Research Protocol

Status: Proposed (May 27, 2026). Originated from a steward observation during the v2 / Path β execution of the Constitutional Ecology and Coordination Architecture Riff that the riff's reference-class engagement in §2.3 — particularly the Holy Roman Empire treatment at §2.3.3 — would have benefited from a structured research pass with source-tier discipline, verification, and a SOURCE_INDEX.md-style output rather than ad-hoc inline citation. The steward asked: how many sources, weighted how, verified how, with what kind of output, and at what scope? This protocol is the agent's answer.

---

The problem this protocol addresses

The project does research three ways today, and the three ways are not currently distinguished:

1. Programmatic source sweeps that produce structured outputs the project commits to maintain. The Government Overreach exchange's two research sweeps produced 53 source digests under sources/ indexed in SOURCE_INDEX.md. This is the project's heaviest research instrument and it has held up well.
2. Targeted research closing a specific exchange gap — e.g., the Round 4 research sweep that produced the seven second-sweep digests on Argentina, fiscal consolidation cases, AI catastrophic-risk literature, AI governance practice, cooperatives, sovereign wealth funds, and Swiss direct democracy. Same instrument as (1), scoped to specific evidence gaps named by an adversarial round.
3. Inline research grounding a riff, an exchange round, or a coherence check — e.g., the v2 §3.5 V-Dem / Freedom House / Bright Line Watch sourcing in the Constitutional Ecology riff, or the named-thinker engagements in §2.5 and reference-class treatments in §2.3. This is much lighter than (1) and (2) — citations live in the document rather than as digests under sources/ — but it carries real epistemic weight in the document.

The three modes share a research substrate but currently apply different (and mostly implicit) standards for source-tier weighting, verification, and external-link requirements. The steward's question is the right one: should they? And: what does the best, most thorough version of agent-driven research look like, so that the question of which mode to apply becomes a deliberate decision rather than an emergent property of context?

This protocol distinguishes the three modes by scope tier, specifies a source-tier weighting that applies to all three, names a verification discipline scaled to scope, and treats SOURCE_INDEX.md as the canonical output form for the heaviest scope tier — with intermediate output forms for the lighter tiers.

This protocol does not replace the existing protocols. It is a peer to the Adversarial Review Protocol, the Coherence Audit Protocol, the Historical Parallel Test Protocol, and the Comparative Alignment Protocol. The Historical Parallel Test Protocol, in particular, is what runs inside a programmatic source sweep when the question being researched is has a structurally similar reform been tried before?

---

Protocol

1. Three research scope tiers — choose deliberately

Before starting research, name which tier the work belongs to. This is a single-sentence judgment, not a process. The tier determines source-count discipline, verification discipline, output form, and time budget.

The Constitutional Ecology riff §2.3 is a T3 example with §2.5 and §11.6 also T3. The steward's intuition that §2.3.3 (Holy Roman Empire) would have benefited from more research is right — and the protocol's answer is if §2.3 had been done as a T1 programmatic sweep it would have produced ~5–10 digests per case (Haudenosaunee, Achaemenid, HRE, US Constitution, EU), with multiple viewpoint buckets per case, indexed in SOURCE_INDEX.md under a new sub-debate "Historical coordination architectures across pluralism". That work was not appropriate to do inside the v2 / Path β session under the steward's "Make it happen, Addi!" timebox, but is the right next move if any of the §2.3 cases becomes load-bearing for Path γ doctrine.

Tier escalation is allowed and expected. A T3 inline pass can be promoted to T2 or T1 when a follow-up artifact (doctrine, public-facing position, external review) raises the epistemic load on the original claims. The promotion is itself a deliberate decision recorded in the exchange or roadmap; it is not automatic.

2. The source-tier weighting (applies to all three scope tiers)

Sources are weighted by their epistemic standing for the kind of claim being made. The same source can be high-weight for one claim and low-weight for another. The weighting is not a ranking of source quality in the abstract; it is a discipline about what kind of source counts as evidence for what kind of claim.

2.1 Source-tier weighting table

2.2 Weighting in practice — the steward's specific question

The steward asked: are there other "peer-reviewed" content like medical journals that are more authoritative, but in other non-medical domains? The answer the protocol gives is yes and names them explicitly:

• Medicine. NEJM, JAMA, Lancet, BMJ, Annals of Internal Medicine, Cochrane Reviews, Nature Medicine.
• Law. Harvard Law Review, Yale Law Journal, Stanford Law Review, Columbia Law Review, Chicago Law Review, plus authoritative reporters (Federal Supplement, regional reporters) and the Restatements of Law.
• Economics and economic policy. NBER, IMF and World Bank working papers (different from their press output), BIS, ECB, Fed regional bank research, the AEA's flagship journals (AER, JEL, JEP), Journal of Political Economy, QJE, Econometrica.
• CS and AI. ACM and IEEE proceedings, the NeurIPS / ICML / ICLR pipelines (with caveat that single-paper findings often need replication), arXiv preprints from authors with established institutional affiliations, and the IETF RFC corpus for protocol-level claims.
• Political science. APSR, AJPS, JOP, World Politics, Comparative Political Studies, Public Opinion Quarterly.
• History. AHR, JAH, Past & Present, William and Mary Quarterly, and the major university-press historical monograph traditions (Cambridge, Oxford, Harvard, Princeton, Chicago, Yale, Stanford university presses).
• Sociology. ASR, AJS, Social Forces, plus Annual Review of Sociology.
• Engineering and standards. IEEE Transactions series, IETF RFCs (internet protocols), ISO and IEC standards, NIST publications (especially for crypto / AI risk / measurement standards), W3C recommendations.
• Statistics and demography. JASA, Biometrika, JRSS, Population Studies, Demography.
• Public health (distinct from medicine). American Journal of Public Health, BMJ Global Health, Lancet Public Health, WHO Bulletin.

The standing rule: when a claim has W1- or W2-equivalent literature available, the project cites that literature. When a claim falls outside any field's W1/W2 boundary (much of policy synthesis falls here), the project cites the strongest W3–W6 sources available, names the weighting honestly, and does not pretend the synthesis is settled where it is not.

2.3 Balance requirements

Every substantive claim that is currently contested in the relevant field must have at least one citation from each credible side of the contest. The SOURCE_INDEX.md viewpoint-bucket convention (pro-market / libertarian, social-democratic / progressive, heterodox / institutional / synthesis) is the project's standing answer for political-economy debates. Other domains may need different bucketing — the principle is balance the sides of an actually contested question, not find one source for each preset bucket regardless of what is actually contested.

For uncontested claims (data citations, basic historical facts, standard institutional definitions), one strong source suffices.

Historically-contested-but-converged subclause (added v1.1, May 27 PM, surfaced by the first §4.1 application on Constitutional Ecology riff §2.3.3 HRE). When a claim was historically contested but the contemporary literature has converged — i.e., the older view is now a historical position rather than a live scholarly minority — the balance requirement is satisfied by naming the convergence honestly rather than manufacturing a counterweight from a now-historical position. False balance against an old consensus is itself a research-integrity failure (it reads to a reviewer as performance rather than analysis). The honest move in such cases: cite two-or-more independent contemporary sources that converge, name the convergence as convergence in the surrounding text, name the older position as the historical foil it is, and flag candidate live contemporary critiques (challenges to the new consensus on its own terms, not warmed-over versions of the old position) as named future-T2 work when those critiques exist but have not yet been engaged. The HRE / Wilson-Whaley revisionist convergence is the worked example.

For uncontested claims, one strong source still suffices.

3. Source-count discipline — by tier

The steward asked: do we limit the number of research items? Yes, but in both directions — there is a floor below which a claim is undersourced and a ceiling above which adding sources is performative rather than informative.

External links: always required, with one exception. Every cited source must include a direct URL (open-access preferred). The only exception is when the source is a published book or paper without an open-access version and the citation provides full bibliographic information (author, title, publisher, year, page or chapter); in that case, the link can point to the publisher's catalog page or a reputable secondary index (OpenLibrary, Google Scholar landing page, JSTOR landing page) rather than the paywalled full text. "3 or more" external links is the wrong framing — the right framing is every source gets a link, period. The floors above apply to the number of sources, which equals the number of links.

4. Verification — sub-agent or in-protocol checks

The steward asked: can additional sub agents (or just within the same protocol) verify the results? Yes, and the discipline scales with tier.

4.1 T1 programmatic sweep — sub-agent verification pass required

Every T1 sweep ends with a verification sub-agent pass before the digests are merged. The sub-agent is given:

• The full set of new digests.
• The original sweep prompt and the sub-debate the digests were authored against.
• The authoring guard and this protocol.

The sub-agent's job is to identify, per digest:

1. Source-existence check. Does the URL resolve? Does the cited author have publicly visible affiliation with the cited work? Is the publication venue real? This is the failure mode the April 2026 coherence audit found — invented section titles and phantom citations. The sub-agent's job is to catch those at sweep time, not at audit time months later.
2. Tier-misclassification check. Is the source claimed at W1 actually W4 (e.g., a think-tank report cited as if it were peer-reviewed)?
3. Quote-fidelity check. Are quoted excerpts present in the cited text, with no insertion or deletion that changes meaning?
4. Viewpoint-imbalance flag. Does the sub-debate have at least one source per viewpoint bucket?
5. Authoring-guard pass. Does every claimed linkage to a core normative document quote the section's actual current title?

The sub-agent produces a verification report appended to the sweep. Digests flagged for any failure mode are not merged until the authoring agent addresses the flag.

4.2 T2 targeted gap-close — sub-agent verification pass or steward spot-check

Same verification pass as T1, but scoped to the gap-close set. Acceptable substitute: steward spot-check of at least one digest or citation cluster, with the steward applying the §4.1 checks personally.

4.3 T3 inline grounding — self-verification by authoring agent

The authoring agent runs the §4.1 checks on its own work before submitting the riff section or exchange round. Documented self-verification: the agent states "§4.1 checks passed for sources cited in §X.Y" at the end of the section. If the section is substantial (5+ sources cited), the steward should spot-check at least one source at the next riff-version boundary.

4.4 What verification does not substitute for

Verification is a citation-integrity check, not a truth check. A verification sub-agent confirms that the source exists, the quote is faithful, and the tier is correct. It does not confirm that the source is right. Source rightness is what the Adversarial Review Protocol, the Historical Parallel Test Protocol, and external human review are for. Conflating verification with adversarial review is itself a failure mode — the source is real and well-cited is a necessary but not sufficient condition for the source supports the conclusion the agent drew from it.

5. Link-and-quote discipline

Every cited source, at every tier, must satisfy:

1. Live URL. Direct link to the cited material. For books without open-access versions, link to the publisher's catalog page or to OpenLibrary / Google Scholar / JSTOR / WorldCat. Archive.org links are acceptable for paywalled or moved content if a direct link is unavailable.
2. Author, title, venue, date. All four, every time. For multi-author works, all authors in the first citation, et al. acceptable in subsequent citations within the same artifact.
3. Page or section locator. For book-length works or long reports, cite the specific page, chapter, or section the claim is drawn from. "As discussed in Wilson (2016)" is undersourced; "Wilson, Heart of Europe (2016), Chapter 12 'Justice'" is sourced.
4. Quote-fidelity for direct quotes. Direct quotes use quotation marks, are verbatim, and are accompanied by the page or paragraph locator. Quotes longer than ~75 words use blockquote formatting per the project's existing convention.
5. Tier label when contested. When a claim is contested in the literature and the project is citing one side, the citation includes the source's viewpoint or methodological tradition (per §2.3) so readers can locate the citation in the contest.

6. Output forms — by tier

6.1 T1 output: source digests + SOURCE_INDEX entries

Every T1 digest follows the existing digest convention (see source-higgs-crisis-ratchet-digest.md as the canonical example) — front matter with source_type, source_title, source_author, source_publication, source_date, source_url, provenance, viewpoint, sub_debates, copyright_notice; body with Status, Why this digest, Source identification, one or more Thematic clusters each containing Core claims, Representative excerpts, Research context (the corroborated / partially corroborated / debated / no-independent-source table), and a Project 2028 mapping subsection per the authoring guard.

SOURCE_INDEX.md is updated in the same commit as the digest is added. The protocol's view: SOURCE_INDEX.md is the canonical T1 output and should be treated as the master reference for the project's organized evidence base. The steward's intuition that the SOURCE_INDEX might be a good output example is exactly right for T1; less right for T2 (where digests are produced but not necessarily reorganizing the index); not right for T3 (where the output is the inline-cited riff section, not a new digest set).

6.2 T2 output: minimal digest set or Research grounding subsection

T2 produces either (a) a small set of digests under sources/ if the gap warrants their durability and the sources are reusable, or (b) a Research grounding subsection inside the exchange round that does the gap-closing. The Research grounding subsection format:

### Research grounding for [round / claim cluster]

Sources consulted (N):

1. [Author, Title, Venue, Date]([URL]) — [Tier label, viewpoint label]. Relevance: [1–3 sentences on what this source establishes for the claim].
2. [...]

Aggregate finding: [1–2 paragraphs on what the sources collectively support, where they disagree, and what remains unsourced.]

Verification: §4.1 checks passed by [authoring agent / verification sub-agent], [date].

The choice between (a) and (b) is a judgment call: if the sources will be cited by future exchanges, prefer (a) for reusability; if the sources are scoped to the immediate gap and unlikely to be cited again, (b) is sufficient.

6.3 T3 output: inline citations with link-and-quote discipline

T3 cites in the body of the riff, exchange round, or coherence check — same link-and-quote discipline as T1 and T2, but no separate digest file, no SOURCE_INDEX.md entry. The end of the cited section names the §4.1 self-verification as a one-line statement.

7. Where this protocol fits with the others

Common pairings:

• Research Protocol → Adversarial Review. A T1 or T2 sweep precedes the round that will be adversarially reviewed; the adversarial round can then challenge claims on the evidence, not just on structure.
• Research Protocol → Historical Parallel Test. When the question is has this been tried before?, the Historical Parallel Test runs inside a T2 gap-close, with the test's 2–3 cases each produced as T2 digests or Research grounding entries.
• Research Protocol → Coherence Audit. The Coherence Audit checks that the documents' citations resolve and are not phantom; this protocol's §4 verification is the prevention side of that audit.
• Research Protocol → Reviewer Packet. A T1 or T2 sweep is often the evidence base a Reviewer Packet (per REVIEWER_PACKET_TEMPLATE.md) draws on for the what the deliverable is responding to section; a reviewer can spot-check the sweep's verification pass as part of their review.

8. The "best, most thorough researcher" question

The steward asked: what does the BEST, most thorough researcher look like? The protocol's answer:

1. Names the scope tier deliberately before starting rather than letting it emerge from context.
2. Picks the right source tier per claim — peer-reviewed for empirical, primary documents for what-was-said, investigative journalism and reference compendia for contemporary factual, named-expert essays for position-articulation, practitioner accounts for tacit-knowledge, and never W7 or W8 for substantive claims.
3. Meets the floor and respects the ceiling for sources per claim cluster per tier.
4. Cites links and quotes faithfully every time, with author / title / venue / date / locator on every citation.
5. Runs a verification sub-agent pass at T1 and T2; runs self-verification at T3 with steward spot-check at version boundaries.
6. Balances contested claims with citations from each credible side.
7. Outputs to the right form — digests + SOURCE_INDEX at T1, digests-or-grounding-section at T2, inline-with-discipline at T3.
8. Knows what verification does not substitute for — verification is citation-integrity, not truth; truth is the job of adversarial review, historical parallel testing, and external human review.

A research pass that does all eight of those is what the project means by thorough. A research pass that does fewer of them is faster but produces evidence the next phase of work has to re-verify, which is a hidden cost the project has paid before (the April 2026 coherence audit found 17 + ~50 phantom citations in digests authored without §4.1-equivalent discipline).

9. When this protocol does not apply

• Steward personal note-taking (SCRATCHPAD.md) does not require protocol compliance; it is the steward's working space.
• Feedback capture (feedback/) follows the feedback/README.md conventions, which include their own consent and attribution discipline distinct from this protocol's source-tier weighting.
• Practitioner outreach (PRACTITIONER_OUTREACH_TEMPLATE.md) is not research; it is a different relational instrument.
• Riff opening-frame writing that captures the steward's thinking before any source engagement does not require sources; the sources arrive at the riff's body sections (T3) or in follow-up sweeps (T1 / T2).

---

Worked example — what §2.3.3 (HRE) would have been at each tier

To make the tier choices concrete, here is what the Constitutional Ecology riff §2.3.3 Holy Roman Empire treatment would have looked like under each scope tier. The actual treatment is T3; this is shown only to make the protocol's scope tiers concrete.

The protocol's view: the actual T3 work was the right tier choice for v2 of the riff under the "Make it happen, Addi!" timebox. A future promotion to T2 for the HRE case specifically would be triggered when the Exchange #24 Round 3 reviewer (under Option E Lane 1 or Lane 2) flags the reference-class engagement as needing more sourcing, or when Path γ doctrine begins drafting. That is the deliberate promotion path the protocol is designed to make legible.

---

Open questions for the protocol itself

1. Do T1 sweeps need formal Round-style structure inside sources/, or is the existing digest + SOURCE_INDEX.md form sufficient? Current view: existing form is sufficient; the structure lives in SOURCE_INDEX.md's sub-debate organization rather than in a Round protocol inside the sweep.
2. Should the verification sub-agent be a different model than the authoring agent by default? Current view: yes when feasible, per the Comparative Alignment Protocol's logic — different models catch different failures. When only one model is available, the sub-agent should at minimum operate in a fresh session with §4.1 as the entire context.
3. How are W6 practitioner accounts integrated when the practitioner is on the project's pseudonymity discipline? Current view: cite per the feedback/ conventions, with the consent state visible in the citation.
4. Does this protocol apply to the external-formation-docs/ repository? Current view: external-formation-docs is itself a research output; this protocol applies to the project's use of external-formation-docs as a source, not to the external-formation-docs repository's own internal organization.
5. Does T3 self-verification need a structured artifact, or is the one-line statement sufficient? Current view: one-line statement is sufficient at T3; if structured artifact is needed, the work is probably T2.

---

Changelog

• 2026-05-27 (PM, v1.1) — Added §2.3 historically-contested-but-converged subclause after first §4.1 application on Constitutional Ecology riff §2.3.3 HRE surfaced that the original §2.3 standing rule would have demanded a false-balance counterweight against a now-historical position. The subclause is the protocol's first revision from real application; it is also the first instance of the protocol's design assumption (that protocols revise in response to use) actually playing out.
• 2026-05-27 — Initial draft. Originated from a steward observation during the Constitutional Ecology riff v2 / Path β execution that §2.3.3 HRE would have benefited from structured research. The protocol is provisional until first applied to a research pass; first application is most likely on either (a) a T1 or T2 sweep for the Constitutional Ecology reference class if Round 3 surfaces the need, or (b) a future Exchange #21 follow-up (F3, F4) that requires fresh sourcing. Steward review pending.