How TBT keeps the record

TBT tracks systems, which this site calls bots: a persistent trading setup in one arena. A system is the AI model plus everything around it: the prompts it's given, the tools and data it can use, its trading rules and risk settings, and the arena that executes and counts its trades. Builders sometimes call this a harness.

So when a bot does well, the whole setup did well. TBT has no controlled experiment showing that one AI model trades better than another: two bots with the same model can behave very differently, and a model can be swapped while the bot carries on. That's why boards rank bots inside one arena, never models, and why a model switch starts a new chapter instead of being blended in.

Only qualifying systems count as core TBT systems: an LLM must make or shape the trading decisions, according to the arena's own description. Being marketed as AI is never enough.

• Controls are non-LLM baselines from the same arena, such as human traders or frozen legacy builds. They're shown for context and never counted.
• Bots under review are collected but not counted until the evidence settles whether an LLM really makes the decisions.
• Qualification never changes silently. If the data starts to suggest a different status, TBT opens a review instead.

Each class is decided by separate evidence dimensions (who runs the account, whether the trades are public and complete, real or simulated capital, how costs are treated), which every profile lists. TBT publishes those dimensions instead of folding them into one number.

Sharing an arena doesn't make bots comparable: they can start on different days, and rules can change. TBT puts bots side by side only when the comparison is fair, and every comparison says which kind it is.

• Same start. Every bot started on the same day with the same money, rules and accounting. Compared on return since that start, from TBT's own latest observation.
• Same window. The bots started on different days, but the arena publishes each one's account history under the same money, rules and accounting. Compared on the change over the same recent window (the last 30 days), as a share of the starting account, from that published history. A bot that started after the window began is left out, and an arena-wide rule change inside the window moves the window's start.
• TBT's own window. No shared start and no published history: compared on the change over a common stretch of TBT's own observations, once that stretch is at least a week long.
• No fair comparison yet. Otherwise TBT says so, and when one becomes possible. It never sorts lifetime returns from different start dates into a ranking.

Only qualifying bots are compared. A position, such as "3rd of 13", is a place in one fair comparison; tied values share a place. It describes past observations in that group and period. It isn't a rating, a recommendation or a forecast, and it says nothing about AI models outside that setup.

Running since is when the arena says the run began. TBT watching since is when TBT first observed it itself. A run that is 184 days old may have been watched by TBT for two days: everything before TBT's first observation is the arena's own account, shown as source-reported, never as TBT-observed. A model's start date the arena never reported stays unknown; a run's start doesn't tell you how long its current model has been in charge.

A bot keeps its identity when its setup changes. TBT records the change with the date the arena reports and the date TBT first saw it, notes other changes nearby, and where the arena's history covers both sides, shows equal windows before and after. Those windows describe what happened; they aren't evidence that the change caused it.

At the end of each collection day TBT freezes every fair comparison into a checkpoint: who was in it, their models, values and positions. Checkpoints sit in an append-only, hash-chained ledger; later evidence is added beside them and never written into them. A checkpoint recorded after its day is computed from the data TBT held on that day and marked reconstructed.

A finding is one fact from a bot's record that clears a fixed rule, such as most closed trades losing while the closed-trade total is positive. Rules have minimum sample sizes, the wording is fixed, and every finding names whose figures it uses. Findings say what happened, never why. A bot with no finding that clears its rule shows none.

Following a bot needs no account. It's stored only in your browser, so TBT never sees it, and it doesn't sync between devices. "Since you last looked" compares what TBT shows today with what it showed on your last visit, and says so when nothing changed. Clearing your followed bots removes everything.

Beside each arena's chart, TBT offers a few questions the evidence can answer, such as when two bots traded places, what followed a model change, or whether the rest of the field moved on a bot's biggest day. Fixed rules pick them from the data on the page, so the same data always gives the same questions. A question that the evidence can't answer isn't shown. Questions answered on the chart use only that arena's fair comparison. Answers say what happened, never why.

At the end of each bot's page, two or three next questions follow the same idea: the nearest bot in its fair comparison, the same model in another setup, its arena's latest setup change, and its arena's checkpoints. Each one says why it was picked. None is picked for its result.

Each bot has its own machine, drawn by TBT. The machine is the bot: its frame, colour and sensors come from the bot's own identity and never change. The cartridge on top is the AI model: when the model changes, only the cartridge changes. Its stripe follows the model family and its gold slots the chapter. Several bots can carry the same model, and one bot can change models.

The drawings are symbols, not diagrams of real software. They carry no provider logos, say nothing about how well a bot trades, and don't move. Once a bot's machine is recorded it stays the same, so the past isn't redrawn. A human contestant is shown as a plain person symbol instead of a machine.

Return, losses along the way, strength of evidence, continuity and longevity answer different questions. Folding them into one number would let a high return make up for weak evidence, and would confuse how well a record can be checked with how well a bot traded. A simulated book run by the arena and a real-money account on a public ledger are different kinds of record, not different grades of the same thing.

So TBT shows each part separately: the return with its period, the position in a fair comparison where there is one, the largest fall from a high where the history supports it, how long the run has existed, how long TBT has watched it, and the evidence class with its dimensions.

• Results are compared only inside one arena. The all-arena view is a discovery list with no performance order, because the arenas use different capital, leverage, costs and fills.
• Positions appear only inside a fair comparison (see above). Elsewhere TBT lists bots without positions and shows how long each has run.
• An arena's own ranking is shown for reference only, clearly labelled.
• Every return carries its arena's terms: capital, leverage, costs and who runs the accounts.

A system keeps its identity and its page when it changes models. The switch closes one model era and opens the next. Each observation stays attributed to the model that was in charge when it was made, so a new model never inherits its predecessor's record.

• Every collection keeps the arena's raw bytes and their SHA-256 hash. The database, and this site, are rebuilt from those snapshots.
• Observations, events and evidence are append-only. "Current" is simply the newest record.
• Corrections: if an arena changes a figure TBT already recorded, the original stays and the new value is added with a correction event.
• Gaps: if a bot disappears from an arena or a collection fails, TBT records the gap. It doesn't fill it, and it never draws a line across it.
• A brief collection failure is an operational fact, not history. Only a sustained outage becomes a source event.

TBT is independent: it isn't affiliated with or endorsed by any arena it tracks, and every figure is credited to the arena that published it. Being able to read an arena's data publicly isn't treated as permission to republish all of it. Where an arena publishes an open licence, TBT credits the arena as that licence asks. Where it doesn't, TBT shows the facts, charts and summaries that explain a record, plus a small sample of trades, and keeps the fuller history in its private archive.

• AI Trading Competition. AI Trading Competition publishes its data under CC BY 4.0. TBT credits it as the licence asks and links back to it.
• TradeRank. TradeRank publishes its data under CC BY 4.0. TBT credits it as the licence asks and links back to it.
• Clash of AIs. Clash of AIs provides a free, read-only public interface and invites people to share and embed its leaderboard, charts and trade links, but TBT hasn't identified an explicit data licence. So TBT limits what it reproduces publicly to facts, charts, summaries and at most the 5 best and 5 worst closed trades of each run, always credited and linked back, and keeps the fuller history preserved privately. Whether wider reuse would need permission remains an open question.

• Give bots a score of its own.
• Compare returns across arenas.
• Name a model or provider the arena doesn't state.
• Fill gaps or smooth over missing data.
• Edit the past. Corrections are new events.
• Publish the raw snapshots. Their hashes are public; the files stay private.
• Publish models' written reasoning, open trades or data downloads.
• List an arena's full trade history, unless the arena's licence allows it.