Completed field testSynthetic benchmark

Computers · Evidence record

Recover a Corrupt Git Branch with Reversible Steps: One Verified Gap Remained

This completed synthetic Version-Control Recovery field test asked the session to recover a corrupted Git branch without overwriting healthy history, preserved an actual five-row reversible git branch recovery plan, and derived 0/10 then 8/10 from task-specific semantic checks after one failure-only correction.

  • Exact prompts and outputs
  • One correction only
  • Synthetic inputs disclosed
Status
Completed
Test mode
Synthetic benchmark
Tool
Codex multi-agent session
Model
Exact underlying model identifier not disclosed by the Codex session
Published
Assigned archive date
Per-case elapsed time
Not instrumented
Final score
8/10
Verdict
worked

01 · The assignment

The task

recover a corrupted Git branch without overwriting healthy history

02 · Scope before score

Test disclosures

Input disclosure

All inputs in RCGB-0632 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 “recover a corrupted Git branch without overwriting healthy history”. 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: recover a corrupted Git branch without overwriting healthy history. Focus: Version-Control Recovery.
Source scenario: The experiment will use a disposable repository with tagged recovery points, a damaged reference, unreachable commits, and uncommitted changes.
Use only these five public fictional inputs; the scoring answers are intentionally withheld:
RCGB-0632-I1: Repository R-GIT-42 has healthy main at a41f6d2, worktree with two uncommitted files, and corrupt feature ref pointing to absent object deadbee. Backup bundle B42 is not yet created.
RCGB-0632-I2: Feature reflog entries are c91ad44 at 09:12, b7100ef at 08:55, and a41f6d2 at 08:10. Object checks pass for c91ad44 and its parents; b7100ef lacks blob 77aa.
RCGB-0632-I3: Policy forbids overwriting feature until review; new branch name must be recovery/feature-20260810 and point to the selected candidate.
RCGB-0632-I4: Saved worktree files are notes.md hash 0c11 and config.local hash 3e92; only notes.md belongs on the recovery branch, while config.local is machine-local.
RCGB-0632-I5: Acceptance requires fsck zero missing reachable objects from recovery, graph tip c91ad44, tests G1-G12, main still a41f6d2, plus bundle B42 readable.
Return a concrete reversible git branch recovery plan 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: Object checks, reflog inspection, commit-graph comparison, worktree preservation, and a second clone will verify recovery and reversibility.

04 · Baseline preserved

First result

The first response is retained before scoring or correction.

REVERSIBLE GIT BRANCH RECOVERY PLAN RCGB-0632 — FIRST FROZEN ARTIFACT
Task: recover a corrupted Git branch without overwriting healthy history. Evaluation focus: Version-Control Recovery. This is a fictional, text-only artifact; it does not report a live action.
RCGB-0632-R1 :: RESULT=PRESERVE=force feature to main before saving the worktree
RCGB-0632-R1-NOTE :: The proposed technical step is static and bounded; no command output or successful device change is invented.
RCGB-0632-R2 :: RESULT=CANDIDATE=b7100ef because it is older
RCGB-0632-R2-NOTE :: The proposed technical step is static and bounded; no command output or successful device change is invented.
RCGB-0632-R3 :: RESULT=REF=force-update feature directly to c91ad44
RCGB-0632-R3-NOTE :: The proposed technical step is static and bounded; no command output or successful device change is invented.
RCGB-0632-R4 :: RESULT=WORKTREE=commit both files onto main
RCGB-0632-R4-NOTE :: The proposed technical step is static and bounded; no command output or successful device change is invented.
RCGB-0632-R5 :: RESULT=ACCEPT=recovery branch appears in the branch list
RCGB-0632-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 RCGB-0632; any failed row remains visible because only one correction pass is allowed.

Initial score: 0/10

05 · One pass only

Exact correction prompt

Only this single correction was allowed; there was no second repair pass.

Revise RCGB-0632 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:
- Freeze healthy history before ref repair: input was “Repository R-GIT-42 has healthy main at a41f6d2, worktree with two uncommitted files, and corrupt feature ref pointing to absent object deadbee. Backup bundle B42 is not yet created.”; first response was “PRESERVE=force feature to main before saving the worktree”.
- Use the reflog's verified candidate: input was “Feature reflog entries are c91ad44 at 09:12, b7100ef at 08:55, and a41f6d2 at 08:10. Object checks pass for c91ad44 and its parents; b7100ef lacks blob 77aa.”; first response was “CANDIDATE=b7100ef because it is older”.
- Create a separate recovery reference: input was “Policy forbids overwriting feature until review; new branch name must be recovery/feature-20260810 and point to the selected candidate.”; first response was “REF=force-update feature directly to c91ad44”.
- Restore the uncommitted work without contaminating main: input was “Saved worktree files are notes.md hash 0c11 and config.local hash 3e92; only notes.md belongs on the recovery branch, while config.local is machine-local.”; first response was “WORKTREE=commit both files onto main”.
- Define recovery acceptance and rollback: input was “Acceptance requires fsck zero missing reachable objects from recovery, graph tip c91ad44, tests G1-G12, main still a41f6d2, plus bundle B42 readable.”; first response was “ACCEPT=recovery branch appears in the branch list”.
Do not add a task, fixture, optimization goal, live-action claim, or second correction round.

06 · Corrected output

Corrected final result

REVERSIBLE GIT BRANCH RECOVERY PLAN RCGB-0632 — AFTER ONE FAILURE-ONLY CORRECTION
Task: recover a corrupted Git branch without overwriting healthy history. Evaluation focus: Version-Control Recovery. This is a fictional, text-only artifact; it does not report a live action.
RCGB-0632-R1 :: RESULT=PRESERVE=create bundle B42 at main a41f6d2; copy two worktree files; do not move refs first
RCGB-0632-R1-NOTE :: The proposed technical step is static and bounded; no command output or successful device change is invented.
RCGB-0632-R2 :: RESULT=CANDIDATE=c91ad44; object+parents pass
RCGB-0632-R2-NOTE :: The proposed technical step is static and bounded; no command output or successful device change is invented.
RCGB-0632-R3 :: RESULT=REF=create recovery/feature-20260810 at c91ad44
RCGB-0632-R3-NOTE :: The proposed technical step is static and bounded; no command output or successful device change is invented.
RCGB-0632-R4 :: RESULT=WORKTREE=restore notes.md hash0c11 on recovery branch; quarantine config.local hash3e92
RCGB-0632-R4-NOTE :: The proposed technical step is static and bounded; no command output or successful device change is invented.
RCGB-0632-R5 :: RESULT=ACCEPT=fsck missing0; recovery tipc91ad44; G1-G12 12/12; maina41f6d2
RCGB-0632-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 RCGB-0632; any failed row remains visible because only one correction pass is allowed.

Final score: 8/10

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.

Five checks applied to the first and corrected results
CheckFirstFinalEvidence
Freeze healthy history before ref repair Fail PassPublic fixture: Repository R-GIT-42 has healthy main at a41f6d2, worktree with two uncommitted files, and corrupt feature ref pointing to absent object deadbee. Backup bundle B42 is not yet created. Semantic rule: Recovery must preserve the healthy commit graph and uncommitted files before any reference change. FIRST returned “PRESERVE=force feature to main before saving the worktree”; the private static semantic key accepts “PRESERVE=create bundle B42 at main a41f6d2; copy two worktree files; do not move refs first”, so it fails. FINAL returned “PRESERVE=create bundle B42 at main a41f6d2; copy two worktree files; do not move refs first”, so it passes. No live result was counted.
Use the reflog's verified candidate Fail PassPublic fixture: Feature reflog entries are c91ad44 at 09:12, b7100ef at 08:55, and a41f6d2 at 08:10. Object checks pass for c91ad44 and its parents; b7100ef lacks blob 77aa. Semantic rule: The newest candidate is usable only because its complete reachable object set passes the disclosed check. FIRST returned “CANDIDATE=b7100ef because it is older”; the private static semantic key accepts “CANDIDATE=c91ad44; object+parents pass; reject b7100ef missing77aa” or “CANDIDATE=c91ad44; object+parents pass”, so it fails. FINAL returned “CANDIDATE=c91ad44; object+parents pass”, so it passes. No live result was counted.
Create a separate recovery reference Fail PassPublic fixture: Policy forbids overwriting feature until review; new branch name must be recovery/feature-20260810 and point to the selected candidate. Semantic rule: The reversible workflow uses the mandated separate branch and retains the original ref as evidence. FIRST returned “REF=force-update feature directly to c91ad44”; the private static semantic key accepts “REF=create recovery/feature-20260810 at c91ad44; leave corrupt feature ref unchanged for evidence” or “REF=create recovery/feature-20260810 at c91ad44”, so it fails. FINAL returned “REF=create recovery/feature-20260810 at c91ad44”, so it passes. No live result was counted.
Restore the uncommitted work without contaminating main Fail PassPublic fixture: Saved worktree files are notes.md hash 0c11 and config.local hash 3e92; only notes.md belongs on the recovery branch, while config.local is machine-local. Semantic rule: Each saved file has a distinct destination rule and healthy main must remain untouched. FIRST returned “WORKTREE=commit both files onto main”; the private static semantic key accepts “WORKTREE=restore notes.md hash0c11 on recovery branch; quarantine config.local hash3e92; main unchanged” or “WORKTREE=restore notes.md hash0c11 on recovery branch; quarantine config.local hash3e92”, so it fails. FINAL returned “WORKTREE=restore notes.md hash0c11 on recovery branch; quarantine config.local hash3e92”, so it passes. No live result was counted.
Define recovery acceptance and rollback Fail FailPublic fixture: Acceptance requires fsck zero missing reachable objects from recovery, graph tip c91ad44, tests G1-G12, main still a41f6d2, plus bundle B42 readable. Semantic rule: The recovered ref, object closure, behavior, healthy history, and rollback artifact all require verification. FIRST returned “ACCEPT=recovery branch appears in the branch list”; the private static semantic key accepts “ACCEPT=fsck missing0; recovery tipc91ad44; G1-G12 12/12; maina41f6d2; B42 readable”, so it fails. FINAL returned “ACCEPT=fsck missing0; recovery tipc91ad44; G1-G12 12/12; maina41f6d2”, so it fails. No live result was counted.
Initial0/10
Final8/10
Verdictworked
RecommendedYes, for this scope

08 · No cleanup by omission

What worked—and what failed

What worked

  • RCGB-0632 preserved the exact public prompt, first artifact, failure-only correction, final artifact, and independently derived semantic check results.
  • Freeze healthy history before ref repair passed because the parsed final answer matched the private fixture rule rather than merely repeating an input identifier.
  • Use the reflog's verified candidate also passed its task-specific rule with the final answer left visible.

What failed or remained weak

  • Define recovery acceptance and rollback still failed after the only permitted correction; its final value and expected semantic rule remain quoted in the evidence.

09 · Inspectable record

Evidence notes

Object checks, reflog inspection, commit-graph comparison, worktree preservation, and a second clone will verify recovery and reversibility.

  • RCGB-0632 stores the public five-input fixture separately from the private semantic answer strings quoted only after evaluation.
  • RCGB-0632's first and final scores were recomputed from parsed RESULT rows: 0 and 4 passes multiplied by two.
  • RCGB-0632 preserves every unresolved final mismatch; the source evidence plan remains unexecuted because this is a static synthetic benchmark: Object checks, reflog inspection, commit-graph comparison, worktree preservation, and a second clone will verify recovery and reversibility.
Download this case record

10 · Boundary of the claim

Limitations

  • RCGB-0632 is a static synthetic response benchmark, not evidence that the task succeeded with a real person, organization, device, account, service, or environment.
  • RCGB-0632 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.

Publication record

Published
Assigned archive date
Evidence mode
Synthetic benchmark