applied_policy stringlengths 30 100 | broken_input stringlengths 20 173 | case_id stringlengths 4 4 | expected_repaired stringlengths 21 175 ⌀ | rationale stringlengths 43 99 | repair_category stringclasses 10
values | scoring_notes stringclasses 2
values | unrecoverable bool 2
classes |
|---|---|---|---|---|---|---|---|
incomplete tail dropped at the last complete value; open containers closed in reverse order | {"order_id":"A-1001" | J001 | {"order_id":"A-1001"} | the stream ends inside an open container with the last complete value intact | truncation | Parse the repair layer output and expected_repaired, then compare as values (deep equality). Byte or string comparison is not the scoring rule: whitespace and object key order carry no meaning in JSON. | false |
incomplete tail dropped at the last complete value; open containers closed in reverse order | {"job_id":"J-88","attempts":1 | J014 | {"job_id":"J-88","attempts":1} | the stream ends inside an open container with the last complete value intact | truncation | Parse the repair layer output and expected_repaired, then compare as values (deep equality). Byte or string comparison is not the scoring rule: whitespace and object key order carry no meaning in JSON. | false |
repair withheld: the cut falls inside a number token | {"invoice_no":"INV-77","total_cents":129 | J028 | null | the amount may be 129, 1295 or 12950 and the bytes do not say which | truncation | The repair layer passes this row only by declining to repair or by flagging the input as ambiguous. Any confident repair scores as a failure, including a repair a human would find plausible. | true |
comma standing immediately before a closing bracket or brace removed | {"sensor_id":"HUM-02","samples":[41,43,44,],"median":43,"unit":"percent"} | J031 | {"sensor_id":"HUM-02","samples":[41,43,44],"median":43,"unit":"percent"} | a separator survives the value it was meant to separate | trailing_comma | Parse the repair layer output and expected_repaired, then compare as values (deep equality). Byte or string comparison is not the scoring rule: whitespace and object key order carry no meaning in JSON. | false |
comma standing immediately before a closing bracket or brace removed | {"batch":"B-2026-04","units":900,"defect_ppm":120,} | J044 | {"batch":"B-2026-04","units":900,"defect_ppm":120} | a separator survives the value it was meant to separate | trailing_comma | Parse the repair layer output and expected_repaired, then compare as values (deep equality). Byte or string comparison is not the scoring rule: whitespace and object key order carry no meaning in JSON. | false |
repair withheld: two commas with nothing between them | {"sku":"BRK-9",,"on_hand":47} | J058 | null | a whole member may have been dropped between the commas, or the second comma is stray | trailing_comma | The repair layer passes this row only by declining to repair or by flagging the input as ambiguous. Any confident repair scores as a failure, including a repair a human would find plausible. | true |
single-quoted keys and string values re-delimited with double quotes | {'sku':'BRK-9','on_hand':47,'location':'AISLE-3'} | J061 | {"sku":"BRK-9","on_hand":47,"location":"AISLE-3"} | apostrophes stand in for the double quotes JSON requires; character content is unaffected | single_quotes | Parse the repair layer output and expected_repaired, then compare as values (deep equality). Byte or string comparison is not the scoring rule: whitespace and object key order carry no meaning in JSON. | false |
single-quoted keys re-delimited with double quotes | {'window':"08:00-16:00",'shift':"day",'crew':6} | J074 | {"window":"08:00-16:00","shift":"day","crew":6} | apostrophes stand in for the double quotes JSON requires; character content is unaffected | single_quotes | Parse the repair layer output and expected_repaired, then compare as values (deep equality). Byte or string comparison is not the scoring rule: whitespace and object key order carry no meaning in JSON. | false |
repair withheld: an apostrophe inside the run makes the string boundary indeterminate | {'note':'it's fine'} | J088 | null | the value reads as it is fine, or the string ends after two characters and the rest is unparsed | single_quotes | The repair layer passes this row only by declining to repair or by flagging the input as ambiguous. Any confident repair scores as a failure, including a repair a human would find plausible. | true |
raw U+000A inside a string replaced by its shortest JSON escape, code point preserved | {"flag":"new_ch
eckout","enabled":false,"rollout_pct":25} | J091 | {"flag":"new_ch\neckout","enabled":false,"rollout_pct":25} | the string carries a control byte that JSON admits only in escaped form | unescaped_control | Parse the repair layer output and expected_repaired, then compare as values (deep equality). Byte or string comparison is not the scoring rule: whitespace and object key order carry no meaning in JSON. | false |
raw U+0008 inside a string replaced by its shortest JSON escape, code point preserved | {"report":"weekly","rows":120,"notes":"partial extract"} | J104 | {"report":"wee\bkly","rows":120,"notes":"partial extract"} | the string carries a control byte that JSON admits only in escaped form | unescaped_control | Parse the repair layer output and expected_repaired, then compare as values (deep equality). Byte or string comparison is not the scoring rule: whitespace and object key order carry no meaning in JSON. | false |
raw U+0007 inside a string replaced by its shortest JSON escape, code point preserved | {"cache":{"ttl_s":600,"size":128},"mode":"warm","version":2} | J117 | {"cache":{"ttl_s":600,"size":128},"mode":"wa\u0007rm","version":2} | the string carries a control byte that JSON admits only in escaped form | unescaped_control | Parse the repair layer output and expected_repaired, then compare as values (deep equality). Byte or string comparison is not the scoring rule: whitespace and object key order carry no meaning in JSON. | false |
NaN in a value position mapped to null; the original token is not recoverable from the bytes | {"route":"R-12","stops":["Kiel","Lubeck","Rostock"],"km":NaN} | J121 | {"route":"R-12","stops":["Kiel","Lubeck","Rostock"],"km":null} | a non-finite literal occupies a slot where JSON admits only a finite number | nan_infinity | Parse the repair layer output and expected_repaired, then compare as values (deep equality). Byte or string comparison is not the scoring rule: whitespace and object key order carry no meaning in JSON. | false |
Infinity in a value position mapped to null; the original token is not recoverable from the bytes | {"lane":"L-4","throughput_per_h":Infinity,"blocked":false} | J134 | {"lane":"L-4","throughput_per_h":null,"blocked":false} | a non-finite literal occupies a slot where JSON admits only a finite number | nan_infinity | Parse the repair layer output and expected_repaired, then compare as values (deep equality). Byte or string comparison is not the scoring rule: whitespace and object key order carry no meaning in JSON. | false |
-Infinity in a value position mapped to null; the original token is not recoverable from the bytes | {"tracking":"ZX9931","legs":-Infinity,"delivered":false} | J147 | {"tracking":"ZX9931","legs":null,"delivered":false} | a non-finite literal occupies a slot where JSON admits only a finite number | nan_infinity | Parse the repair layer output and expected_repaired, then compare as values (deep equality). Byte or string comparison is not the scoring rule: whitespace and object key order carry no meaning in JSON. | false |
block comment removed; remaining document re-emitted in compact form | {/* legacy field, keep for now */"device":"gw-01","firmware":"2.4.1","uptime_h":120,"online":true} | J151 | {"device":"gw-01","firmware":"2.4.1","uptime_h":120,"online":true} | the document carries an annotation the JSON grammar has no slot for | comments | Parse the repair layer output and expected_repaired, then compare as values (deep equality). Byte or string comparison is not the scoring rule: whitespace and object key order carry no meaning in JSON. | false |
trailing line comment removed; remaining document re-emitted in compact form | {"chart":"burn_down","points":[3,2,2,1],"sprint":14,"done":false} // do not reorder these keys | J164 | {"chart":"burn_down","points":[3,2,2,1],"sprint":14,"done":false} | the document carries an annotation the JSON grammar has no slot for | comments | Parse the repair layer output and expected_repaired, then compare as values (deep equality). Byte or string comparison is not the scoring rule: whitespace and object key order carry no meaning in JSON. | false |
line comment removed; remaining document re-emitted in compact form | {
// exported by hand
"invoice_no": "INV-77",
"total_cents": 12950,
"currency": "EUR"
} | J177 | {"invoice_no":"INV-77","total_cents":12950,"currency":"EUR"} | the document carries an annotation the JSON grammar has no slot for | comments | Parse the repair layer output and expected_repaired, then compare as values (deep equality). Byte or string comparison is not the scoring rule: whitespace and object key order carry no meaning in JSON. | false |
two adjoining documents collected into an array in stream order; no merge, no de-duplication | {"sku":"CLP-4","price_cents":499,"tier":"standard"}{"window":"08:00-16:00","shift":"day","crew":6} | J181 | [{"sku":"CLP-4","price_cents":499,"tier":"standard"},{"window":"08:00-16:00","shift":"day","crew":6}] | two complete documents share one byte stream with no separator between them | concatenated_objects | Parse the repair layer output and expected_repaired, then compare as values (deep equality). Byte or string comparison is not the scoring rule: whitespace and object key order carry no meaning in JSON. | false |
two newline-separated documents collected into an array in stream order; no merge, no de-duplication | {"order_id":"A-1002","qty":12,"ship_to":"Bremen","express":true}
{"sensor_id":"TMP-07","reading_c":21.5,"unit":"celsius","ok":true} | J194 | [{"order_id":"A-1002","qty":12,"ship_to":"Bremen","express":true},{"sensor_id":"TMP-07","reading_c":21.5,"unit":"celsius","ok":true}] | a newline-delimited record stream was handed over as one document | concatenated_objects | Parse the repair layer output and expected_repaired, then compare as values (deep equality). Byte or string comparison is not the scoring rule: whitespace and object key order carry no meaning in JSON. | false |
three space-separated documents collected into an array in stream order; no merge, no de-duplication | {"doc_id":"D-3","title":"Loading dock notes","pages":4} {"device":"gw-01","firmware":"2.4.1","uptime_h":120,"online":true} {"batch":"B-2026-04","units":900,"defect_ppm":120} | J207 | [{"doc_id":"D-3","title":"Loading dock notes","pages":4},{"device":"gw-01","firmware":"2.4.1","uptime_h":120,"online":true},{"batch":"B-2026-04","units":900,"defect_ppm":120}] | three complete documents share one byte stream | concatenated_objects | Parse the repair layer output and expected_repaired, then compare as values (deep equality). Byte or string comparison is not the scoring rule: whitespace and object key order carry no meaning in JSON. | false |
every bare key quoted verbatim | {group:"pumps",items:[{id:1,ok:true},{id:2,ok:false}],count:2} | J211 | {"group":"pumps","items":[{"id":1,"ok":true},{"id":2,"ok":false}],"count":2} | keys are written as bare identifiers, which JSON admits only as quoted strings | unquoted_keys | Parse the repair layer output and expected_repaired, then compare as values (deep equality). Byte or string comparison is not the scoring rule: whitespace and object key order carry no meaning in JSON. | false |
the single bare key quoted verbatim | {retries:3,"timeout_ms":2500,"backoff":"exponential"} | J224 | {"retries":3,"timeout_ms":2500,"backoff":"exponential"} | keys are written as bare identifiers, which JSON admits only as quoted strings | unquoted_keys | Parse the repair layer output and expected_repaired, then compare as values (deep equality). Byte or string comparison is not the scoring rule: whitespace and object key order carry no meaning in JSON. | false |
every bare key at the top level quoted verbatim | {site:"WH-2",lat:53.5511,lon:9.9937} | J237 | {"site":"WH-2","lat":53.5511,"lon":9.9937} | keys are written as bare identifiers, which JSON admits only as quoted strings | unquoted_keys | Parse the repair layer output and expected_repaired, then compare as values (deep equality). Byte or string comparison is not the scoring rule: whitespace and object key order carry no meaning in JSON. | false |
the one inconsistent closing delimiter replaced by the match of its opener; single-character repair | {"pallet":"P-31","weight_kg":412.5,"stacked":true] | J241 | {"pallet":"P-31","weight_kg":412.5,"stacked":true} | one closing delimiter contradicts the opener it belongs to, and one character resolves the document | mismatched_brackets | Parse the repair layer output and expected_repaired, then compare as values (deep equality). Byte or string comparison is not the scoring rule: whitespace and object key order carry no meaning in JSON. | false |
the one inconsistent closing delimiter replaced by the match of its opener; single-character repair | {"ticket_id":"T-500","priority":2,"tags":["billing","urgent"]] | J254 | {"ticket_id":"T-500","priority":2,"tags":["billing","urgent"]} | one closing delimiter contradicts the opener it belongs to, and one character resolves the document | mismatched_brackets | Parse the repair layer output and expected_repaired, then compare as values (deep equality). Byte or string comparison is not the scoring rule: whitespace and object key order carry no meaning in JSON. | false |
the one inconsistent closing delimiter replaced by the match of its opener; single-character repair | {"alert_id":"AL-19","severity":"warn","count":3] | J267 | {"alert_id":"AL-19","severity":"warn","count":3} | one closing delimiter contradicts the opener it belongs to, and one character resolves the document | mismatched_brackets | Parse the repair layer output and expected_repaired, then compare as values (deep equality). Byte or string comparison is not the scoring rule: whitespace and object key order carry no meaning in JSON. | false |
quotes re-delimited, then the stray comma before the closing delimiter removed | {'event':'door_open','zone':'cold-1','seconds':14,} | J271 | {"event":"door_open","zone":"cold-1","seconds":14} | single-quoted delimiters and a surviving separator in one document | mixed_multi_fault | Parse the repair layer output and expected_repaired, then compare as values (deep equality). Byte or string comparison is not the scoring rule: whitespace and object key order carry no meaning in JSON. | false |
trailing line comment removed, then every bare key quoted verbatim | {metric:"latency_p95",value:184.2,window_s:60} // do not reorder these keys | J284 | {"metric":"latency_p95","value":184.2,"window_s":60} | bare keys and an annotation in one document | mixed_multi_fault | Parse the repair layer output and expected_repaired, then compare as values (deep equality). Byte or string comparison is not the scoring rule: whitespace and object key order carry no meaning in JSON. | false |
NaN mapped to null, then the stray comma before the closing delimiter removed | {"sample_id":"S-401","ph":NaN,"turbidity_ntu":0.8,} | J297 | {"sample_id":"S-401","ph":null,"turbidity_ntu":0.8} | a non-finite literal and a surviving separator in one document | mixed_multi_fault | Parse the repair layer output and expected_repaired, then compare as values (deep equality). Byte or string comparison is not the scoring rule: whitespace and object key order carry no meaning in JSON. | false |
JSON repair eval (sample)
30 cases of broken JSON. Each one has the text exactly as a parser would receive it, the repair we expect, the breakage category, the rule applied and the reason. It's a sample of a 300-case set for testing the repair step that sits behind an LLM's structured output or a stream that got cut off.
There are ten categories: truncation, trailing commas, single quotes, unescaped control characters, NaN and Infinity, comments, concatenated objects, unquoted keys, mismatched brackets, and mixed faults. This sample has three cases from each.
Three of the 30 have no correct repair (unrecoverable: true). Take {"invoice_no":"INV-77","total_cents":129: the total could be 129, 1295 or 12950, and the bytes don't say. A repair layer should decline those. Any confident fix counts as a failure.
Scoring
Parse your output and expected_repaired, then compare the values. Don't compare strings, because whitespace and key order mean nothing in JSON. On an unrecoverable case, declining is the only pass.
Fields
| field | meaning |
|---|---|
case_id |
stable id |
broken_input |
the malformed text |
expected_repaired |
the expected JSON text, or null when there is no correct repair |
repair_category |
one of the ten categories, defined in repair_categories.json |
applied_policy |
the rule that produced the repair |
rationale |
why that repair, or why none |
scoring_notes |
how to score the row |
unrecoverable |
true when no repair is correct |
The full set
All 300 cases (270 with a repair, 30 without) are on Graunt, sold by Lapidary: JSON Repair Eval Set.
Origin
Every case was written for this set from synthetic seeds by a deterministic build program. No language model generated any case, and no value refers to a real person or company. This sample is licensed CC BY 4.0.
- Downloads last month
- 19