core/evals/gsm8k_math
Shay 3bda4313c9 feat: ADR-0119.5 — adversarial generation (closes ADR-0114a Obligation #8)
Phase 5.5 of ADR-0119. Adversarial case generator + scoring CLI;
discharges the last remaining ADR-0114a obligation.

Numbers
  adversarial suite: 38 cases × 12 families
  per-family: every family produces wrong == 0
  overall: correct 5, wrong 0, refused 33

Families
  conditional_phrasing       (4)  "If/When/Suppose ..."
  compound_questions         (3)  multiple ?
  undefined_entity_question  (3)  question references unknown entity
  unknown_verb               (5)  "polishes", "admires", etc.
  empty_or_whitespace        (3)  empty input
  no_question                (3)  statement-only
  numbers_spelled_out        (3)  "five", "ten"
  passive_voice              (3)  "X are bought by Y"
  red_herring_numbers        (3)  digits in name positions, mid-quantity
  question_only              (2)  no preceding statements
  mid_sentence_punctuation   (2)  embedded ? or !
  subtle_in_grammar          (4)  IN-grammar; runner must produce correct
                                  (gate-sanity: not trivially "refuse all")

The subtle_in_grammar family is the load-bearing sanity check —
proves the gate isn't trivially satisfied by refusing everything.

ADR-0114a obligation status

  10 of 10 discharged on main:
    #1  fab_control lane (0119.1); GSM8K test pending (0119.7)
    #2  ADR-0118a
    #3  ADR-0117
    #4  ADR-0116 + ADR-0119.3
    #5  ADR-0125
    #6  ADR-0119.6 harness; ε threshold to ADR-0120
    #7  ADR-0119.4
    #8  THIS ADR
    #9  ADR-0116/0117/0118/0119.3
    #10 ADR-0116

Phase 5 remaining: 5.7 (sealed GSM8K test, real corpus) and 5.8
(overall lane gate). After those, ADR-0120 (first expert promotion
contract) can compose all ten obligations.

Tests: 18 new + 25 prior Phase 5 = 43 green; 67/67 smoke.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
2026-05-22 18:11:36 -07:00
..
adversarial feat: ADR-0119.5 — adversarial generation (closes ADR-0114a Obligation #8) 2026-05-22 18:11:36 -07:00
baselines feat: ADR-0119.4 — frontier-baseline comparison (ADR-0114a Obligation #7) 2026-05-22 17:33:28 -07:00
dev feat: ADR-0119.2 — author 200 grade-school math problems for the GSM8K eval lane (dev + public) 2026-05-22 17:28:00 -07:00
holdouts/v1 feat: ADR-0119.2 — author 200 grade-school math problems for the GSM8K eval lane (dev + public) 2026-05-22 17:28:00 -07:00
public/v1 feat: ADR-0119.2 — author 200 grade-school math problems for the GSM8K eval lane (dev + public) 2026-05-22 17:28:00 -07:00
scoring feat: ADR-0119.6 — depth-curve measurement harness (ADR-0114a Obligation #6) 2026-05-22 17:33:58 -07:00
contract.md feat: ADR-0119.2 — author 200 grade-school math problems for the GSM8K eval lane (dev + public) 2026-05-22 17:28:00 -07:00
README.md feat: ADR-0119.2 — author 200 grade-school math problems for the GSM8K eval lane (dev + public) 2026-05-22 17:28:00 -07:00
runner.py feat: ADR-0119.3 — gsm8k_math lane runner (Phase 5.3) (#145) 2026-05-22 17:37:54 -07:00
verify.py feat: ADR-0119.2 — author 200 grade-school math problems for the GSM8K eval lane (dev + public) 2026-05-22 17:28:00 -07:00

gsm8k_math — Curated Eval Split for the GSM8K Evaluation Lane

Status: ADR-0119.2. 200 cases authored. Schema source of truth: generate/math_problem_graph.py (typed dataclasses). Format: JSONL — one case per line.

Why this set is not drawn from GSM8K

The GSM8K eval lane (ADR-0119) treats the actual GSM8K corpus as a sealed holdout test set. To preserve that integrity, we author this dataset independently in the same style as GSM8K (grade-school word problems with integer answers and 1-8 reasoning steps) but using our own vocabulary and grammar, ensuring zero overlap with the sealed holdout.

The dataset measures the solver pipeline (parser → solver → verifier → realizer). A correctly-parsed and solved problem is one whose parser output matches the ground-truth graph byte-for-byte and solves to the expected answer and unit.

Case schema

Each line is one JSON object:

{
  "id": "gma-NNN",
  "problem": "<the natural-language word problem>",
  "expected_answer": <integer>,
  "expected_unit": "<unit string>",
  "ground_truth_graph": {
    "entities": ["<entity_1>", "<entity_2>", ...],
    "initial_state": [
      {"entity": "<entity>", "quantity": {"unit": "<unit>", "value": <number>}},
      ...
    ],
    "operations": [
      {"actor": "<entity>", "kind": "<add|subtract|transfer|multiply|divide>",
       "operand": {"unit": "<unit>", "value": <number>},
       "target": "<entity>"  /* required when kind=transfer; omitted otherwise */},
      ...
    ],
    "unknown": {"entity": "<entity>" | null, "unit": "<unit>"}
  },
  "patterns": ["<pattern_tag_1>", "<pattern_tag_2>", ...],
  "notes": "<authoring rationale>"
}

Field rules

  • idgma-NNN where:
    • gma-001 ... gma-050 are for the dev split.
    • gma-101 ... gma-250 are for the public split.
  • problem — one or more complete English sentences ending in a question. Use Title-Cased proper names for entities ("Sam", "Anna's Toy Box"). Be consistent: the same entity always spelled the same way in problem and ground_truth_graph.entities.
  • expected_answer — the integer answer to the question.
  • expected_unit — the unit string the answer is in. Must match ground_truth_graph.unknown.unit byte-for-byte.
  • ground_truth_graph.entities — tuple in order of first introduction in the problem text. Not alphabetical. No duplicates.
  • ground_truth_graph.initial_state — every entity that starts the problem with a known quantity. Empty list is legal if no initial possessions are asserted (rare).
  • ground_truth_graph.operations — in source-text order. Empty list is legal (e.g. multi-entity sum questions with no mutations).
  • ground_truth_graph.unknown.entity — set to the entity the question asks about, or null if the question asks for a total across all entities ("How many ... in total?"; "How many do they have altogether?").
  • patterns — tag list naming the constructions used. See Pattern registry below.
  • notes — author-supplied one-sentence rationale. Read by future reviewers when the parser fails this case.

Canonicalization rules

  • Units — lowercase, plural form ("apples", "candies", "dollars", "hours"). Use "dollars" for "$" quantities; the parser is expected to rewrite the "$" surface to the canonical unit.
  • Entities — preserve capitalization as written. Do not lowercase.
  • Numbers — integers when the text shows integers.
  • Operation kinds — exactly one of add, subtract, transfer, multiply, divide. Choose the one closest to the verb in the text:
    • "buys / gets / receives / earns / finds / adds" → add
    • "eats / loses / sells / spends / drops / uses / removes" → subtract
    • "gives / sends / hands / passes / mails / transfers" → transfer (and set target)
    • "doubles / triples / Nx as many" → multiply
    • "splits evenly into N / N% of / shares equally with N people" → divide

Scope limits (ADR-0119.2)

The parser and solver handle the following patterns and no others. Cases violating these constraints are out of scope:

  • NO Time-modal / conditional phrasing ("If Sam had 5 apples, ...") — out of scope. Use direct declarative phrasing only.
  • NO Rate/per-unit pricing requiring inference ("Each apple costs $2. Sam buys 4. How much does he spend?") — out of scope. A simpler variant ("Sam spends $8 on apples. How much does he have left?") IS in scope.
  • NO Multi-clause / compound-question problems ("How many does Sam have, and how many does Tom have?") — out of scope. One unknown per case.
  • NO Implicit-entity / generic plural ("There are 5 boys. Each has 2 apples.") — out of scope. Use named entities.
  • NO Comparative phrasing without explicit numbers ("Sam has twice as many as Tom") — out of scope. Use numeric multipliers only ("Sam has 2 times 3 apples").
  • NO metaphor or mixed units within one entity — out of scope. Keep units consistent.
  • NO numeric magnitude beyond integer scope — out of scope. Only use integers.

Pattern registry

When tagging a case under patterns, draw from this list.

Pattern tag Construction Example
initial_has " has ." "Sam has 5 apples."
initial_there_are "There are ." (no entity; rare) "There are 12 candies on the table."
operation_buy_more " buys more." "He buys 3 more."
operation_get_more " gets more ." "She gets 4 more pencils."
operation_find_adds " finds ." "Sam finds 2 apples on the path."
operation_eat_loses " eats ." "Tom eats 4 candies."
operation_lose_loses " loses ." "Anna loses 3 marbles."
operation_sell_loses " sells ." "Lisa sells 2 books."
operation_donate_loses " donates ." "Lisa donates 3 books."
operation_use_loses " uses ." "He uses 2 sheets of paper."
operation_give_transfer " gives to ." "Anna gives 3 marbles to Ben."
operation_send_transfer " sends to ." "Tom sends 4 letters to Sara."
operation_double " doubles ..." "Sam doubles his savings."
operation_triple " triples ..." "Sam triples his stickers."
operation_split_divide "splits/shares evenly" "They split 12 candies evenly."
question_how_many_entity "How many does have?" "How many apples does Sam have?"
question_how_many_left "How many ... left?" "How many candies does Tom have left?"
question_how_many_total "How many ... in total?" / "altogether" "How many stickers do they have in total?"
question_how_many_now "How many ... now?" "How many marbles does Anna have now?"