Collect content ideas without losing the evidence

1. Separate discovery records from accepted evidence

sources.json records a source ID, event ID, URL, source type and available text. The repost points to its original, the viral assertion explicitly lacks supporting measurements, and the unavailable demo has no text. All links are .invalid placeholders and all source statements are synthetic.

That structure makes the starting evidence inspectable. Discovery can identify a topic worth investigating without certifying its factual claims. A live version should retain original authorship, dates and access gaps with the real sources before building an editorial angle.

2. Require an editorial contribution and a next action

The bounded prompt asks for exactly two grounded ideas, held leads and gaps. An idea must identify a reader question, an original contribution, source identities, duplicates, missing evidence and a next research action. published must stay false. The task does not authorize writing an article or posting anything.

A useful queue helps decide what to research next. It is not a list of rewritten source summaries. The missing-evidence field makes the next work concrete, while the contribution field forces a proposed explanation or worked example beyond repeating the source material.

3. Inspect the two grounded ideas

The accepted output consolidates S1 and repost S2 into one receipt-pattern idea, retaining S1 as the original and S2 as a duplicate. Its reader question asks how to distinguish an attempted task, a validated result and destination readback. Its next action calls for a concrete workflow and checkable record format.

The separate S3 idea proposes a decision ledger and a worked recovery example. Both ideas retain missing evidence rather than claiming those future examples already exist. These are model-generated research suggestions; an editor still needs to check wording, source fidelity, demand and the proposed contribution before turning them into finished guidance.

4. Hold unsupported and unavailable material apart

The anonymous 99-percent cost-savings assertion stays in heldLeads. Its reason names missing invoices, workload definition, baseline and measurement method. The inaccessible demo appears in gaps. Neither becomes one of the two grounded ideas or a verified savings claim.

The output contains no publication action, and the source hash stayed unchanged. Those observations establish the intended separation on this fixture. They do not establish that every viral claim is detected or that a future article will be accurate simply because its lead came through the same queue.

How to reproduce and extend the exercise

Use the separate-folder replay instructions and compare each idea with its cited source IDs. Check that a repost has not inflated the number of distinct events, that unavailable material remains visible and that the next research step could actually resolve the missing evidence. The original prompt and accepted output are included.

To expand the workflow, start with a narrow audience and a few real, independently usable sources. Research the strongest ideas, build original examples and verify claims before publishing. This trial did not test live web discovery, audience demand, article originality at publication scale, or external publishing.

Inspect and reproduce

The source fixtures, original model outputs, prompts and execution receipts are below. Use a separate folder and your existing Hermes setup. Replace <trial-folder> in prompts with that folder. A local test run is not evidence of unattended production reliability.

{
  "published": false,
  "ideas": [
    {
      "eventId": "receipt-pattern",
      "readerQuestion": "How can a small team tell whether an agent task merely ran, produced a valid result, and actually left that result where it belongs?",
      "originalContribution": "Frame agent work as a three-part receipt: attempt, validation, and destination readback, with each checkpoint recorded separately.",
      "sourceIds": [
        "S1"
      ],
      "duplicateSourceIds": [
        "S2"
      ],
      "missingEvidence": [
        "A concrete workflow example showing the three checkpoints",
        "A defined record format or acceptance criteria for the readback"
      ],
      "nextResearchAction": "Find or create a small, reproducible example that records an attempt, validates its output, and reads back the intended destination."
    },
    {
      "eventId": "decision-recovery",
      "readerQuestion": "What should a project record preserve so a later contributor can recover why a decision stands or was replaced?",
      "originalContribution": "Use a decision ledger with explicit states for accepted choices, superseded choices, and proposals that were considered but never adopted.",
      "sourceIds": [
        "S3"
      ],
      "duplicateSourceIds": [],
      "missingEvidence": [
        "A worked example of recovering a decision after the project context changes",
        "Evidence about who updates the ledger and when"
      ],
      "nextResearchAction": "Develop a sample decision history with one accepted choice, one superseded choice, and one rejected proposal, then test whether a new contributor can reconstruct the rationale."
    }
  ],
  "heldLeads": [
    {
      "eventId": "viral-cost-claim",
      "sourceId": "S4",
      "reason": "The anonymous claim of 99 percent savings has no invoices, workload definition, baseline, or measurement method, so it cannot support a result or tutorial."
    }
  ],
  "gaps": [
    {
      "sourceId": "S5",
      "reason": "The referenced demo is unavailable and contains no usable evidence."
    }
  ]
}