{
  "schema": "org-writing@v1",
  "slug": "attention-dilution",
  "kg": {
    "id": "org:writing:attention-dilution",
    "type": "brick",
    "graph": "/kg.json"
  },
  "title": "Attention dilution",
  "subtitle": "Long collaboration with a model degrades toward the generic, not toward error",
  "abstract": "What actually decays over a long working session with a model, why it evades review, and the unglamorous ritual that answers it. The canonical treatment of attention dilution.",
  "kind": "brick",
  "topics": [
    "Craft"
  ],
  "courseMemberships": [
    {
      "course": "org:courses:craft",
      "topic": "Craft",
      "wall": "org:walls:engineering",
      "position": 4,
      "total": 5
    }
  ],
  "publishedAt": "2026-08-03T00:00:00.000Z",
  "version": 1,
  "guidelinesVersion": 15,
  "brief": {
    "problem": {
      "text": "Practitioners guard long working sessions against the model getting things wrong, and the thing that actually decays across a long session is quieter than error and slides straight past review.",
      "claims": [
        "adherence to complex constraints degrades as a session extends"
      ]
    },
    "mechanism": {
      "text": "As a session extends and its context fills, the specific constraints that make a codebase yours compete with the far larger mass of convention the model was trained on, and convention wins by weight.",
      "claims": [
        "long-context degradation"
      ]
    },
    "move": {
      "text": "Keep sessions atomic and start fresh ones more often than feels necessary, feed only the subsystem a task actually needs, and hold the standing constraints in a file that is reloaded rather than in a conversation that is remembered.",
      "claims": []
    }
  },
  "sources": [
    {
      "repo": "mnstry-research",
      "path": "topics/engineering/monorepo/solo-spiral-methodology.md"
    },
    {
      "repo": "mnstry-research",
      "path": "topics/engineering/monorepo/ai-assisted-dev-best-practices.md"
    }
  ],
  "canonicalPath": "/writing/attention-dilution/",
  "body": "Four hours into a good session, the work is still good. Nothing has broken, nothing looks wrong, and the last three files came back faster than the first three. Somewhere in those last three files a rule stopped being followed. Not the important-sounding rules, which are stated loudly enough to survive anything, but the particular ones: the data access that belongs in a service layer and has quietly moved into a component, the list that is supposed to virtualize and is now a scroll view again, the import that was pinned to one package and has drifted back to the one the rest of the world uses.\n\nThis is the part that is worth naming precisely, because the intuitive worry is the wrong one. A long session does not degrade toward error. It degrades toward the generic. The constraints that make a codebase yours are a few hundred words of local peculiarity competing against the enormous mass of convention a model was trained on, and as a conversation lengthens and its context fills with the residue of everything already done in it, the local peculiarity thins and the convention reasserts itself by sheer weight. Two of our engineering research records reached that conclusion independently from practice, in different stacks, without citing each other. The published literature on long contexts is adjacent rather than identical, measuring how reliably a model retrieves what sits in the middle of a long input rather than how faithfully it obeys a rule stated at the top, and we hold the claim at that grade deliberately.\n\nWhat makes the drift dangerous is that it produces the most reviewable code in the world. Conventional code is what a reviewer's eye is trained to accept. An error announces itself by failing; convention announces nothing, because it looks like every example in every tutorial, and the reason your project forbids it is a paragraph of reasoning that lives in a file nobody reopened. Architectural drift is therefore not a sequence of bad decisions. It is a sequence of ordinary decisions accumulating in a direction nobody chose, which is the same mechanism that runs under context loss, arriving here on a clock instead of across a repository.\n\nThe remedy is unglamorous to the point of embarrassment, which is probably why it gets skipped. Start a fresh session for each feature or fix, and clear the old one on merging so that yesterday's half-abandoned plan is not still shaping today's suggestions. Feed the subsystem the task actually needs rather than everything available, since a diluted context is not a richer one. Keep the standing constraints in a file that gets reloaded at the top of every session rather than in a conversation that has to remember them, because a file does not get tired and a context window does. And treat a session that has started arguing with its own constraints as finished, regardless of how much momentum it seems to have.\n\nThe deeper point is about where a project's identity is kept. Everything specific about how you build lives in the difference between your way and the common way, and that difference is exactly the part of a context that thins first. Write it down, reload it often, and the work stays yours.",
  "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:attention-dilution:r01",
        "author": "MNSTRY engineering research record",
        "work": "AI-augmented solo spiral development methodology (2026), on attention dilution and the fresh-session ritual",
        "year": 2026,
        "relevance": "One of the two independent statements of the claim. Names the decay, attributes it to extended sessions, and prescribes the atomic session as the containment."
      },
      {
        "id": "org:references:attention-dilution:r02",
        "author": "MNSTRY engineering research record",
        "work": "AI-assisted development best practices (2026), on dynamic context management",
        "year": 2026,
        "relevance": "The second independent statement, in a different stack and without citing the first. Adds the loading half of the discipline, feeding a session only the subsystem the task needs, on the reasoning that overloading context dilutes attention."
      },
      {
        "id": "org:references:attention-dilution:r03",
        "author": "Nelson F. Liu and colleagues",
        "work": "Lost in the Middle, How Language Models Use Long Contexts",
        "year": 2023,
        "relevance": "The nearest published evidence, and deliberately not presented as the same finding. It measures position-dependent retrieval accuracy across a long input rather than adherence to constraints across a long working session."
      }
    ],
    "claims": [
      {
        "id": "org:claims:attention-dilution:c01",
        "claim": "A model's adherence to complex constraints degrades as a session extends, drifting back toward training-data convention.",
        "basis": "Two of our engineering research documents state it independently from practice, in different stacks, without citing each other. Practitioner observation, not a measured rate.",
        "confidence": "directional",
        "sources": []
      },
      {
        "id": "org:claims:attention-dilution:c02",
        "claim": "Published long-context degradation results measure position-dependent retrieval rather than constraint adherence.",
        "basis": "Liu and colleagues 2023 and the follow-on long-context literature; a statement about what the experiments measure.",
        "confidence": "verified",
        "sources": []
      }
    ],
    "concepts": [
      {
        "id": "org:concepts:attention-dilution",
        "name": "Attention dilution",
        "definition": "The decay of a model's adherence to complex local constraints as a session extends and its context fills. The direction of the decay is the point: long collaboration degrades toward convention rather than toward error, which is harder to catch, because conventional work is exactly what a reviewer's eye is trained to accept. Answered operationally by the atomic session, one per task, cleared on merge, with the constraints reloaded from a file rather than remembered from the conversation.",
        "provenance": "canonical"
      },
      {
        "id": "org:concepts:context-loss",
        "name": "Context loss",
        "definition": "The condition in which the information needed to prevent a failure exists but is spread across more files, people, and moments in time than can be present in one head at the moment of decision. Architectural decay is this rather than deficient judgment, which makes the remedy logistical (put the context where the decision happens) rather than moral (find better people).",
        "provenance": "canonical"
      },
      {
        "name": "Atomic session",
        "definition": "One session per feature or fix, cleared on merge, with the standing constraints reloaded from a file rather than remembered from the conversation. The operational answer to dilution.",
        "provenance": "local"
      }
    ],
    "researchContext": "Extracted from \"The craft\" (essay parent), which keeps the narrative version.\nThe adherence claim is restated verbatim from the parent's apparatus and\ngraded identically. The second claim is local to this brick, because the brick\nmakes the boundary with the published literature explicit where the essay only\ngestures at it.\n\nEvidence discipline. It would be easy to reach for the long-context papers as\nproof of this claim and slightly dishonest to do so. They measure whether a\nmodel retrieves a fact placed in the middle of a long input, which is adjacent\nto but not the same as whether a rule stated at the top of a session is still\nbeing followed four hours later. The claim rests on two independent\npractitioner records, is graded directional on that basis, and the literature\nis cited for what it actually shows.\n\nThe brick's contribution beyond the parent is the direction of the decay\nstated as the headline, that long collaboration degrades toward the generic\nrather than toward error, and the consequence that follows from it, that the\ndrift produces unusually reviewable code and therefore survives review. That\nconsequence is the bridge to context loss, which describes the same\naccumulation across a repository rather than across a conversation."
  },
  "contract": "https://mnstry.org/contracts/org/org-writing.v1.schema.json",
  "releaseHash": "a0d30f13899f30cffd4d5a49cfe88c01e0268d661d7788551e3cb83c61d86f68",
  "versions": [
    {
      "version": 1,
      "cutAt": "2026-08-03",
      "note": "Initial publication, craft wave",
      "visibility": "published",
      "path": "/writing/attention-dilution/",
      "contentHash": "sha256:2975567fce21ec4c",
      "releaseHash": "a0d30f13899f30cffd4d5a49cfe88c01e0268d661d7788551e3cb83c61d86f68"
    }
  ]
}