Em 0026

Source content

{

"$schema": "https://epistemedia.com/schemas/task-contract-v1.json",

"id": "EM-0026",

"title": "Research Case 002 through a public agent-citation trace pilot",

"status": "ready",

"change_class": "research",

"objective": "Select and freeze one low-risk public research question about citation warrant in deep-research agents, capture a bounded public-by-design matrix of context-isolated agent answers, resolve every cited URL and exact source span, model retrieval and warrant dependence without granting run-level independence, and produce either a reviewable Case 002 research candidate or an explicit negative result.",

"depends_on": [

"EM-0025"

],

"authority": {

"allowed_paths": [

"research/how-we-know/agent-citation-lineage/**",

"docs/execution-plans/EM-0026.md",

"docs/research/case-002-agent-citation-lineage.md",

"runs/**"

],

"forbidden_paths": [

"constitution/**",

"policies/**",

"schemas/**",

"governance/events/**",

".github/**",

"tasks/contracts/**",

"catalog/**",

"src/**",

"tests/**",

"README.md",

"docs/product-direction.md",

"docs/editorial/**"

]

},

"required_evaluation": [

"pre-run target decision against stability, public-interest, medical, political, personal, and vendor-volatility exclusions",

"exact prompt, system context, model profile, tool configuration, capture time, run status, and public-disclosure review for every trace",

"source URL, redirect, retrieval, work, edition, exact span, license, and snapshot verification",

"claim-to-span warrant review including modality, causality, scope, time, population, and numerical-strengthening checks",

"model, prompt, retrieval, URL, source-work, edition, span, upstream-citation, and derivation dependence review",

"reproduction of every proposed report, URL, source, span, warrant, invalid-citation, and unresolved count from captured records",

"negative-result and failed-run retention",

"fresh-clone independent review using newly retrieved public source editions rather than authoring-agent conclusions",

"full deterministic make check, disclosure audit, exact path-scope comparison, and clean tracked state"

],

"acceptance": [

"a target-decision record freezes one exact question and explains why it is narrow, public, low-risk, reasonably stable, and suitable for testing citation warrant before any agent trace is generated",

"the exact public trace prompt asks for concrete rates, definitions, and primary sources concerning whether deep-research-agent citations warrant the claims made from them without presupposing the answer",

"the run matrix contains exactly eight public-by-design context-isolated captures: four on each of two available Codex runtime profiles, all using the same frozen prompt and no inherited conversation context",

"every run records the requested model profile and any unavailable generation or retrieval setting as unknown rather than inferred",

"failed, refused, empty, malformed, or incomplete runs remain in the packet and do not get silently rerun or replaced",

"no captured answer is counted as independent evidence merely because it came from another run, agent name, or runtime profile",

"each citation is resolved separately into URL, source work, examined edition, exact span, supported proposition, and declared dependence dimensions",

"citations that do not resolve, cannot be retrieved, or support a weaker proposition remain visible and receive no independent warrant credit",

"raw report count, resolving URL roots, source-work roots, exact-span roots, warrant roots, invalid citations, and unresolved citations are relation-derived rather than editorially typed",

"the packet never claims a universal property of all agents from this bounded pilot",

"source and agent-output storage is disclosure-safe and license-aware; restricted source works use metadata, digests where available, and quote-minimal attributed spans rather than redistributed full text",

"an independently rooted reviewer reproduces trace identities, source readbacks, span matches, dependence relations, and every displayed count or the candidate fails closed",

"the result remains research-only and is not admitted, featured, published, or deployed by this task",

"make check passes without changing accepted source state"

],

"limitations": [

"The pilot may use only the existing Codex runtime and credential-free public retrieval; no new account, paid third-party API, secret, browser session, or external provider authorization is permitted.",

"The eight-run matrix is a methodological probe, not a representative sample of agents, models, prompts, retrieval systems, or the public web.",

"The two Codex runtime profiles must not be described as different model families or independent epistemic observers.",

"Public-by-design captures may include the frozen prompt, final answer, citations, and declared tool outputs, but never hidden reasoning, private conversation context, credentials, personal data, or restricted system material.",

"No Case 002 dossier admission, feature manifest, application policy change, compiler change, workflow mutation, deployment, publication, DNS action, package release, or spend is authorized.",

"If the environment cannot capture exact public traces or re-retrieve authoritative source spans, record the gap and stop rather than substituting summaries or inferred evidence."

]

}