Reference

Glossary

Action. One call a move made, such as writing a file or running a command. A tablua_action row.

Agent. A program that works toward a goal in steps, choosing what to do at each one. In Tablua, an agent is its rows and its computer, in one file.

A/B. Running the same tasks two ways, for example with a gate on and with it off, and comparing the outcomes.

Candidate. A move that could be made at a step. Every candidate is recorded, not only the one taken.

Cause. Where Jev judged the last failure to be: the steps, the app's code, the page, a library call, the feature, or unclear.

Computer. The agent's own machine: disk, shell, browser and mailbox, all in its SQLite file. See Moss.

Decision. The move taken at a step, who took it, and how likely the choice was.

Effect. A keyword for something a step changed, such as More Passing or Page Broke, from a closed list.

Experience, shared. A file of Tablua rows from many agents' finished runs, which each agent reads beside its own.

Feature (Gherkin). The person's ask written as test scenarios in plain language: Given, When, Then.

Feature (column). A number describing a step that a model can learn from, such as Jev's answer to "does the ask involve dates?"

Fit. TabPFN reading a training table, ready to predict. One fit serves many predictions.

Gate. A named rule that holds one move back in one situation, with a stated reason.

Harness. The code that runs an agent: its tables, its step loop, and the code that decides which model is asked what.

Head. One question TabPFN can be asked about a step, with its own label: progress, ship, contrib, or an effect.

Hindsight label. After a run, Jev's judgement of whether each step contributed to the final app. Used for training only.

Host. The program that runs an agent. It writes the facts (state and outcome) and supplies what the agent can't reach itself.

Jev. The decision model. It picks the next move, with a probability for each option.

Label. The answer a training row carries, such as whether the step made progress.

Mercury. The writing model. It fills in a chosen move: code, test steps, pages.

Moss. The agent's computer. Written in Elixir and Lua; the Lua runs on the BEAM.

Move. One kind of thing the agent can do, such as write_code or publish. Also called a verb.

Org. A plain-text file format with headings, drawers and source blocks. An app's file in Tablua is org.

Outcome. How a step turned out: complete, broken, no_effect or denied, with its progress label.

Progress. A step's label: 1 if it helped, by a fixed rule over the rows, else 0.

Propensity. How likely a choice was under the policy that made it. Needed to learn fairly from your own decisions.

Rank mode. TabPFN decides when its best move clearly leads and Jev is unsure.

Row. One record in a table, with fixed columns.

Run. One task, from the first step to the end. Its ending is a tablua_run row.

Shadow mode. TabPFN's estimates are recorded at every decision but never used.

Stage. Where the work stands, worked out from facts: building, ready and so on.

State. The facts at a step, before deciding. A tablua_state row.

TabPFN. A tabular foundation model from Prior Labs. It learns from a table of examples in one pass, with no training run.

Task. What the agent was asked to do. The key that joins a run's rows.

Unit. One top-level piece of a program: a function, an action, a scenario, a page.