{
  "schemaVersion": 1,
  "caseId": "codex-compression-2026-09-26",
  "status": "historical-experiment-record",
  "date": "2026-09-26",
  "publishedOn": "2026-10-15",
  "canonicalUrl": "https://mintstorage.app/guides/codex-optimization-case-study/",
  "operator": {
    "name": "DZG Studio LLC / Mint team",
    "relationship": "Product maker; owner-operated engineering experiment on real private data, not an independent customer review"
  },
  "provenance": {
    "excerptUrl": "https://mintstorage.app/research/codex-compression-2026-09-26-source.txt",
    "excerptSha256": "d21840f0b270bc4a3f7591ab07c1ad70c7b771eff7ff8421ba0d3b0b6b77384c",
    "excerptEditedOn": "2026-10-16",
    "excerptEditNote": "Internal identifiers and implementation details were left out of the excerpt; every measurement is unchanged.",
    "evidenceType": "Dated experiment note from the Mint team’s records",
    "rawRunLogsAvailableInPublishedRecord": false,
    "rerunOnPublicationDate": false
  },
  "sample": {
    "source": "Independent copies of four large real Codex session files from the owner’s Mac",
    "compressedFiles": 4,
    "logicalThreads": 3,
    "reportedFileSizes": [
      "540 MB",
      "564 MB",
      "877 MB",
      "1.15 GB"
    ],
    "additionalReadDependency": "Two compressed files belong to one thread; its 44 MB first file was also supplied to Codex for reopening. It is not a fifth compressed file.",
    "randomSample": false,
    "representativeOfAllMacs": false,
    "liveOriginalsModified": false
  },
  "execution": {
    "operation": "A different, file-level compression method in a development build; ordinary reads return the original bytes",
    "buildType": "Development build from the historical experiment; not a test of the October 15 public 1.0.99 binary",
    "entryPoint": "Mint’s compression run directly in a development build, without the app interface; Optimize drawer interaction not tested",
    "compressed": 4,
    "operationSeconds": 12.6,
    "operationTimeIncludes": "Compression, file identity preservation and complete byte read-back"
  },
  "measurement": {
    "unit": "Decimal GB = 1,000,000,000 bytes; MB = 1,000,000 bytes",
    "reportedInputGB": 3.13,
    "inputSizePrecision": "Rounded total in the original lab note; exact before and after byte totals were not preserved in this published record",
    "allocatedReductionBytes": 1285251072,
    "recordedDfIncreaseKiB": 1255124,
    "dfIncreaseMatchesAllocatedReduction": false,
    "reportedAfterToBeforeBlockRatios": [
      0.445,
      0.606,
      0.624,
      0.627
    ],
    "approximateReductionPercent": 41,
    "percentageBasis": "1,285,251,072 / (3.13 × 1,000,000,000), rounded to a whole percent; approximate because the original input total is rounded",
    "averageReductionBytesPerCompressedFile": 321312768,
    "averageReductionBytesPerLogicalThread": 428417024,
    "averageThreadDenominator": 3,
    "averagingNote": "Total measured reduction divided by four compressed files or three logical threads. Not an average of individual percentage reductions.",
    "recordedDfIncreaseBytes": 1285246976,
    "allocatedReductionMinusDfIncreaseBytes": 4096,
    "dfComparisonNote": "The original note calls the values equal, but converting its recorded df increase to bytes reveals a 4,096-byte difference. They agree within 4 KiB, not exactly. The source excerpt keeps both original figures unchanged; raw run logs were not recovered to resolve the discrepancy."
  },
  "validation": {
    "sha256IdenticalFiles": 4,
    "codexRead": {
      "threads": 3,
      "turns": 1598,
      "items": 2381,
      "matchedUncompressedCopies": true,
      "originalContentKeptAfterResume": true
    },
    "undoTestedInThisExperiment": false,
    "guiInteractionTested": false,
    "realModelResponseTested": false,
    "twentyFourHourRegrowthMeasured": false
  },
  "interpretation": {
    "methodDistinctFromOctober15ScreenshotOptimization": true,
    "combineSavingsAcrossExperiments": false,
    "historicalResultNotCurrentVersionBenchmark": true,
    "lossless": "All four compressed copies returned exactly the original bytes; no screenshot thumbnail conversion or media deletion was used."
  },
  "limitations": [
    "One Mac and four deliberately selected large files, not a random population sample or minimum savings promise",
    "Historical development-build engine experiment, not newly rerun on the current public release",
    "Exact input total is rounded in the source; the approximate 41% must not be presented as an exact pooled allocation percentage",
    "No Undo, real model response, desktop UI or 24-hour regrowth check in this experiment",
    "The separately published October 15 case tests screenshot optimization and Undo on different files; its checks do not retroactively validate this experiment",
    "Private conversations and original run logs are not published; the dated source excerpt and aggregate measurements document the historical result"
  ]
}