What the desk's hash-chained record holds, and which of Bitget's data surfaces this console answers from are answering right now. Live checks ran 06 Nov 2026 17:08 UTC; the slow sweeps say when they ran. Console · Research task · What we got wrong · JSON
In plain words: ARGUS runs a practice trading desk with no real money. This page shows how many decisions it has written down, whether any of them has been changed since (the "hash chain" — none has), and whether Bitget's data is reaching the console right now. The tables are for checking the work; you do not need them to use the console.
| Surface | Checked | Result | Detail | Time |
|---|---|---|---|---|
| Bitget market API | all tickers | answered | 884 USDT-margined contracts quoted; NVDA last 188.97 | 399 ms |
| Bitget market API | order book, 50 levels | answered | 50 bid and 50 ask levels for NVDA | 260 ms |
| Bitget market API | hourly candles | answered | 24 hourly NVDA candles, newest 06 Nov 17:00 UTC | 259 ms |
| bitget-mcp-server | equity quote (agent.bitget.com/mcp) | answered | US equity quote for NVDA: 188.5513 | 1,050 ms |
| bitget-signal | technical-analysis Skill | answered | technical-analysis Skill answered for NVDA, RSI 37.5 | 242 ms |
| Surface | Checked | 24 h | 7 d |
|---|---|---|---|
| Bitget market API | all tickers | 100.0% of 3 | 100.0% of 27 |
| Bitget market API | order book, 50 levels | 100.0% of 3 | 100.0% of 27 |
| Bitget market API | hourly candles | 100.0% of 3 | 100.0% of 27 |
| bitget-mcp-server | equity quote (agent.bitget.com/mcp) | 100.0% of 3 | 100.0% of 27 |
| bitget-signal | technical-analysis Skill | 100.0% of 3 | 100.0% of 27 |
Share of recorded probes that answered; the newest round was 2026-11-06 14:27 UTC, and the history's hash chain holds.
| Sweep | Result |
|---|---|
| Language model reader (Qwen) | not paused on this server instance |
| bitget-signal, the desk's calls | 6 of 19 calls (12 tools) returned data; 1 of 5 Skills returned data from at least one tool — swept 2026-09-26 |
| Understanding questions | a blind writer (never read this repository), 240 rows, 12 languages: this console, with no language model, read 125/240 (52.1%) correctly before 2026-09-25 and 196/240 (81.7%) after (204/240 (85.0%) with a saved book); its question classifier alone 219/240; with Qwen reading first 204/240 (85.0%) — measured 2026-09-25 |
| Numbers read from questions | of 47 blind-corpus questions that state a percentage, amount or multiple, 46 carry every one into the analysis; 1 loses one — each listed in data/figurecheck.json, none silently |
| bitget-signal, covered | of the desk's 19 calls, Bitget's server answered 6; ARGUS read 13 more straight from the sources those Skills name (11 the same source, 2 a named substitute), labelled as that source, never as the Skill — swept 2026-09-26 |
| bitget-signal, reliability | 2 of 19 tools answered all 3 attempts (crypto_derivatives, technical_analysis); 1 of 5 Skills have a reliable tool — swept 2026-09-26 |
| bitget-mcp-server, used by answers | 19 of 67 catalog entries are read by the console's answers; 15 of those answered in the sweep below |
| bitget-mcp-server, every entry | 37 of 67 catalog entries answered (55%) — swept 2026-09-28 |
| Corrected, not trusted | bitget-signal returns MACD's signal line and histogram swapped on 10 of 10 symbols checked; every technicals answer here recomputes from Bitget candles and corrects it |
| Research sources beyond Bitget | 12 of 12 (SEC EDGAR, FRED, news feeds and others) answered — probed 2026-09-24 |
| Our own calls, graded | 160 answers recorded since 2026-09-27, hash-chained; beta off by 0.09 on average over 52 calls four weeks on; direction Brier 0.241 against 0.245 for the plain base rate, 356 horizons — graded 2026-11-06 |
| Our weekend bands, graded | 30 weekend bands recorded before the reopen; 30 graded by Monday's open, 80% inside their 10-90 band (80% is calibrated) |
| Our thesis verdicts, graded | 198 verdicts on "keeps rising over the next week" recorded before the week; came true — supported 18 of 80; contradicted 10 of 72, 46 pending |
| Our event-move calls, graded | 15 stated moves for CPI and Fed hours recorded before the release; 10 graded, off by 0.31% on average, 80% bigger than the ordinary hour, 5 pending |
| Self-audit: does the lean rank the move? | information coefficient +0.042 at ~2h over 58 decision cycles (t +0.92, p 0.36, alphalens's per-period method) — no evidence the confidence ranks the next move; generated 2026-10-04 |
| Self-audit: were the refusals' leans right? | 741 of 1433 went the right way (51.7%, 48.1-55.4% over 135 cycles) against 51.6% for always "up" — not better than the naive call; median forgone edge -10.8bps after fees, as of 2026-11-06 |
| Self-audit: does noise in the evidence flip it? | 6 snapshots re-decided under seeded evidence perturbation (HELM's worst case): action held 100%, lean held 83% with every flip inside the unperturbed runs' own disagreement — a small sample, 2026-09-25 |
| Self-audit: same evidence, same answer? | 1 state replayed 5 times with the cache off: lean agreement 100%, action agreement 80% at worst — one state, so a spot check, not a rate (2026-09-13) |
| Self-audit: does the model already know the outcome? | Qwen recalled 32 of 42 dated facts (76% against 20% chance), with no upper boundary found through 2026H2 — so the decision-maker refuses to decide any frame before 2026-09-01 (agents/meta_pm.py) |
| Part measured alone: supervisor | 219/240 (91.2%) — questions routed to the right kind, blind 12-language set, classifier alone (2026-09-25) |
| Part measured alone: structured data | 50/50 (best rival 46/50) — FinanceBench metrics-generated questions answered from XBRL (2026-09-25) |
| Part measured alone: unstructured documents | 25/27 right, 3/3 unanswerable refused; 66/71 sentences supported, 0 fabricated citations — answerable questions right; sentences supported by their cited passage (2026-09-25) |
| Part measured alone: analytics | 17/17 caught, 0/4 false alarms (backtest-truth 5/17) — planted backtest defects caught, false alarms on clean code |
| Part measured alone: personalisation | recall 15/20 (mem0 18/20); false memories 0/12 (mem0 7/12) — last blind set: facts recalled; memories invented from talk about others (2026-09-30) |
| Part measured alone: reflection | 147/156 (94.2%) — content lines carrying a provenance label, unseen questions (2026-09-25) |
The desk decides four times a day during US market hours; the next cycle is due 06 Nov 17:30 UTC.