Methodology · research protocol
A governed, auditable construction framework.
The method separates source capture, substantive classification, deterministic verification, and scientific authority. Multi-agent review is one evidence-reconciliation layer within that larger protocol.
Control principles
Claims remain bounded by their evidence state.
A successful source-frame audit does not promote an unfinished classification layer, and agreement among agents does not authorize a scientific release.
- Separate roles
- Retrieval, screening, challenge, evidence, and audit tasks have different inputs, outputs, and failure rules.
- Bounded context
- Annual and task-level partitions keep work reviewable, resumable, and attributable to exact source artifacts.
- Fail-closed gates
- Missing evidence, changed dependencies, ambiguous inputs, or failed replay stop promotion instead of producing a plausible substitute.
- Human authority
- Agents may recommend; the principal investigator retains authority over construct-changing corrections and publication.
Figure 03 · Construction sequence
From official publications to a country–product exposure index
A count appears only where a retained artifact supports it. Unfinished stages are shown with dashed boundaries and no implied observation count.
- 01788,470
Candidate publications
Frozen source frame · 26/26 annual audits passed
Verified foundation - 02144,849
Routed publications
High-recall full-text routing layer · dedicated audit not started
Retained count - 03—
Restrictions observations
11 years fresh PASS; annual closure continues
In validation - 04—
Country–product linkage
Classification, temporal rules, and concordances to be specified
In development - 05—
GVN exposure index
Formula, network model, sensitivity tests, and release pending
Not released
Multi-agent assurance architecture
Auditable evidence reconciliation, not autonomous authority.
Independent agent passes retrieve, classify, and challenge candidate evidence. Deterministic software checks custody and consistency; unresolved cases proceed to human adjudication.
Agents propose and reconcile. Software verifies. Humans authorize.
- 01Retrieve
Bounded official-source inventories
- 02Classify
Provisional evidence decisions
- 03Challenge
Independent miss and boundary review
- 04Reconcile
Identity, date, provenance, and arithmetic checks
- 05Adjudicate
Human resolution of substantive ambiguity
- 06Authorize
Principal-investigator release authority
Table 01 · Research roles
Tasks, boundaries, and verification controls
Roles may run in parallel only where their inputs, outputs, and write boundaries are disjoint.
| Stage | Research task | Primary control |
|---|---|---|
| Collection | Run bounded FederalRegister.gov API queries, retain exact responses, create stable publication identities, and record query provenance. | Query inventory, response custody, and identity checks |
| Screening | Apply the substantive boundary to complete text and record explicit reasons for inclusion or exclusion. | Bounded criteria, evidence receipt, and reason codes |
| Challenge | Audit likely misses, compare boundary cases, reopen affected strata after detected errors, and preserve superseded decisions. | Independent review and preserved supersession history |
| Evidence | Retrieve the correct official document, locate the operative passage, and bind targets and mechanisms to exact evidence. | Official-document provenance and exact evidence anchors |
| Audit and orchestration | Bind inputs and outputs by hash, check receipts and dependencies, reject drift, and preserve nonaccepted pilots as diagnostics. | Deterministic schemas, hashes, arithmetic, and fail-closed gates |
| Human research authority | Define the construct, ratify material corrections, approve validation, and authorize the public release boundary. | Recorded adjudication and principal-investigator approval |
Methodological disclosure
What the framework does not claim.
- Not two human codersSeparate agent review passes are not conventional independent human inter-coder reliability.
- Not a statistical zero-error claim“Zero unresolved” is an internal closure gate, not proof that classification error is literally zero.
- Not the universe of U.S. policyThe Federal Register is official and auditable, but it does not contain every U.S. foreign-economic action.
- Not yet the GVN indexProduct concordance, network exposure, index construction, and external validation remain separate analytical work.
Next methodological contract
What must be fixed before the index can be released.
- Unit and product classification
Define the country–product–year unit, granularity, concordances, and time-varying classification rules.
- Policy-to-product mapping
Translate operative scope without inventing specificity absent from the source evidence.
- Network model
Select the network data source, propagation direction, weights, and treatment of missing links.
- Index construction
Publish the formula, normalization, aggregation, and sensitivity tests.
- External validation
Test construct validity, temporal behavior, edge cases, and reproducibility against a frozen release.