LLM agents are increasingly asked to read markets and explain their calls, but their outputs are hard to audit: numbers appear in prose with no source, backtests quietly use data that did not exist yet, and track records can be edited after the fact.
My work treats verifiability as the core design constraint. I turn every number an agent reports into a structured claim that points at the exact value in a stored, point-in-time source snapshot, check each claim before anything is published, and record every call in an append-only, hash-chained ledger so the record can be scored forward only.
I also study how to build such systems cheaply and reliably with agents themselves: a builder/critic loop in which low-cost models write code and a stronger model plans and reviews, recording every attempt as labelled data for distillation.
Every reported number is a claim with a JSON pointer into a hashed source snapshot, verified before publication.
Agents only see data fetched before their decision time; calls are scored only after they are recorded. No look-ahead leakage.
Append-only, hash-chained call ledgers with database-level guards, so a track record cannot be edited after the outcome.
Builder/critic twin loops that produce labelled trajectories, corrections and failure modes for training smaller models.
Frontier models now post IQ-style scores above 99% of people yet stay far below humans on learning new tasks (ARC-AGI-3), long-horizon autonomy and continual learning. A growth-aware, multi-dimensional comparison with an interactive companion. Data snapshot 27 Sep 2026.
Method and early results from GlassBox: the baseline discrepancy audit, the verification gate, and the live forward-only record.
A. Thakare, S. Gondane, N. Prasad, S. Chigale.
S. Gondane, A. Thakare, C. Deshpande, O. Gupta.
An evidence-first multi-agent investment-research committee. Release s3-part1 adds a point-in-time snapshot store for every data fetch, structured claims with exact source pointers, a verification gate that re-checks every number (source match, point in time, staleness, snapshot integrity, units), a compliance filter, a gate-health dashboard and an append-only hash-chained ledger. Research preview only; not investment advice.
Building production software with a free model as builder and a strong model as planner and reviewer. Every run is traced; accepted runs, rejections and before/after correction pairs form a labelled dataset. Includes CPU benchmarks of small local coding models (Qwen2.5-Coder 1.5B at ~42 tokens/s on a 4-core laptop CPU).
An earlier multi-agent investment system built around a V(n) value function, with dashboards and analytics. The groundwork that GlassBox's verification-first design grew out of.
Infrastructure for an agent that assists scientific discovery workflows.
Dated releases, results and service, each linked to evidence where it exists. Newest first; updated Sep 27, 2026. Entries are never edited after the fact; corrections are new entries. Follow via feed.
A multi-dimensional, growth-aware comparison of frontier AI models and human intelligence, with an interactive companion of charts and projections.
Persistent researcher identifier linking publications to this profile.
Snapshot-integrity and unit checks added to the verification gate; an A7 compliance filter with 113 rule tests; a dashboard of the daily verification pass rate.
Every number in the first live committee decisions re-checked against its source snapshot; decisions provably unchanged by the verification layer.
Point-in-time snapshot store, structured claims, A6 verification gate, hash-chained call ledger, feature flags and kill switches.
Qwen2.5-Coder 0.5B and 1.5B (Q4_K_M, llama.cpp): ~68 and ~42 tokens/s generation on an i7-1165G7.
Repository opened with a README comparing it honestly with TradingAgents.
Reviewing and talks will appear here.