Methodology
How AskVerum works
AskVerum is an epistemic trust platform: multiple independent models, structured synthesis, claim verification, and a transparent confidence score — designed for research, policy, and high-stakes decisions. This page explains what you get and how to read it. It is not a full technical specification.
1. Investigation flow
Each investigation runs as a single pipeline — typically about one to two minutes end-to-end. Stages run in a fixed order; some verification and bias work may complete shortly after the first results view.
Understand the question
We classify intent, apply domain context, and break complex questions into focused sub-questions where needed — before any model answers.
Independent model investigation
Multiple AI systems answer in parallel without seeing each other’s outputs. Disagreement is preserved; we do not average opinions into a single hidden answer.
Synthesis & structured verdicts
A separate synthesis stage turns raw answers into a readable brief and structured verdict signals used for scoring — not a chat transcript.
Trust, bias, and verification
A governed trust score combines triangulation, source verification, measurement quality, and bias checks. Verification against authoritative sources runs where the domain supports it.
Your investigation report
You receive an executive summary, findings, trust indicators, claim checks, and optional dig-deeper angles — shareable with your team.
2. AskVerum Trust Score™
The Trust Score (VTS-2.4) is a 0–100 verification-led headline. VC2 claim checks against sources are the primary input (33% weight), with model triangulation, measurement fidelity, and swarm coverage as supporting signals. Universal weights apply across domains; only uncertainty ceilings and safety caps vary by topic. Perspective coverage (VBE) is shown separately — soft bias bands do not multiply down the headline.
Four ideas to read the score by:
- Verification (primary). VC2 claim checks against public and authoritative sources drive the headline (33% weight). Contradictions are surfaced, not hidden.
- Triangulation. How much independent models align — and where they diverge. Supporting signal, not the whole story.
- Measurement quality. We distinguish synthesis-grade verdicts from fallback gap-fill so scores reflect instrument quality, not just model prose.
- Perspective (VBE, separate). Bias evaluation looks for framing, omission, and echo-chamber patterns on its own bar. Soft bias does not multiply down the headline; critical hard-fail caps for safety only.
Display labels (for example Strongly supported, Contested, Low confidence) follow a separate presentation policy so wording stays consistent even as the scoring engine is recalibrated.
Glossary (what the letters mean)
VTS — AskVerum Trust Score
The 0–100 headline governance composite on an investigation. Verification-led (not “models agreed”), with bias and uncertainty ceilings. It is not clinical or legal truth.
VVR — verification against sources
Claim-level checks that extract discrete claims and cross-check them against authoritative public sources where the domain pack supports it. You see per-claim status — not a stamp on the whole brief.
VBE — AskVerum Bias Evaluation
A separate perspective-coverage pass (framing, omission, echo chambers). Soft bands are shown beside the score; only critical safety hard-fails cap the headline.
VC2 — Verification Core version 2
The claim-checking core that feeds VTS. In VTS-2.4, VC2 is the primary weighted input (33%) to the headline.
Swarm coverage
A supporting signal for how broadly independent answering and verification agents covered the question set — not a popularity vote on truth.
Worked example (non-certifying)
Suppose a policy brief claims “State X cut maternal mortality 40% after programme Y.” Independent models may agree on the direction of the trend while disagreeing on the magnitude and on whether programme Y caused it. VVR may find a government registry that supports a mortality decline, a peer-reviewed paper that contests attribution, and no source for the exact “40%” figure. The Trust Score can still land in a mid band if verification coverage is partial — and the dossier should surface the contested claim explicitly. That score means the investigation met governance thresholds for transparency; it does not certify that the 40% claim is true.
3. Bias evaluation
Bias runs as a structured review across multiple perspective dimensions — political framing, source balance, omission, geographic lens, and related axes. Results appear on your investigation when complete. Critical perspective hard-fail can cap the trust headline for safety; ordinary bias patterns are shown separately and do not silently rewrite the synthesis narrative.
4. Source verification
Where domain packs allow, we cross-check selected claims against authoritative registries and public sources (legal, medical, research, government, and financial connectors vary by topic and region). You see per-claim status — supported, partial, contradicted, or unverifiable — not a binary “true/false” stamp on the whole brief.
5. Models & independence
AskVerum uses a curated, multi-vendor model roster. Different providers handle classification, answering, synthesis, bias, and verification tasks so no single vendor dominates the pipeline. Rosters are updated as models improve; tier and domain determine depth (for example live search on paid plans for time-sensitive main questions).
6. What we do not publish here
To protect product integrity we do not publish internal weight matrices, penalty coefficients, model routing tables, or registry wiring on the marketing site. Enterprise and research partners can request methodology briefs under NDA. Algorithm versions are tracked in product metadata on each investigation.
7. Limitations
Consensus ≠ truth
Multiple models can agree on an incorrect answer. A high trust score means strong epistemic support under our governance model — not certified fact. Verify consequential claims independently.
Time-sensitive topics
Parametric models have knowledge cutoffs. Paid tiers may add live investigation on the main question for news, markets, and policy — still read cited limits in the report.
Language & jurisdiction
English and major Indian languages are best supported. Compliance checks are India-first; global packs extend routing over time.
Not professional advice
AskVerum outputs are decision-support only. They are not a substitute for qualified legal, medical, financial, or other expert counsel.
Methodology questions: info@ask-verum.com · Last updated October 2026 · Trust engine VTS-2.4