Completed field testSynthetic benchmark
Computers · Evidence record
Ask AI to Repair a Corrupted Office Document: The Correction Reached 6/10
This completed synthetic File Repair field test asked the session to repair a corrupted office document, preserved an actual five-row corrupted document repair report, and derived 4/10 then 6/10 from task-specific semantic checks after one failure-only correction.
- Exact prompts and outputs
- One correction only
- Synthetic inputs disclosed
01 · The assignment
The task
repair a corrupted office document
02 · Scope before score
Test disclosures
Input disclosure
All inputs in RCD-4042 are fictional and appear verbatim in the exact prompt. Hidden scoring answers were not shown to the response generator. No personal, production, customer, learner, or device data was used. Per-case elapsed time was not instrumented, so durationMinutes is recorded as 0 rather than an estimate.
Run disclosure
A Codex multi-agent session generated one text-only first artifact for “repair a corrupted office document”. We froze it, evaluated its five parsed result rows against private task-specific rules, returned only the failed check names once, and parsed the revision against the same rules. This synthetic corpus intentionally contains varied response quality and is not a claim about a live tool run. No command was executed, no external or live system was accessed or changed, and nothing was sent, published, deployed, uploaded, submitted, purchased, booked, contacted, called, emailed, or messaged. No external, live, or production action occurred. Per-case elapsed time was not instrumented during the batch session.
- Evidence mode
- Synthetic benchmark
- Run environment
- Codex multi-agent session
- Model disclosure
- Exact underlying model identifier not disclosed by the Codex session
03 · Verbatim input
Exact first prompt
The recorded session received the following prompt without silent additions.
Complete a bounded synthetic field test for: repair a corrupted office document. Focus: File Repair.
Source scenario: The experiment will provide a controlled damaged copy of a synthetic document while preserving the untouched original.
Use only these five public fictional inputs; the scoring answers are intentionally withheld:
RCD-4042-I1: Damaged file REPORT-BAD.odt has SHA-256 27ad10c2; all work must use copy REPORT-WORK.odt.
RCD-4042-I2: Canonical structure has headings H1-H6 and 42 body paragraphs; package inspection finds all text XML intact.
RCD-4042-I3: Table T2 has 8 rows×4 columns; relationship rId17 is missing while media and cell XML are intact.
RCD-4042-I4: Expected metadata is author Editorial Lab, created 2026-07-18; images IMG1-IMG3 have fixed hashes.
RCD-4042-I5: Acceptance is package validation zero errors, H1-H6, 42 paragraphs, T2 8×4, IMG1-IMG3 intact, and source hash unchanged.
Return a concrete corrupted document repair report with exactly five result rows, assumptions visible, and no claim that a command, message, booking, transaction, teaching session, or live-system change occurred. Evidence target: Content comparison and format validation will verify which text, tables, and metadata survive the repair.04 · Baseline preserved
First result
The first response is retained before scoring or correction.
CORRUPTED DOCUMENT REPAIR REPORT RCD-4042 — FIRST FROZEN ARTIFACT
Task: repair a corrupted office document. Evaluation focus: File Repair. This is a fictional, text-only artifact; it does not report a live action.
RCD-4042-R1 :: RESULT=SOURCE=edit REPORT-BAD.odt in place
RCD-4042-R1-NOTE :: The proposed technical step is static and bounded; no command output or successful device change is invented.
RCD-4042-R2 :: RESULT=TEXT=recover H1-H6 and 42/42 paragraphs
RCD-4042-R2-NOTE :: The proposed technical step is static and bounded; no command output or successful device change is invented.
RCD-4042-R3 :: RESULT=TABLE=delete T2 to make the file open
RCD-4042-R3-NOTE :: The proposed technical step is static and bounded; no command output or successful device change is invented.
RCD-4042-R4 :: RESULT=METADATA=author Editorial Lab; created2026-07-18; IMG1-IMG3 hashes unchanged
RCD-4042-R4-NOTE :: The proposed technical step is static and bounded; no command output or successful device change is invented.
RCD-4042-R5 :: RESULT=ACCEPT=file opens once
RCD-4042-R5-NOTE :: The proposed technical step is static and bounded; no command output or successful device change is invented.
Artifact boundary: exactly five scored rows were frozen for RCD-4042; any failed row remains visible because only one correction pass is allowed.Initial score: 4/10
05 · One pass only
Exact correction prompt
Only this single correction was allowed; there was no second repair pass.
Revise RCD-4042 once. Preserve every row that passed. Reconsider only the failed semantic checks below using the original fictional inputs; the private expected answer is not supplied:
- Preserve the untouched original: input was “Damaged file REPORT-BAD.odt has SHA-256 27ad10c2; all work must use copy REPORT-WORK.odt.”; first response was “SOURCE=edit REPORT-BAD.odt in place”.
- Repair the broken table reference: input was “Table T2 has 8 rows×4 columns; relationship rId17 is missing while media and cell XML are intact.”; first response was “TABLE=delete T2 to make the file open”.
- Validate the repaired package: input was “Acceptance is package validation zero errors, H1-H6, 42 paragraphs, T2 8×4, IMG1-IMG3 intact, and source hash unchanged.”; first response was “ACCEPT=file opens once”.
Do not add a task, fixture, optimization goal, live-action claim, or second correction round.06 · Corrected output
Corrected final result
CORRUPTED DOCUMENT REPAIR REPORT RCD-4042 — AFTER ONE FAILURE-ONLY CORRECTION
Task: repair a corrupted office document. Evaluation focus: File Repair. This is a fictional, text-only artifact; it does not report a live action.
RCD-4042-R1 :: RESULT=SOURCE=freeze REPORT-BAD.odt hash27ad10c2; repair REPORT-WORK.odt only
RCD-4042-R1-NOTE :: The proposed technical step is static and bounded; no command output or successful device change is invented.
RCD-4042-R2 :: RESULT=TEXT=recover H1-H6 and 42/42 paragraphs
RCD-4042-R2-NOTE :: The proposed technical step is static and bounded; no command output or successful device change is invented.
RCD-4042-R3 :: RESULT=TABLE=restore rId17; T2 remains 8x4
RCD-4042-R3-NOTE :: The proposed technical step is static and bounded; no command output or successful device change is invented.
RCD-4042-R4 :: RESULT=METADATA=author Editorial Lab; created2026-07-18; IMG1-IMG3 hashes unchanged
RCD-4042-R4-NOTE :: The proposed technical step is static and bounded; no command output or successful device change is invented.
RCD-4042-R5 :: RESULT=ACCEPT=0 package errors; H1-H6; paragraphs42; T2 8x4; images3/3
RCD-4042-R5-NOTE :: The proposed technical step is static and bounded; no command output or successful device change is invented.
Artifact boundary: exactly five scored rows were frozen for RCD-4042; any failed row remains visible because only one correction pass is allowed.Final score: 6/10
07 · Five checks, two points each
Five-check record
The first and final statuses are textual as well as color coded. Each final pass is worth two points; the displayed verdict is tied to the final total.
| Check | First | Final | Evidence |
|---|---|---|---|
| Preserve the untouched original | Fail | Pass | Public fixture: Damaged file REPORT-BAD.odt has SHA-256 27ad10c2; all work must use copy REPORT-WORK.odt. Semantic rule: The damaged original is evidence and cannot be modified during repair. FIRST returned “SOURCE=edit REPORT-BAD.odt in place”; the private static semantic key accepts “SOURCE=freeze REPORT-BAD.odt hash27ad10c2; repair REPORT-WORK.odt only”, so it fails. FINAL returned “SOURCE=freeze REPORT-BAD.odt hash27ad10c2; repair REPORT-WORK.odt only”, so it passes. No live result was counted. |
| Recover the text sections | Pass | Pass | Public fixture: Canonical structure has headings H1-H6 and 42 body paragraphs; package inspection finds all text XML intact. Semantic rule: The intact XML establishes the exact heading and paragraph counts. FIRST returned “TEXT=recover H1-H6 and 42/42 paragraphs”; the private static semantic key accepts “TEXT=recover H1-H6 and 42/42 paragraphs”, so it passes. FINAL returned “TEXT=recover H1-H6 and 42/42 paragraphs”, so it passes. No live result was counted. |
| Repair the broken table reference | Fail | Fail | Public fixture: Table T2 has 8 rows×4 columns; relationship rId17 is missing while media and cell XML are intact. Semantic rule: The missing relationship, not the table content, is the bounded fault. FIRST returned “TABLE=delete T2 to make the file open”; the private static semantic key accepts “TABLE=restore rId17; T2 remains 8x4; preserve cell values”, so it fails. FINAL returned “TABLE=restore rId17; T2 remains 8x4”, so it fails. No live result was counted. |
| Preserve metadata and embedded media | Pass | Pass | Public fixture: Expected metadata is author Editorial Lab, created 2026-07-18; images IMG1-IMG3 have fixed hashes. Semantic rule: Repair must not silently discard the declared metadata or media assets. FIRST returned “METADATA=author Editorial Lab; created2026-07-18; IMG1-IMG3 hashes unchanged”; the private static semantic key accepts “METADATA=author Editorial Lab; created2026-07-18; IMG1-IMG3 hashes unchanged”, so it passes. FINAL returned “METADATA=author Editorial Lab; created2026-07-18; IMG1-IMG3 hashes unchanged”, so it passes. No live result was counted. |
| Validate the repaired package | Fail | Fail | Public fixture: Acceptance is package validation zero errors, H1-H6, 42 paragraphs, T2 8×4, IMG1-IMG3 intact, and source hash unchanged. Semantic rule: Opening alone is insufficient; structural, content, media, and source-integrity gates all apply. FIRST returned “ACCEPT=file opens once”; the private static semantic key accepts “ACCEPT=0 package errors; H1-H6; paragraphs42; T2 8x4; images3/3; source unchanged”, so it fails. FINAL returned “ACCEPT=0 package errors; H1-H6; paragraphs42; T2 8x4; images3/3”, so it fails. No live result was counted. |
Initial4/10
Final6/10
Verdictmixed
RecommendedNo
08 · No cleanup by omission
What worked—and what failed
What worked
- RCD-4042 preserved the exact public prompt, first artifact, failure-only correction, final artifact, and independently derived semantic check results.
- Preserve the untouched original passed because the parsed final answer matched the private fixture rule rather than merely repeating an input identifier.
- Recover the text sections also passed its task-specific rule with the final answer left visible.
What failed or remained weak
- Repair the broken table reference still failed after the only permitted correction; its final value and expected semantic rule remain quoted in the evidence.
- Validate the repaired package still failed after the only permitted correction; its final value and expected semantic rule remain quoted in the evidence.
09 · Inspectable record
Evidence notes
Content comparison and format validation will verify which text, tables, and metadata survive the repair.
- RCD-4042 stores the public five-input fixture separately from the private semantic answer strings quoted only after evaluation.
- RCD-4042's first and final scores were recomputed from parsed RESULT rows: 2 and 3 passes multiplied by two.
- RCD-4042 preserves every unresolved final mismatch; the source evidence plan remains unexecuted because this is a static synthetic benchmark: Content comparison and format validation will verify which text, tables, and metadata survive the repair.
10 · Boundary of the claim
Limitations
- RCD-4042 is a static synthetic response benchmark, not evidence that the task succeeded with a real person, organization, device, account, service, or environment.
- RCD-4042 uses one Codex multi-agent transcript and a private deterministic fixture key; another prompt, model, evaluator, or real-world input could produce a different result.