Use cases
Where the dial, the ensemble, and Source Mode actually pay off.
Three concrete scenarios. Each one shows how the architecture — twelve verifiers, four tiers, NotebookLM-equivalent bounded answers, constitutional governance — maps to a specific work pattern that single-model tools can’t handle.
01 · Pharma
Drug repurposing intel that cites the page.
Pharma analysts evaluating repurposing candidates need answers that come from the FDA brief, the company 10-K, and the ClinicalTrials.gov registry — not the model’s training data. One hallucinated dosage schedule kills the project, and possibly the patient.
The setup
The analyst pins the relevant sources to the session: FDA approval letter, latest 10-K, ClinicalTrials.gov entries, supplier earnings transcripts. Source Mode goes ON. Tier is set to Full power — budget exists because the answer carries financial consequence.
What happens per query
- Pharma Compliance (Persona 10) activates as the lead reviewer — this persona only fires when the prompt touches medical / pharmaceutical content.
- The other 11 personas cross-evaluate the same answer — Test Oracle (Grok 4) checks the answer against the supplied sources’ explicit data points; Security Auditor (Opus 4.7) checks for accidentally exposed PII / clinical-trial subject IDs.
- Constitutional Tribune enforces principle P-07 «honour Source Mode» — if the answer extrapolates beyond the pinned sources, it’s vetoed.
- Every claim ships with
[SOURCE N]tags. The trail JSON shows which page of which document each claim came from.
«The agent won’t hallucinate dosage. It cites the FDA page or refuses. That’s the difference between ‘maybe useful’ and ‘trust me with the IRB submission.’»
02 · Regulated code
Code review against the actual control set.
Engineers shipping into a regulated repo (financial services, healthcare, defence) need every change reviewed against the explicit policy. «The agent thinks this is fine» isn’t evidence; «The agent cites SOC 2 CC6.6 and ISO 27001 A.9.4.1, both of which permit this pattern» is.
The setup
The team pins the SOC 2 control matrix, the internal security policy PDF, and the redaction rule set as session sources. Source Mode is on; Krentix can only reason from those documents. Tier starts at Medium for routine reviews, switches to Full when a control violation is suspected.
What that catches
- Persona 5 Security Auditor (Opus 4.7 at full tier) checks every code change against the pinned control set, citing the specific control number that permits or forbids the pattern.
- Persona 11 Trading Risk activates on financial keywords, catching position-sizing errors, accidentally hardcoded customer account numbers, or wrong-instrument trades.
- Persona 9 Constitutional Tribune enforces principle P-08 «privacy by default» — vetoes any change that adds new PII collection without explicit justification.
- The Supreme Court (cross-provider review of vetoes) ensures a hallucinated «this violates X» doesn’t become a denial-of-service that blocks legitimate work.
«Compliance review used to take three days, two engineers, and a senior reviewer. Krentix produces a draft review with cited controls in 90 seconds. Senior still signs it — but the input is shaped, not raw.»
03 · Research
NotebookLM’s job, with a verification ensemble on top.
Researchers pin papers, preprints, and internal write-ups to a session. Krentix synthesises across them, cites every claim, and refuses to extrapolate. Same job NotebookLM does — with twelve verifiers, four cost tiers (including $0/day forever for students), and a real provenance trail per claim.
The setup
Pin 10–20 PDFs (papers, internal notes, course materials). Source Mode on. Tier is Zero by default — works on free providers (Cerebras OSS / Moonshot Kimi / OpenRouter free tiers / local Ollama) so a research student can run unbounded sessions for nothing.
Why Krentix beats NotebookLM here
- Twelve verifiers vote on each answer — NotebookLM uses one model lineage; Krentix at zero tier already routes through three distinct OSS lineages (Qwen / GPT-OSS / GLM), each catching different classes of error.
- Refuse-on-extrapolation is enforced by veto — principle P-07 means the Tribune blocks any answer that drifts off-source. NotebookLM hopes the underlying model behaves; Krentix verifies.
- Public benchmark harness — HumanEval 89.0%, MBPP+ 81.5% measured publicly, harness in repo. NotebookLM ships no public score.
- Self-hosting — the bridge runs on the researcher’s machine, sources never leave their disk, no Google Drive lock-in.
«NotebookLM with sharper teeth, an honest score, and zero vendor lock.»