Open Methodology

How to inspect Veridi’s methodology

The complete Veridi methodology is published as a set of Markdown documents. These are the same files the AI follows when performing a fact-check; there is no hidden layer or proprietary process. What you see is what runs.

Source-of-truth discipline: what this methodology does not do

Veridi’s verdicts trace to real sources retrieved at the time of assessment, not to a pre-indexed corpus and not to the assessor model’s prior knowledge. Three commitments follow:

• No corpus replication. Veridi does not maintain a local copy of academic, governmental, journalistic, or institutional source material. Every assessment calls live search backends to retrieve current sources by URL and cites them in the output. The methodology improves over time; the source landscape is queried fresh each time.
• No model-knowledge substitution. Before final output, a retrieval-was-performed gate (Step 12g in the verification skill) inspects every EVIDENCE entry and tags it Retrieved (R) from a source by URL, or Substrate (S) meaning the model’s training-corpus knowledge. (S) entries are removed; if removal empties the EVIDENCE section, the verdict downgrades to INSUFFICIENT EVIDENCE or UNVERIFIABLE rather than committing to a directional answer. The substrate-side gate is soft, and a hard gate at the app layer separately discards responses that fail the per-tier search-floor minimums.
• Rights stay with rights-holders. Veridi does not redistribute, cache, or repackage the content of the sources it cites. Users follow citation links to read the source themselves at the source’s own URL, under whatever terms the rights-holder has set. Methodology versions are independent of the live source corpus, and Veridi does not claim a license to any of it.

Two practical consequences worth knowing for buyers and reviewers:

• No air-gapped deployment. A Veridi installation cannot run without live network access to the LLM API and search backends; the methodology cannot function on a fully isolated machine. The strongest disconnection the product supports is an off-network management channel for self-hosted enterprise customers (no inbound update connection from Veridi), but the deployment’s data plane remains connected to its LLM and search providers.
• The methodology is the product, not the corpus. Veridi customers pay for documented procedures, decision trees, source-hierarchy logic, gaming countermeasures, and calibrated confidence frameworks. The substrate (Claude, search backends) is provided by external vendors at the customer’s expense; the methodology is what Veridi delivers and updates.

This posture is deliberate. AI tools that rely on training-data recall or cached source material inherit hallucination risk, provable-accountability gaps, and unresolved rights-management exposure. Veridi’s structural choice to require live retrieval and to cite by URL forecloses those problems at the methodology level, at the cost of operating only when the network is available.

The methodology files

Core system

Evidence evaluation

Gaming countermeasures

Two distinct anti-gaming layers, one threat catalog and one procedural checklist. The file below contains both; each is enumerated in plain language in the “Veridi gaming countermeasures in detail” section further down.

Domain specialists

Supporting frameworks

Operational

Validation

Veridi gaming countermeasures in detail

Two anti-gaming layers, named here in plain language so the difference is visible without opening the file.

The 13 attack vectors name specific tactics that have been observed in the wild for manipulating fact-checks. The methodology defines a detection procedure for each one, including red flags and an analytical response. Vectors are the threat catalog: “here are the moves an adversary makes.”

The 15-item quick checklist is a procedure run before any Veridi verdict above 70% confidence. Each item is a yes/no question that a verdict has to clear. The checklist is the screening protocol: “here is what every high-confidence answer has to pass.” Several items map to specific vectors (e.g., the framing item maps to Vector 7); others are general hygiene that catches multiple vectors at once.

The 13 attack vectors

The 15-item quick checklist

Before issuing any Veridi verdict above 70% confidence:

1. Original evidence trail verified (not just derived sources; trace to origin).
2. Sources traced to independent origins (different reporting, ownership, access, timestamps).
3. Timestamps checked for coordination (suspiciously similar publication times?).
4. Language similarity checked (identical phrasing across “independent” sources?).
5. Claim falsifiability assessed (is the claim structured to resist verification?).
6. Preprint timing and credentials checked, if applicable.
7. Breaking event ceiling applied, if the claim is under 72 hours old.
8. Evidence standards symmetric (same standard applied to the claim and counter-claims?).
9. Tier integrity verified (does the effective tier match the publication tier?).
10. Framing assessed separately from facts (do true sub-claims create a false composite impression?).
11. Publication timing clustering checked (suspicious coordination window?).
12. Multi-clause claims decomposed (true anchors distinguished from false payloads?).
13. Data availability verified (is the relevant government data source still publishing?).
14. Institutional reliability checked (does the claim rely on a government source assessed at IRI Level 2 or higher?).
15. Promotional or advocacy framing checked (does the claim embed a product, service, organization, or methodology evaluation within an apparently factual assertion?).

What to look for

If you’re reviewing the methodology, here are the most important things to evaluate:

1. Decision tree consistency. Do the verdict decision trees produce the same result regardless of which path you take? Are there contradictions between the trees and the output format definitions?

2. Source hierarchy completeness. Are there source types that don’t fit cleanly into the four tiers? Are there edge cases where the confidence ceiling produces unreasonable results?

3. Gaming countermeasure coverage. Are there disinformation techniques not covered by the 13 Veridi attack vectors (enumerated in “Veridi gaming countermeasures in detail” above)? Can you construct a claim that should be detected but wouldn’t be? Is the 15-item quick checklist missing a procedural step it should include?

4. IRI assessment methodology. Are the degradation levels well-defined? Are the comparison anchors genuinely independent? Could the IRI itself be gamed?

5. Confidence calibration. Are the field reliability coefficients well-sourced? Does the interaction between tier ceilings and field coefficients produce sensible results across edge cases? Is the confidence-in-verdict vs. likelihood distinction clear and consistently applied?

6. Evidence directness and assumptions. Are the indirectness types (population, context, temporal, metric) well-defined? Are assumptions documented with meaningful consequence-if-wrong statements?

Pragma methodology files

Pragma is the policy evidence synthesis component: given a policy question, it evaluates the evidence base, assesses transferability, identifies value disputes, and produces a structured recommendation with calibrated confidence. Location: Pragma/

Core system

Evidence evaluation

Policy-specific frameworks

Domain specialists (13)

Validation

Cumulative validation: 53 PASS, 2 PARTIAL, 0 FAIL across 55 test claims.

Pragma gaming countermeasures in detail

Same two-layer pattern as Veridi, applied to policy analysis instead of fact-checking.

The 14 attack vectors name tactics that distort policy evidence synthesis: cherry-picking jurisdictions to make a policy look better than the distribution supports, dressing up correlation as causation, treating a pilot as if it scales, and so on. Vector 14 (Commitment Exploitation) was added 2026-05-01 and is unique to Pragma’s normative framework.

The 15-item quick checklist is the screening protocol run before any Pragma recommendation above Speculative confidence. Each item maps to one of the vectors.

The 14 attack vectors

The 15-item quick checklist

Before issuing any Pragma recommendation above Speculative confidence:

1. Jurisdiction distribution checked (is the cited case the best-performing outlier or representative?).
2. Causal mechanism verified (interventional or merely correlational?).
3. Implementation scale verified (pilot or at-scale?).
4. Counterfactual explicit (comparison to status quo, best alternative, or doing nothing?).
5. Transferability assessed systematically (all 7 dimensions scored, not just disclaimed?).
6. Trade-offs quantified (magnitude stated, not just mentioned?).
7. Inaction costs assessed (same evidence standard applied to doing nothing?).
8. Full option set considered (all evidence-supported options, not just politically palatable ones?).
9. Value choices explicit (any “the evidence shows we should” statement that is actually a value choice?).
10. Scale effects considered (would the intervention behave differently at the recommended scale?).
11. Evidence level accurately stated (causal language used only for interventional evidence?).
12. Discount rate explicit (time horizon justified, not assumed?).
13. Evidence independence verified (research groups, funders, datasets genuinely independent?).
14. Commitment exploitation checked (self-serving policy framed as benefiting disadvantaged populations or protecting the environment?).
15. Comparison anchors consulted (for government data sources, IRI checked, comparison anchors used?).

Praxis methodology files

Praxis is the individual action component: given a policy goal and a person’s profile, it identifies the highest-leverage actions available to that specific person. Location: Praxis/

Core system

Evidence and risk

Change pathways (9)

Each pathway file defines the mechanism, leverage conditions, evidence base, actions by engagement level, risk profile, sustainability characteristics, and cross-pathway synergies.

Validation

Cumulative validation: 34 PASS, 6 PARTIAL, 0 FAIL across 40 test scenarios.

Praxis gaming countermeasures in detail

Same two-layer pattern as Veridi and Pragma, applied to individual-action recommendations.

The 6 attack vectors name tactics that distort what a specific person is told to do: framing every option as futile, recommending feel-good substitutes for higher-impact actions, routing the user toward an organization whose funding does not match its public framing. The vector taxonomy is smaller than Veridi’s or Pragma’s because the action-recommendation attack surface is smaller (the methodology already rejects suggesting low-impact substitutes via the Impact Honesty rule built into Leverage Matching).

The 9-item quick checklist is the screening protocol run before any action recommendation at Standard tier or above. Several items map back to Veridi or Pragma vectors when the action depends on a fact-check or a policy claim.

The 6 attack vectors

The 9-item quick checklist

Before issuing any Praxis action recommendation at Standard tier or above:

1. Helplessness claims evidence-checked (does the “nothing works” framing hold up against Chenoweth data and documented successful movements?).
2. Impact level honestly classified (each action classified at its true impact level: Direct / Contributing / Infrastructure / Expressive, not inflated?).
3. Aggregation mechanism verified (does the action connect to a collective effort with a measurable pathway to outcomes, or is it isolated activity?).
4. Organization provenance checked (funding transparent, leadership identifiable, membership genuine?).
5. Urgency verified against calendar (actual deadline like a legislation date or election, or emotional urgency?).
6. Action plan community-centered (positions the user as a participant in collective effort, not a solo hero?).
7. Beneficiary match confirmed (for equity- or environment-framed actions, does the primary beneficiary match the claimed beneficiary population?).
8. Pragma gaming vectors cross-checked (for recommendations that depend on policy evidence, applicable Pragma vectors cleared?).
9. Promotional or advocacy framing checked (does the user’s goal embed a product, service, organization, or methodology evaluation within an apparently actionable request?).

The pipeline

Veridi (is this true?) feeds into Pragma (what should we do?) feeds into Praxis (what can I do?). Each system can also be used independently. Pipeline integration has been validated across 10 end-to-end scenarios, 30/30 stage executions PASS. Confidence decreases transparently at each stage as uncertainty accumulates.

Reporting issues

If you find an inconsistency, gap, or vulnerability in the methodology, we want to hear about it. The methodology improves through exactly this kind of external scrutiny.

Contact: veridi.org contact form