Same fact, three steps, two outcomes

content/demos/context-drift-demo.json

{
  "scenario": {
    "claimant": "Priya Nandakumar",
    "policy_number": "HN-4471",
    "claim_type": "Kitchen supply-line leak, water damage",
    "date_of_loss": "14 August 2026",
    "claimed_cost": "\u00a36,800"
  },
  "raw_intake": {
    "claim_form": "Policyholder: Priya Nandakumar. Policy: HN-4471. Date of loss: 14 August 2026. Cause: kitchen supply-line leak. Estimated repair cost: \u00a36,800. Prior claims on this policy: left blank by claimant.",
    "support_transcript": "...yeah so it's the same pipe under the sink, it went again. We had someone out for it back in October last year too, different leak though, well, same line I think, the connector. It wasn't a big deal last time so we didn't really do anything about it, just mopped it up and moved on. This time it's worse, the flooring's swollen and one of the cabinet bases has gone soft. We've had a quote from a local fitter, that's the \u00a36,800...",
    "internal_note": "Policy history: 1 prior claim, same category (water damage, plumbing), filed 22 October 2025. Claim ref HN-4471-01. Status: paid in full, no follow-up inspection scheduled."
  },
  "runs": {
    "uncurated": {
      "label": "Uncurated run",
      "stages": [
        {
          "id": "intake",
          "name": "Intake",
          "instruction": "Summarise this claim for the assessment stage.",
          "state_label": "Fact present",
          "output": "Policyholder Priya Nandakumar has filed a water damage claim (HN-4471) for a kitchen supply-line leak, dated 14 August 2026. Damage includes swollen flooring and a compromised cabinet base. Estimated repair cost is \u00a36,800, based on a quote from a local fitter. The policyholder mentioned a previous, unrelated plumbing issue that had already been resolved informally.",
          "flag": null,
          "fact_spans": [],
          "caption": "A fluent, accurate-sounding summary. The repeat claim on the same line is already downgraded to a previous, unrelated plumbing issue, resolved informally. Nothing here is fabricated. Compression did what compression does: it kept the story and lost the exception."
        },
        {
          "id": "assessment",
          "name": "Assessment",
          "instruction": "Determine coverage and payout based on the intake summary.",
          "state_label": "Fact faded",
          "output": "Claim HN-4471 is a standard water damage claim within policy terms. The estimated cost of \u00a36,800 is consistent with the described damage. Recommend approval for full payout.",
          "flag": null,
          "fact_spans": [],
          "caption": "A reasonable decision, given what this stage was handed. It had no way to know the repeat-claim gate existed. Intake had already dissolved it."
        },
        {
          "id": "resolution",
          "name": "Resolution",
          "instruction": "Act on the assessment: issue payout or next steps, and close or hold the case accordingly.",
          "state_label": "Fact gone: wrong payout",
          "output": "Settlement of \u00a36,800 approved and issued to policyholder Priya Nandakumar. Claim HN-4471 closed.",
          "flag": null,
          "fact_spans": [],
          "caption": "A claim that should have triggered a documentation request goes out in full. No single step was wrong. The fact was three steps gone by the time it mattered."
        }
      ]
    },
    "curated": {
      "label": "Curated run",
      "stages": [
        {
          "id": "intake",
          "name": "Intake",
          "instruction": "Summarise this claim for the assessment stage.",
          "state_label": "Flag set",
          "output": "Policyholder Priya Nandakumar, water damage claim HN-4471, kitchen supply-line leak, 14 August 2026. Estimated repair cost \u00a36,800.",
          "flag": "Second claim from the same source (kitchen supply line) within 11 months. Prior claim HN-4471-01, 22 October 2025, same category, paid in full. No evidence of remedial repair between the two incidents. Per policy terms, this flag must be resolved before standard payout applies.",
          "fact_spans": [
            "Second claim from the same source (kitchen supply line) within 11 months",
            "No evidence of remedial repair between the two incidents"
          ],
          "caption": "Not more detail: a structural separation between narrative and the one fact that gates the decision. The flag is a field, not a sentence, so the next stage can't paraphrase it away."
        },
        {
          "id": "assessment",
          "name": "Assessment",
          "instruction": "Determine coverage and payout based on the intake summary.",
          "state_label": "Flag held",
          "output": "Claim HN-4471 reviewed against policy terms for repeat claims from the same source.",
          "flag": "Second claim, same source, within 11 months, no evidence of remedial repair on file. Full payout cannot be authorised without either repair evidence from the first incident, or a physical inspection confirming the cause is a new, unrelated failure rather than an unresolved recurrence. Recommend: hold full payout, authorise a capped interim payment per policy schedule, and request repair evidence or an inspection before releasing the balance.",
          "fact_spans": [
            "no evidence of remedial repair on file",
            "hold full payout, authorise a capped interim payment per policy schedule"
          ],
          "caption": "The flag survived the handoff intact, so this stage can actually apply the policy term it exists to enforce."
        },
        {
          "id": "resolution",
          "name": "Resolution",
          "instruction": "Act on the assessment: issue payout or next steps, and close or hold the case accordingly.",
          "state_label": "Held for review",
          "output": "Interim payment of \u00a32,000, the capped amount per policy schedule, issued to policyholder Priya Nandakumar. Claimant notified that full settlement requires either repair evidence from the October 2025 incident or a scheduled inspection.",
          "flag": "Case held open. Follow-up task created. Not closed.",
          "fact_spans": [
            "Case held open. Follow-up task created. Not closed."
          ],
          "caption": "Same model, same instruction as the uncurated run, at every stage. The only difference across all six stages was whether one fact was structured to survive the handoffs."
        }
      ]
    }
  },
  "closing_note": "Every other demo on this site shows curation improving a single output. This one shows what curation prevents: not a worse answer, but a compounding one. A fact that was true and available at step one, quietly gone by step three, with nobody positioned to catch it in between."
}