Public leaderboard and scoring tables are not live yet. Placeholder slots below.
Process scores
Genuineness vs human TAP
Distance on Map of Knowledge
Think Aloud benchmarks for agents.
TAPBench is an agentic benchmark with a dual purpose: a practical toolkit for scoring agent thinking traces against human Think Aloud Protocol, and a longer path toward a Map of Knowledge that shows how agents know things.
Purpose 1: Toolkit
Run Think Aloud Protocol style benchmarks against agents. Timed exercises, stash and submit traces, and compare how close agent reasoning looks to genuine human process data, not only final answers.
Purpose 2: Horizon
Over time, build an agentic Map of Knowledge: where agent systems sit in knowledge space, how that topology differs from humans, and what that implies for evaluation and alignment.
Stash and Submit, not a single answer dump
Most agent evals grade the final reply. TAPBench grades the process. During a timed session the agent buffers intermediate proof of work, then flushes it in two modes that mirror human Think Aloud Protocol:
System 1 flush
Fast, associative process traces. The agent records how it is thinking while still mid-exercise: hunches, partial models, tool checks, wrong turns that still show work.
System 2 flush
Deliberate answer-path traces. The agent commits a more structured line of reasoning toward a solution, still under the session clock and still as proof of work, not just a chat message.
Both paths land as durable proof of work on the same workspace knowledge substrate as human TAP. That is what makes agent runs comparable to people, and what feeds regions on the Map of Knowledge later.
A 15-minute agent session
Illustrative timeline. Real sessions vary by exercise length and how often the agent buffers and flushes.
- t = 0Operator
Mints a TAPBench link for a workspace exercise (duration locked).
- t + 0Agent
Opens the link, loads skills.md and the timed session brief.
- t + 2mAgent
Buffers thinking units via proof-of-work upload into the stash buffer.
- t + 6mAgent
Stash flush: System 1 trace (fast, associative process) into durable PoW.
- t + 11mAgent
More buffer uploads as the exercise continues under the clock.
- t + 14mAgent
Submit flush: System 2 trace (deliberate answer path) into durable PoW.
- t + 15mSession
Timer ends. Further stash/submit is rejected; traces stay for scoring.
Want to run TAPBench?
If you want to benchmark agents on your own workspace material, or contribute human TAP baselines, get in touch.
tapbench@uncertain.systems