q459 NOT_EQUAL, minidev-pg meta-llama-3-70b-instruct-2

The question

Which card costs more converted mana, "Serra Angel" or "Shrine Keeper"?

the hint the set supplies: "Serra Angel" refers to name = 'Serra Angel'; "Shrine Keeper" refers to name = 'Shrine Keeper'; card costs more converted mana when the value of convertedManaCost is greater

NOT_EQUAL states that these two statements disagree on this data under this rule. It does not state which of them is wrong.

read by hand, 2026-09-04: B harmless: one distinct row repeated; DISTINCT added by the prediction; a list under twice the answer with every row present

the gold's LIMIT 1 name; the prediction returns that one name 88 times, once per printing

a maintainer's reading of this question, out of classification.json, copied from plans/reports/prediction-mode-260904-real-predictions/classification.json. It is not a verdict and nothing above it was computed from it.

The two statements name their result columns differently. A projection is compared by position and declared type, so these names did not decide the verdict. They are part of the canonical rendering each result is hashed under, so the two result hashes differ with them.

gold: name
prediction: case

The statements

gold

SELECT name FROM cards WHERE name IN ('Serra Angel', 'Shrine Keeper') ORDER BY convertedManaCost DESC NULLS LAST LIMIT 1

convertedmanacost: descending, nulls last

sha256:d18bf362ebfed07d19ab1135279a8cb77c68d0c2a9977f53ba244c4f4e03ae82

second

SELECT 
    CASE 
        WHEN a.convertedmanacost > b.convertedmanacost THEN 'Serra Angel'
        WHEN a.convertedmanacost < b.convertedmanacost THEN 'Shrine Keeper'
        ELSE 'Both cards have the same converted mana cost'
    END 
FROM 
    cards a, cards b 
WHERE 
    a.name = 'Serra Angel' 
    AND b.name = 'Shrine Keeper';

this statement states no ordering of its own

sha256:9499f62728803096a462e5d6a1294055e9a2b019375b148e46b1a7ae8a035f4d

The marked tokens are where the two texts differ. Two statements that differ everywhere can return the same rows, and two that differ in one token can return other rows; the verdict above is read off the results.

The rows they differ in

in gold, not in the prediction, 0 rows

from counterexample.json, 0 rows, up to 25 shown per side

side	nametext
no rows

in the prediction, not in gold, 87 rows

from counterexample.json, 87 rows, up to 25 shown per side, 1 on this page

side	times	casetext
second	87	Serra Angel

What the benchmark would have said

BIRD's own check: 1

set(second_rows) == set(gold_rows), float4/float8 cells as Python float as psycopg2 returns them, numeric as Decimal

https://github.com/bird-bench/mini_dev/blob/main/evaluation/evaluation_ex.py

gold_rows: 1
second_rows: 88
gold_distinct_rows: 1
second_distinct_rows: 1

the test-suite check: 0

result_eq: equal row counts and equal column counts, each row unordered as a quick rejection, then the two equal as a list when the gold text holds ORDER BY and as a multiset otherwise, under some permutation of the columns; DISTINCT is not stripped and re-executed, and the cells are PostgreSQL's as psycopg2 returns them

ruiqi-zhong/test-suite-sql-eval, exec_eval.py, result_eq, at commit 48cb78ec: https://github.com/ruiqi-zhong/test-suite-sql-eval/blob/48cb78ecf7f610620206283846c76751b18a1326/exec_eval.py

order_matters: true
gold_rows: 1
second_rows: 88
gold_columns: 1
second_columns: 1

the class: multiplicity

The class states what makes these two results unequal under this rule, read off the two results and nothing else. It does not state which of the two statements is wrong.

The gold returned 1 rows and the prediction 88. 1 rows occur in both results at least once.

gold_types: ["text"]
second_types: ["text"]
multiset_equal: false
set_equal: true
order_equal: false
shorter_result_is_a_prefix: true

The results

gold, 1 row

from counterexample.json, 1 row

nametext
Serra Angel

second, 88 rows

from counterexample.json, 88 rows, 25 on this page

casetext
Serra Angel
Serra Angel
Serra Angel
Serra Angel
Serra Angel
Serra Angel
Serra Angel
Serra Angel
Serra Angel
Serra Angel
Serra Angel
Serra Angel
Serra Angel
Serra Angel
Serra Angel
Serra Angel
Serra Angel
Serra Angel
Serra Angel
Serra Angel
Serra Angel
Serra Angel
Serra Angel
Serra Angel
Serra Angel

The probes

A smell is a mechanical reason to read this gold statement again. It is a heuristic: it does not state that the statement is wrong, and a maintainer decides.

ordering-over-numeric-text not applicable

this statement orders by a text column holding only numbers, and ordering it as a number gives a different answer, so the gold may be sorting 9.5 above 10

no ORDER BY key resolves to a text column

what it measured

{
  "heuristic": true,
  "reason": "no ORDER BY key resolves to a text column",
  "keys": [
    {
      "key": "convertedmanacost",
      "column": "cards.convertedmanacost",
      "declared_type": "real",
      "not_applicable": "the column is not declared as text"
    }
  ]
}

arbitrary-cut quiet

this statement cuts its result at a LIMIT that does not decide which rows come back, so a different but equally correct statement can return other rows and score zero

what it measured

{
  "heuristic": true,
  "cut": 1,
  "offset": 0,
  "distinct_kept": false,
  "unbounded_sql": "SELECT name, convertedmanacost AS attestql_ordering_key_0 FROM cards WHERE name IN ('Serra Angel', 'Shrine Keeper') ORDER BY convertedmanacost DESC NULLS LAST",
  "unbounded_rows": 46,
  "projected_columns": [
    "name"
  ],
  "ordering_key_columns": [
    "attestql_ordering_key_0"
  ],
  "ordering_keys": [
    {
      "key": "convertedmanacost",
      "direction": "desc",
      "nulls": "last",
      "nulls_first_in_effect": false,
      "returned_rows_null_in_this_key": 0,
      "fires": false
    }
  ],
  "tied_at_the_cut": {
    "positions": [
      0,
      1,
      2,
      3,
      4,
      5,
      6,
      7,
      8,
      9
    ],
    "tied_rows": 44,
    "distinct_projected_answers": 1,
    "rows": [
      [
        {
          "type": "str",
          "value": "Serra Angel"
        }
      ],
      [
        {
          "type": "str",
          "value": "Serra Angel"
        }
      ],
      [
        {
          "type": "str",
          "value": "Serra Angel"
        }
      ],
      [
        {
          "type": "str",
          "value": "Serra Angel"
        }
      ],
      [
        {
          "type": "str",
          "value": "Serra Angel"
        }
      ],
      [
        {
          "type": "str",
          "value": "Serra Angel"
        }
      ],
      [
        {
          "type": "str",
          "value": "Serra Angel"
        }
      ],
      [
        {
          "type": "str",
          "value": "Serra Angel"
        }
      ],
      [
        {
          "type": "str",
          "value": "Serra Angel"
        }
      ],
      [
        {
          "type": "str",
          "value": "Serra Angel"
        }
      ]
    ]
  },
  "case": null
}

not-a-function-of-the-data quiet

rerun over the same rows in another physical order this statement gives another answer, so its result depends on how the rows are stored and not only on the data

what it measured

{
  "heuristic": true,
  "rule": "R-ORD",
  "baseline_result_hash": "sha256:d18bf362ebfed07d19ab1135279a8cb77c68d0c2a9977f53ba244c4f4e03ae82",
  "baseline_result": {
    "columns": [
      {
        "name": "name",
        "declared_type": "text"
      }
    ],
    "row_count": 1,
    "truncated": false,
    "rows_shown": 1,
    "rows": [
      [
        {
          "type": "str",
          "value": "Serra Angel"
        }
      ]
    ],
    "result_hash": "sha256:d18bf362ebfed07d19ab1135279a8cb77c68d0c2a9977f53ba244c4f4e03ae82"
  },
  "planner_statistics": {
    "cards": {
      "last_analyze": "2026-09-14 12:50:59.330452+00",
      "last_autoanalyze": null,
      "n_mod_since_analyze": 0
    }
  },
  "shuffle": {
    "seed": "1",
    "row_limit": 300000,
    "tables": [
      "cards"
    ],
    "tables_not_shuffled": [],
    "tables_skipped_for_size": {
      "laptimes": 400524,
      "legalities": 427907,
      "posthistory": 303155,
      "trans": 1056320,
      "yearmonth": 383282
    },
    "tables_not_reached_by_a_copy": {}
  },
  "shuffled_copies": {
    "run": true,
    "verdict": "equal",
    "differs": false,
    "result_hash": "sha256:d18bf362ebfed07d19ab1135279a8cb77c68d0c2a9977f53ba244c4f4e03ae82",
    "result": {
      "columns": [
        {
          "name": "name",
          "declared_type": "text"
        }
      ],
      "row_count": 1,
      "truncated": false,
      "rows_shown": 1,
      "rows": [
        [
          {
            "type": "str",
            "value": "Serra Angel"
          }
        ]
      ],
      "result_hash": "sha256:d18bf362ebfed07d19ab1135279a8cb77c68d0c2a9977f53ba244c4f4e03ae82"
    }
  },
  "plan_variant": {
    "run": false,
    "reason": "the plan variant was not asked for"
  }
}

duplicate-full-row not applicable

this statement returns the same whole row more than once and never says DISTINCT, so a statement answering the same question once per row disagrees on multiplicity alone

what it measured

{
  "heuristic": true,
  "rows": 1,
  "distinct_rows": 1,
  "repeated_rows": 0,
  "largest_repeat": 1,
  "result_bounded": false,
  "distinct_stated": false,
  "set_operation": false,
  "not_applicable": "a result of fewer than two rows has nothing to repeat"
}

The evidence records

gold: evidence-gold.json

SELECT name FROM cards WHERE name IN ('Serra Angel', 'Shrine Keeper') ORDER BY convertedManaCost DESC NULLS LAST LIMIT 1

statement read from: data/questions/mini_dev_postgresql.json
digest: sha256:d2731292f20b8d8569cd956dd747ffe1df13cd625076263e38ae9ebcef50b1ab
origin: https://bird-bench.oss-cn-beijing.aliyuncs.com/minidev.zip (sha256 cc48ba16838204e4e214512030cb572eeb5f7bcdd999bae4b9b6ff12ec13b92f, downloaded 2026-09-07), member minidev/MINIDEV/mini_dev_postgresql.json, 2024-06-19

result_hash sha256:d18bf362ebfed07d19ab1135279a8cb77c68d0c2a9977f53ba244c4f4e03ae82 recomputed from this JSON: match

record_hash sha256:8efa192744634b36092a676def4848c4e9486bbc05059e69019114caa55a56f9 recomputed from this JSON: match

the result this record holds, 1 row

from evidence-gold.json, 1 row

nametext
Serra Angel

what ran, and where

run: audit-b3dcd959-e541-44a3-bfa5-f4272d8f10ef
executed at: 2026-09-14T12:52:57.436357+00:00
data as of: 2026-09-14T12:51:43.767965+00:00
backend at checkout: PostgreSQL 16.15 (Debian 16.15-1.pgdg13+2) on aarch64-unknown-linux-gnu, compiled by gcc (Debian 14.2.0-19) 14.2.0, 64-bit | server=172.17.0.2/32:5432 | database=bird
backend that answered: PostgreSQL 16.15 (Debian 16.15-1.pgdg13+2) on aarch64-unknown-linux-gnu, compiled by gcc (Debian 14.2.0-19) 14.2.0, 64-bit | server=172.17.0.2/32:5432 | database=bird
database role: auditor
replay rule: R-ORD
question set version: sha256:d2731292f20b8d8569cd956dd747ffe1df13cd625076263e38ae9ebcef50b1ab
validator: audit:libpg_query-parse
checks run: parses_as_exactly_one_statement, the_one_statement_is_a_select, no_placeholder_without_a_bound_parameter
statement timeout: 120000 ms
rows: 1 row

the session it ran under

engine: postgresql
time_zone: Etc/UTC
date_style: ISO, MDY
interval_style: postgres
extra_float_digits: 1
database_collation: en_US.utf8
work_mem: 4096
hash_mem_multiplier: 2

recorded beside them

statement_timeout: 0
server_version: 16.15 (Debian 16.15-1.pgdg13+2)
server_version_num: 160015
transaction_read_only: on
max_parallel_workers_per_gather: 2
server_encoding: UTF8
search_path: public
datlocprovider: c
daticulocale
datcollversion: 2.41

the rendering and the data

version: attestql/audit/4
numeric_scale: 6
timestamp_format: %Y-%m-%dT%H:%M:%S.%fZ
timezone: UTC
null_rendering: NULL
encoding: utf-8
schema digest: sha256:561b171d86e9868bbc46192590383dbedd283a399e48674c300eec7f1cb5f477
source file sha256: sha256:31b1da211849d24a57c9af7636da46a5b82fc8a3ca1542bb3ebd8775e9a31cec
rows in public.cards: 56822

second: evidence-second.json

SELECT 
    CASE 
        WHEN a.convertedmanacost > b.convertedmanacost THEN 'Serra Angel'
        WHEN a.convertedmanacost < b.convertedmanacost THEN 'Shrine Keeper'
        ELSE 'Both cards have the same converted mana cost'
    END 
FROM 
    cards a, cards b 
WHERE 
    a.name = 'Serra Angel' 
    AND b.name = 'Shrine Keeper';

statement read from: data/preds-pg/predict_mini_dev_meta-llama-3-70b-instruct-2_postgresql.json
digest: sha256:32d005d4529798c3be39a5af000454728404e840ce2205506bb9514232ad6330
origin: https://raw.githubusercontent.com/bird-bench/mini_dev/b3d4bcbbae9a96934ad812551eb400c7a3b23c12/llm/exp_result/sql_output_kg/predict_mini_dev_meta-llama-3-70b-instruct-2_postgresql.json, 2024-06-19

result_hash sha256:9499f62728803096a462e5d6a1294055e9a2b019375b148e46b1a7ae8a035f4d recomputed from this JSON: match

record_hash sha256:c99471d8d47740b2eb7bbb7e56ebf1f2e0128fea45d4c27493cfa70ed9e6de65 recomputed from this JSON: match

the result this record holds, 88 rows, 50 of them here

from evidence-second.json, 88 rows, 50 on this page

casetext
Serra Angel
Serra Angel
Serra Angel
Serra Angel
Serra Angel
Serra Angel
Serra Angel
Serra Angel
Serra Angel
Serra Angel
Serra Angel
Serra Angel
Serra Angel
Serra Angel
Serra Angel
Serra Angel
Serra Angel
Serra Angel
Serra Angel
Serra Angel
Serra Angel
Serra Angel
Serra Angel
Serra Angel
Serra Angel
Serra Angel
Serra Angel
Serra Angel
Serra Angel
Serra Angel
Serra Angel
Serra Angel
Serra Angel
Serra Angel
Serra Angel
Serra Angel
Serra Angel
Serra Angel
Serra Angel
Serra Angel
Serra Angel
Serra Angel
Serra Angel
Serra Angel
Serra Angel
Serra Angel
Serra Angel
Serra Angel
Serra Angel
Serra Angel

The whole result is in evidence-second.json, beside this page.

what ran, and where

run: audit-b3dcd959-e541-44a3-bfa5-f4272d8f10ef
executed at: 2026-09-14T12:52:57.469502+00:00
data as of: 2026-09-14T12:51:43.767965+00:00
backend at checkout: PostgreSQL 16.15 (Debian 16.15-1.pgdg13+2) on aarch64-unknown-linux-gnu, compiled by gcc (Debian 14.2.0-19) 14.2.0, 64-bit | server=172.17.0.2/32:5432 | database=bird
backend that answered: PostgreSQL 16.15 (Debian 16.15-1.pgdg13+2) on aarch64-unknown-linux-gnu, compiled by gcc (Debian 14.2.0-19) 14.2.0, 64-bit | server=172.17.0.2/32:5432 | database=bird
database role: auditor
replay rule: R-ORD
question set version: sha256:d2731292f20b8d8569cd956dd747ffe1df13cd625076263e38ae9ebcef50b1ab
validator: audit:libpg_query-parse
checks run: parses_as_exactly_one_statement, the_one_statement_is_a_select, no_placeholder_without_a_bound_parameter
statement timeout: 120000 ms
rows: 88 rows

the session it ran under

engine: postgresql
time_zone: Etc/UTC
date_style: ISO, MDY
interval_style: postgres
extra_float_digits: 1
database_collation: en_US.utf8
work_mem: 4096
hash_mem_multiplier: 2

recorded beside them

statement_timeout: 0
server_version: 16.15 (Debian 16.15-1.pgdg13+2)
server_version_num: 160015
transaction_read_only: on
max_parallel_workers_per_gather: 2
server_encoding: UTF8
search_path: public
datlocprovider: c
daticulocale
datcollversion: 2.41

the rendering and the data

version: attestql/audit/4
numeric_scale: 6
timestamp_format: %Y-%m-%dT%H:%M:%S.%fZ
timezone: UTC
null_rendering: NULL
encoding: utf-8
schema digest: sha256:561b171d86e9868bbc46192590383dbedd283a399e48674c300eec7f1cb5f477
source file sha256: sha256:31b1da211849d24a57c9af7636da46a5b82fc8a3ca1542bb3ebd8775e9a31cec
rows in public.cards: 56822

Running these again

gold

re-run this statement read-only against PostgreSQL 16.15 (Debian 16.15-1.pgdg13+2) on aarch64-unknown-linux-gnu, compiled by gcc (Debian 14.2.0-19) 14.2.0, 64-bit | server=172.17.0.2/32:5432 | database=bird under the session settings and over the data this record's fixture digest names, and compare the two results under R-ORD

second

This question's run

run: audit-b3dcd959-e541-44a3-bfa5-f4272d8f10ef
server: PostgreSQL 16.15 (Debian 16.15-1.pgdg13+2) on aarch64-unknown-linux-gnu, compiled by gcc (Debian 14.2.0-19) 14.2.0, 64-bit | server=172.17.0.2/32:5432 | database=bird
question set: mini_dev_postgresql from https://bird-bench.oss-cn-beijing.aliyuncs.com/minidev.zip (sha256 cc48ba16838204e4e214512030cb572eeb5f7bcdd999bae4b9b6ff12ec13b92f, downloaded 2026-09-07), member minidev/MINIDEV/mini_dev_postgresql.json
replay rule: R-ORD

the run this question belongs to

The JSON this page was rendered from

counterexample.json attestql/audit/counterexample/2
evidence-gold.json the gold's evidence record
evidence-second.json the prediction's evidence record
smells.json attestql/audit/smells/2

The question

The statements

◀ gold

▶ second

The rows they differ in

in gold, not in the prediction, 0 rows

in the prediction, not in gold, 87 rows

What the benchmark would have said

BIRD's own check: 1

the test-suite check: 0

the class: multiplicity

The results

The probes

ordering-over-numeric-text not applicable

arbitrary-cut quiet

not-a-function-of-the-data quiet

duplicate-full-row not applicable

The evidence records

gold: evidence-gold.json

recorded beside them

second: evidence-second.json

recorded beside them

Running these again

gold

second

This question's run

The JSON this page was rendered from

gold

second