{
  "schema": "org-writing@v1",
  "slug": "reconstruction-not-explanation",
  "kg": {
    "id": "org:writing:reconstruction-not-explanation",
    "type": "brick",
    "graph": "/kg.json"
  },
  "title": "Reconstruction, not explanation",
  "subtitle": "An account of a decision that cannot be rerun is a story, and the thing that usually blocks the rerun is a library version",
  "abstract": "Why accountability for an automated decision means reconstructing the state of the machine at the moment it decided, and why the opacity that actually defeats an audit is operational rather than neural.",
  "kind": "brick",
  "topics": [
    "Measurement",
    "Provenance at machine scale"
  ],
  "courseMemberships": [
    {
      "course": "org:courses:measurement",
      "topic": "Measurement",
      "wall": "org:walls:ethics",
      "position": 6,
      "total": 7
    },
    {
      "course": "org:courses:provenance",
      "topic": "Provenance at machine scale",
      "wall": "org:walls:engineering",
      "position": 4,
      "total": 5
    }
  ],
  "publishedAt": "2026-08-03T00:00:00.000Z",
  "version": 1,
  "guidelinesVersion": 15,
  "brief": {
    "problem": {
      "text": "Regulators and wounded people ask why a system decided what it decided, and the standard answer is an account produced after the fact by the same system, which is a story about the decision rather than the decision itself.",
      "claims": [
        "traceability obligations for high-risk systems are written into the European Union's AI Act"
      ]
    },
    "mechanism": {
      "text": "An explanation is generated after the event and can be checked against nothing, while a reconstruction is a rerun that can come out different, so only the second one is capable of failing and therefore only the second one is evidence.",
      "claims": [
        "more often operational than neural"
      ]
    },
    "move": {
      "text": "Record the whole envelope at the moment of inference, versions, artifact hashes, environment, seeds, and every time-dependent source, then declare in advance which grade of replay the system can honour.",
      "claims": []
    }
  },
  "sources": [
    {
      "repo": "mnstry-research",
      "path": "topics/methodology/assessment/58-forensic-provenance-determinism-replay.md"
    }
  ],
  "canonicalPath": "/writing/reconstruction-not-explanation/",
  "body": "The complaint about automated decisions has settled into a familiar shape. The network is a black box, its reasoning is inaccessible in principle, and the best anyone can do is ask the system to explain itself afterward. The complaint is not wrong about the weights, and record-keeping and traceability obligations for high-risk systems are written into the European Union's AI Act partly because of it. But go and try to audit an actual decision taken eighteen months ago, and the wall you hit is almost never the weights. It is that nobody recorded which release of a numerical library was installed, or what seed the sampler drew from, or which version of the time zone database was in force when a timestamp got converted.\n\nExplanation and reconstruction are different obligations wearing the same word. An explanation is produced after the event, by or about the same system under question, and there is nothing to check it against; it is a plausible account of feature importance that would have sounded equally plausible had the decision gone the other way. A reconstruction is a rerun. You rebuild the computational state at the moment of inference, feed the same input through it, and see what comes out, which means the exercise can come out different and can therefore fail. Only the thing capable of failing is evidence.\n\nWhat makes reconstruction hard is nearly always operational rather than neural, and the specifics are almost comically mundane. Floating-point addition is not associative, and a parallel reduction on a graphics processor makes no promise about the order in which it sums, so identical weights on a different card or a different driver can differ in the low bits and, at a threshold, in the answer. The time zone database changes several times a year, because jurisdictions keep moving their clocks and the database keeps recording it, so any computation conditioned on local time is silently pinned to a version nobody wrote down. Add the container image, the sampling parameters, the artifact hashes of whatever rubric or item bank was in force, and you have a list that is tedious rather than mysterious. Every item on it is recordable at the moment of the decision and unrecoverable a year later.\n\nWhich puts the honest system in an unusual position of having to grade itself in advance. Some decisions can be replayed exactly, given the captured environment. Some can be replayed only in distribution, matching within a stated tolerance, which is a real guarantee as long as the tolerance is published rather than discovered during the dispute. And some cannot be replayed at all, which is a fact about the system and not an accusation, provided it is declared before anyone asks rather than after. A product that labels a decision non-replayable has told a regulator something true and testable. An explanation has never once been able to do that.\n\nThe reason this belongs to measurement rather than to compliance is that it is the same commitment as the error bar, arriving in a different layer. An instrument that reports its own uncertainty is stating the limits of what its number can be asked to mean. A system that records its own envelope is stating the limits of what its decision can be asked to justify. Both are ways of refusing to be believed on charm. And a decision that can be run again in front of the person it was taken about is the only kind that ever really answered them.",
  "apparatus": {
    "note": "The human-facing essay is deliberately practical; this apparatus carries the full references, evidence-graded claims, article-local concepts, and research context behind it. Canonical concept definitions come from the concept registry.",
    "references": [
      {
        "id": "org:references:reconstruction-not-explanation:r01",
        "author": "European Union",
        "work": "Regulation on artificial intelligence (the AI Act), record-keeping and technical documentation provisions for high-risk systems",
        "year": 2024,
        "relevance": "The regulatory anchor for the demand. Traceability and record-keeping are obligations rather than good practice, which is what turns the reconstruction question from an engineering preference into a compliance one."
      },
      {
        "id": "org:references:reconstruction-not-explanation:r02",
        "author": "National Institute of Standards and Technology",
        "work": "AI Risk Management Framework",
        "year": 2023,
        "relevance": "The voluntary counterpart, which makes documentation, traceability, and accountability central to its guidance for organisations deploying these systems."
      },
      {
        "id": "org:references:reconstruction-not-explanation:r03",
        "author": "IANA",
        "work": "Time Zone Database, released several times a year as jurisdictions change their rules",
        "year": 2024,
        "relevance": "The most mundane and most instructive of the hidden non-determinism sources. A computation conditioned on local time cannot be replayed without the database version that was in force."
      },
      {
        "id": "org:references:reconstruction-not-explanation:r04",
        "author": "MNSTRY research record",
        "work": "Forensic provenance and deterministic replay, an auditability framework for scoring engines (internal synthesis, self-graded medium confidence)",
        "year": 2026,
        "relevance": "The framing this brick takes: auditability as the reconstruction of computational state rather than post-hoc rationalisation, the minimal provenance envelope, and the three declared grades of replay. The document's own field-level tables are internal and are not reproduced."
      }
    ],
    "claims": [
      {
        "id": "org:claims:reconstruction-not-explanation:c01",
        "claim": "Record-keeping and traceability obligations for high-risk systems are written into the European Union's AI Act, and the NIST AI Risk Management Framework makes documentation and traceability central to its guidance.",
        "basis": "The AI Act's high-risk provisions and the published NIST framework. Reported at the level of the obligation existing, with no reading offered of how any particular deployment falls under it.",
        "confidence": "verified",
        "sources": []
      },
      {
        "id": "org:claims:reconstruction-not-explanation:c02",
        "claim": "Floating-point addition is not associative, and parallel reductions on graphics hardware make no ordering guarantee, so identical weights can produce different low-order bits across hardware or driver versions.",
        "basis": "IEEE 754 arithmetic and the standard reproducibility literature for accelerated computation.",
        "confidence": "verified",
        "sources": []
      },
      {
        "id": "org:claims:reconstruction-not-explanation:c03",
        "claim": "The IANA time zone database is revised several times a year as jurisdictions change their rules, so any computation conditioned on local time depends on the database version in force.",
        "basis": "The database's own release history.",
        "confidence": "verified",
        "sources": []
      },
      {
        "id": "org:claims:reconstruction-not-explanation:c04",
        "claim": "The opacity that defeats an audit in practice is more often operational than neural, a missing library version, seed, container digest, or time zone database rather than the inscrutability of the weights.",
        "basis": "Our forensic-provenance research record, which argues it directly and grades itself medium confidence. A claim about where audits fail in practice rather than a measured distribution of audit failures, and graded accordingly.",
        "confidence": "directional",
        "sources": []
      },
      {
        "id": "org:claims:reconstruction-not-explanation:c05",
        "claim": "Replay guarantees can be declared in advance at three grades, exact given the captured environment, statistically equivalent within a published tolerance, and non-replayable with disclosure.",
        "basis": "The taxonomy proposed in the same research record and adopted here as our position. A design proposal, not a measured finding.",
        "confidence": "directional",
        "sources": []
      }
    ],
    "concepts": [
      {
        "id": "org:concepts:operational-opacity",
        "name": "Operational opacity",
        "definition": "The condition in which an automated decision cannot be audited because the environment around it went unrecorded, an absent seed, an unpinned library, a container image nobody hashed, a time zone database that has since been revised, rather than because a network's reasoning is inaccessible in principle. The remedy is a provenance envelope captured at the moment of inference and a replay grade declared in advance, since an account of a decision that cannot be rerun can be checked against nothing.",
        "provenance": "canonical"
      },
      {
        "name": "Declared replay grade",
        "definition": "A system's advance statement of which reconstruction it can honour, exact, statistically equivalent within a published tolerance, or none. A declaration made before a dispute is a testable commitment; the same statement made afterward is an excuse.",
        "provenance": "local"
      }
    ],
    "researchContext": "The source research document is a synthesis in our own vault, machine-drafted\nand self-graded at medium confidence, and this brick treats it as a framing\nrather than as evidence. Its regulatory and technical anchors were taken up\nonly where they stand on their own (the AI Act's record-keeping obligations,\nthe NIST framework, floating-point non-associativity, the time zone\ndatabase's release cadence), and its central thesis, that the opacity\nblocking audits is operational rather than neural, is graded directional\nbecause it is a claim about practice from a single medium-confidence source.\nThe document's detailed field tables for the provenance envelope are\ninternal engineering material and are deliberately not reproduced.\n\nThe boundary with the two pasts is exact and worth keeping. That brick owns\nthe record's beliefs over time, when a fact was true and when the system came\nto believe it. This one owns the machine's state at a single moment, and\nnever argues the bitemporal case. The parent essay's error bar is the\nargument this brick extends into the audit layer, which is why the closing\nparagraph names the parallel rather than restating the parent."
  },
  "contract": "https://mnstry.org/contracts/org/org-writing.v1.schema.json",
  "releaseHash": "103d0c2e2e4ab0ed40ba24771afd7a6ee910533650751620f8fc8c0007a7c418",
  "versions": [
    {
      "version": 1,
      "cutAt": "2026-08-03",
      "note": "Initial publication, provenance wave",
      "visibility": "published",
      "path": "/writing/reconstruction-not-explanation/",
      "contentHash": "sha256:e9e45cadd73cf6a6",
      "releaseHash": "103d0c2e2e4ab0ed40ba24771afd7a6ee910533650751620f8fc8c0007a7c418"
    }
  ]
}