The AIVIS
Methodology
AIVIS scores are not model opinions. They are the output of a multi-stage retrieval and
validation pipeline designed so you can trace every finding back to its source.
This page is the source of record. Every claim made across aivis.biz
defers here. Model allocation, retrieval logic, semantic telemetry architecture, and BRAG
audit definitions are canonical as written below.
BRAG, The Audit Layer Under Everything
Every score, every citation event, every fix recommendation runs through BRAG before it reaches you. It is not a scoring heuristic, it is a verifiable chain of evidence with evidence IDs attached.
Results are grounded in scraped, evidence-linked source IDs, not model opinion. Deterministic semantic telemetry means measurements are exact: resolved against the engine's actual interpretation of your entity at query time, not sampled or approximated. Every output is independently validated before publication.
Four-Pillar Citation Operating System
Most platforms stop at measurement. AIVIS closes the full loop: measure → prepare → verify → act → prove. Each pillar is a distinct capability; together they form a system no lookalike platform currently replicates.
AIVIS doesn't sample. Every measurement is resolved against the engine's actual interpretation of your entity at query time, exact, not estimated. Results are reproducible and evidence-linked.
Content is structured for citation readiness, then the model's interpretation is verified against intent before publish. Preparation without verification is half the job.
When a gap is detected, AIVIS opens a repo PR with the fix automatically, schema markup, entity disambiguation, structured claim correction. No manual handoff from insight to implementation.
Immutable, timestamped record of every citation event: which engine, which source, which claim triggered the citation, and when. BRAG evidence IDs attach to every entry.
Immutable Citation History
The Cite Ledger is your audit trail. Every citation event is timestamped and attached to a BRAG evidence ID, traceable to the source, the claim, and the engine that triggered it. Public Observer share links surface a redacted view; full ledger access requires authentication.
Sample view. Actual ledger entries include full source URLs, evidence IDs, and prompt context. Available to authenticated accounts only.
Pipeline Model Assignment
Model routing is deterministic and documented here. All claims elsewhere on aivis.biz that reference model names defer to this table as the source of record.
| Role | Primary model | Fallback / pipeline |
|---|---|---|
|
Observer Citation monitoring |
Gemini 4 31B |
Gemini 4 31B
→
Gemini 4 26B MoE
→
Nemotron 3 Super 120B
→
MiniMax M2.5
→
Nemotron 3 Nano 30B
→
GPT-OSS 120B
|
|
Alignment Entity alignment checks |
GPT-5 Nano | Free-model fallback routing |
|
Signal + Score Fix Triple-check pipeline |
GPT-5 Mini |
GPT-5 Mini
generates
→
Claude Sonnet 4.6
critiques
→
Grok 4.1 Fast
validates
|
Observer uses free-model routing, cost is not a signal of quality here. The 6-model fallback chain ensures continuity; the primary model handles the majority of production load. Signal pipeline uses a fixed three-model sequence: one model making claims is not a methodology, three models checking each other is.
How a Score Is Produced
Trace any AIVIS score through these steps. Each step has a corresponding evidence artifact in Cite Ledger.
Target entity is queried across monitored AI engines. Responses are scraped; source citations are extracted and logged with evidence IDs. Deterministic telemetry resolves the engine's interpretation at query time.
Retrieved data is passed to Alignment (GPT-5 Nano). Entity signals are validated: is the model correctly associating the entity with its defined claims? Gaps are flagged with structured gap IDs.
Content is structured for citation readiness. The AI engine's interpretation of the prepared content is verified against intended meaning before any recommendation or publication proceeds.
GPT-5 Mini generates the signal output. Claude Sonnet 4.6 critiques it for accuracy, entity consistency, and citation readiness. Grok 4.1 Fast validates the final result. All three outputs are logged.
Where a structural issue is confirmed, an automated agent opens a repository pull request with the specific fix, schema, markup, entity disambiguation, or claim restructuring. No manual handoff.
Every citation event, whether a successful cite or a gap, is written to Cite Ledger as an immutable, timestamped record with attached BRAG evidence IDs. This is the proof layer.
What Each Page Claims, and Doesn't
To prevent misinterpretation, scope is defined explicitly here.
| Page | What it is | What it isn't |
|---|---|---|
| Compare | A citation positioning snapshot grounded in Cite Ledger history, showing where your entity appears relative to category peers across AI engines. | An independent benchmark publication. Rankings are retrieval-grounded, not editorially curated. |
| Observer (public share) | A view-only snapshot of a monitoring run. Redacted for security. | A full audit. Telemetry signals, evidence IDs, and implementation details are withheld from public links. |
| Methodology (this page) | The canonical source of record. All model names, pipeline steps, and BRAG definitions are authoritative as written here. | Marketing copy. Claims here supersede any conflicting language elsewhere on the site. |
Last updated: May 2026 · Canonical URL: aivis.biz/methodology · All other pages on aivis.biz defer to this document for model allocation, retrieval logic, telemetry architecture, and BRAG audit definitions. Implementation details in Observer public share links are intentionally redacted; authenticated accounts access full Cite Ledger history.