The operational record of building a governance-first AI platform.
Log #1
Logged by Rakesh Maheswaran, Founder, NiraNexus-OSIn brief
NiraNexus-OS runs a 3-round adversarial tribunal, not single-pass model averaging. Four frontier models debate (Opening, Cross-Examination, Rebuttal) before any verdict reaches the user. Every claim is labelled VERIFIED, DISPUTED, or UNVERIFIED against retrieved evidence. This log is the operational record of building that system as a solo founder across months of hardening in production.
AI outputs are brittle. A single model hallucinates and there's no check. Just a confident wrong answer delivered at speed. The industry's solution, running three models and averaging their outputs, doesn't fix this. It just averages the hallucination into something that sounds even more convincing.
NiraNexus-OS doesn't do that. It runs adversarial deliberation across four frontier models. Opening, Cross-Examination, Rebuttal. Three rounds. One verdict. No single-model hallucination survives.
I named NiraNexus after my daughter Nishka. The name carries what I want this to be: something personal, something that lasts. Infrastructure that doesn't trust a single model's output, ever.
Key Takeaways
Averaging three wrong answers gives you a confident wrong answer. If two models make the same error, the aggregate accepts it blindly. And now it's harder to spot because it sounds like consensus.
Our pipeline forces models to challenge each other directly. A model makes a claim. Another model cross-examines that claim. The original model rebuts or concedes. Only claims that survive all three rounds reach the final synthesis stage. A process that took months to harden across dozens of operating sessions. According to Irving et al. (2018), debate as a safety mechanism lets agents identify flaws in each other's reasoning that a single agent cannot self-detect. The theoretical foundation we extended into a production system with persistent deliberation records, source-model attribution, and evidence-grounded labeling.
The infrastructure story that followed is now documented across the Log. Log #2 covers the fire-and-forget database writes that silently lost 57 verdicts. Log #3 traces how a pipeline ordering bug turned a workaround into the failure itself. Log #4 shows how three yield points now guard every debate. Execution integrity through governed checkpoints.
Opening. Each model states its position independently, citing retrieved evidence. No model sees another model's response.
Cross-Examination. Models review peer claims and challenge factual accuracy, logical consistency, and source quality. According to the Consilium Protocol (2024), structuring this phase with engineered cognitive personas produces more rigorous cross-examination than open-ended critique. A claim with no retrievable URL is rejected outright across all four models.
Rebuttal. The original model defends or concedes. Substantive challenges that cannot be defended are withdrawn. Surviving claims advance.
One verdict emerges from the synthesis layer. Not averaged. Cross-examined.
The biggest surprise during development wasn't model hallucination. It was infrastructure. Fire-and-forget database writes lost entire deliberation verdicts. A Promise.race timeout killed synthesis mid-debate. The models worked. The plumbing didn't. Every subsequent session has been about hardening the operating layer around the intelligence, not the intelligence itself.
It labels it.
VERIFIED means retrieved evidence supports the claim. DISPUTED means conflicting evidence was found. UNVERIFIED means insufficient data exists. The system surfaces its own limitations. It doesn't fabricate confidence levels when the evidence doesn't support them.
This is the difference between a model that sounds right and a system that tells you what it actually knows.
This is the operational record. Not theory. Not a blog about "trends in AI."
Architecture decisions that worked and ones that didn't. Production failures that took 8 hours to diagnose and 5 minutes to fix. The pre-code gate that grew from a handful to double-digit mechanical checks. What the framework looks like after months of solo development. What shipped, what broke, what we learnt.
src/lib/debate-engine.ts, src/lib/prompts.ts, src/lib/types.tsIt doesn't trust a single model. Four models debate across three adversarial rounds. Claims are cross-examined by peers. Only surviving claims reach the verdict, and every claim is labelled against the evidence.
Claims cite retrieved URLs. Sources without retrievable URLs are rejected. Evidence is labelled VERIFIED, DISPUTED, or UNVERIFIED. The system surfaces what it knows and what it doesn't.
The first live NiraNexus component. A 3-round adversarial deliberation engine running across four frontier models. Users submit prompts and receive synthesised verdicts with full deliberation records.
Veritas, the professional deliberation engine for legal, accounting, and consulting domains, is in development. View the technical specification for Veritas.