-timeline", "level": 2, "budget_calls": 8, "example_question": "After applying review_notes.txt: clips play back to back in the listed order. What is the total program length in frames, and as a timecode HH:MM:SS:FF at the timeline frame rate?", "answer_format": "a JSON object {\"frames\": 4321, \"timecode\": \"00:02:52:21\"}", "files": ["review_notes.txt", "timeline.edl"]}
+{"task_id": "headline-style-L1", "domain": "media-content-production", "family": "headline-style", "level": 1, "budget_calls": 8, "example_question": "Which headlines in headlines.tsv break the house style guide? Read every page of the guide.", "answer_format": "a JSON object {\"violations\": [\"H02\", \"H09\"]} listing every breaking headline id", "files": ["headlines.tsv", "style_guide_p1.txt"]}
+{"task_id": "headline-style-L2", "domain": "media-content-production", "family": "headline-style", "level": 2, "budget_calls": 8, "example_question": "Which headlines in headlines.tsv break the house style guide? Read every page of the guide.", "answer_format": "a JSON object {\"violations\": [\"H02\", \"H09\"]} listing every breaking headline id", "files": ["headlines.tsv", "style_guide_p1.txt", "style_guide_p2.txt"]}
 {"task_id": "subtitle-qc-L3", "domain": "media-content-production", "family": "subtitle-qc", "level": 3, "budget_calls": 12, "example_question": "Check episode.srt against style.txt. Which cue numbers break the reading-speed rule? Answer with the cue numbers in increasing order, separated by commas (empty string if none).", "answer_format": "a string such as \"3,8,15\"", "files": ["episode.srt", "style.txt"]}