Skip to content
PegBit Studio · Research

Research, documented in the open

PegBit Studio is the applied research arm of PegBit Technologies. Every programme below is written up the same way: the question, why it matters, the method, what we have actually shown, and what is still open. We publish results when we have them, and we say plainly when we do not.

PegBit WitnessOmosAwaazExamSurf
Programme 01 · PegBit WitnessStage: Pilot

Verifiable records of what AI agents do

Question
Can anyone independently verify what an autonomous AI agent did, without having to trust the organisation that ran it?
Why it matters
Agents now run commands, touch data and make decisions. When something goes wrong, the only record is often held by the same party being questioned, and logs can be edited. Monitoring tools usually also require sending private data to another cloud.
Approach
  • Every prompt, answer and tool call is sealed into a SHA-256 hash chain, so editing, dropping or reordering a record breaks the chain.
  • A split-plane design keeps private data in the customer's own environment. Only hashes are sealed.
  • Canonical hashing follows RFC 8785 and timestamps follow RFC 3161.
  • Detection of breakout and runaway-loop behaviour, and a one-click human circuit breaker that freezes an agent.
  • An offline verifier that checks an evidence bundle in a browser with no internet connection.
Evidence so far
  • The sealing engine, breakout and loop detection, circuit breaker and offline verifier are built and tested internally.
  • A staging pilot is running with GlobalPrayerAltar.
  • A public sample console, loaded with sample evidence, can be opened and tried.
Open work
  • Publish the offline verification method as an open, documented protocol.
  • Run on low-power edge hardware with no cloud connection.
  • Pilots with community and public-interest teams running AI agents.
See PegBit Witness →
Programme 02 · OmosStage: Research

Testing market ideas against evidence before acting

Question
Can written rules for several markets at once survive strict testing before they ever touch real money?
Why it matters
Many trading tools sell confidence, not evidence. Ideas get tuned to look good on past data and then fail live, and they usually watch one market at a time, missing what other markets signal.
Approach
  • One engine studies shares, government bonds, currencies, gold and crypto together.
  • Each idea is written down before it is tested.
  • A rule must hold in 800 of 1,000 resampled histories and in 2 of 3 separate periods, then sit one locked exam.
  • Rules that pass practise daily on paper accounts, under strict risk controls.
Evidence so far
  • All results so far are backtests or paper trading. No client money is managed.
Open work
  • A longer, documented paper-trading record.
  • Publish the testing method so others can examine it.
See Omos →
Programme 03 · AwaazStage: Pilot

Live commentary for matches nobody calls

Question
Can live match data become natural spoken commentary, in several languages, for games that have none?
Why it matters
Most women's and grassroots matches are played with no commentary. Fans following a scorecard get numbers, not a story, and blind and low-vision fans get nothing.
Approach
  • Read live scores and ball-by-ball data as they happen.
  • Turn each moment into a short exchange between a caller and an expert, and speak it with neural voices.
  • Store every line, so a replay sounds exactly like the live call.
Evidence so far
  • Open beta covering three sports (cricket, football, basketball) and five languages (English, Arabic, Urdu, Hindi, Filipino).
Open work
  • More languages and sports.
  • Structured feedback from blind and low-vision listeners.
See Awaaz →
Programme 04 · ExamSurfStage: Pilot

Personal AI tutoring for national exams

Question
Can a pixel-accurate exam simulation plus a personal AI tutor help Nigerian students prepare for JAMB and WAEC?
Why it matters
Students underperform not only from knowledge gaps, but from unfamiliarity with computer-based exam software and a lack of affordable, personalised tutoring.
Approach
  • Rebuild exam preparation around a pixel-accurate simulation of the real exam, so practice mirrors exam day.
  • Give every student an AI tutor and AI grading of written answers.
Evidence so far
  • Open beta since August 2026, free to students.
  • No outcome results are published yet.
Open work
  • Measure whether students improve, and publish what we find.
See ExamSurf →
What we have published

Technical write-ups, each marked with how well it is evidenced, are in our engineering notes. Product changes are in the changelog.

Research partners and funders

Work with PegBit Studio

We welcome grant funders, universities, public-interest teams and compute sponsors. Project proposals and technical documentation are available on request. Where a funder requires it, we can release specific research outputs as open source, while PegBit’s own platform products stay proprietary.

Start a conversation →Three ways to work with us →