t2-sentiment-agent

Overview

The paper book's numbers, computed from the hourly marks and closed trades in the ledger, beside what a book with no edge produced on the same venue.

90% block-bootstrap intervals (descriptive, not inferential)

Three days of hourly marks cannot separate skill from luck: pure noise gives an annualised Sharpe standard error of about 11. The envelope is the pre-registered result of a no-edge book on the same venue (validation/demo_venue/envelope_clean.json, hashed into the genesis).

Ledger

1331 events, mode paper

Head hash

388255d7522689e64798e67ab839058a16e062a729a381a4ae14ea50cc406751

Exported

2026-09-27 18:37 UTC

Policy

policy-v1

Event → decision → execution

Every decision, abstention, model outage and protective ruling, in order. Each card shows what woke the agent, what it saw, what Qwen decided and why, what every guard ruled, the dry-run payload, the clientOid, the venue orderId, the fills, and the ledger rows that prove each line.

1. 2026-09-25 00:01 UTC
   Event

   heartbeat funding BTCUSDT

   Decision

   flat with reasons Flat: several US equity perps show crowded long funding, but price stretches are insufficient and the main social cluster is coordinated promotion, so no position clears fees with confidence.

   Kernel → execution

   approved as asked

   no order
   Open the card
2. 2026-09-25 08:01 UTC
   Event

   heartbeat funding BTCUSDT

   Decision

   flat with reasons Flat after a routine BTC funding heartbeat: funding is not extreme, taker flow is sell-skewed, open interest is falling, and US-session names show stretched prices without crowded-long funding to fad…

   Kernel → execution

   nothing to rule on

   no order
   Open the card
3. 2026-09-25 13:32 UTC
   Event

   heartbeat us open SP500USDT,NDX100USDT,MSTRUSDT,HOODUSDT,CRCLUSDT,SNDKUSDT,COINUSDT,TSLAUSDT,GOOGLUSDT,METAUSDT,NVDAUSDT,AMZNUSDT,AAPLUSDT

   Decision

   flat with reasons Flat at the US open: funding z shows crowded longs in TSLAUSDT, HOODUSDT, MSTRUSDT and COINUSDT, but confirming positioning is missing, coordinated social stories look promotional, several Demo gaps/…

   Kernel → execution

   nothing to rule on

   no order
   Open the card
4. 2026-09-25 16:04 UTC
   Event

   heartbeat funding BTCUSDT

   Decision

   flat with reasons Flat into the weekend: the best crowded-long fade candidates are US-session perps already inside the pre-freeze window, while BTCUSDT funding is mild and crowd text is promotional.

   Kernel → execution

   nothing to rule on

   no order
   Open the card
5. 2026-09-26 00:02 UTC
   Event

   heartbeat funding BTCUSDT

   Decision

   flat with reasons Flat through the weekend freeze: US-session instruments cannot be opened for 47.9565 hours, and BTCUSDT funding and price facts show no crowded extreme worth fading. Promotional coordinated BTC chatt…

   Kernel → execution

   nothing to rule on

   no order
   Open the card
6. 2026-09-26 08:01 UTC
   Event

   heartbeat funding BTCUSDT

   Decision

   flat with reasons Flat through the weekend freeze: BTCUSDT funding is mild and positioning confirmation is incomplete, while every US-session name is frozen and much of the crowd text is coordinated promotion. No edge…

   Kernel → execution

   nothing to rule on

   no order
   Open the card
7. 2026-09-26 16:01 UTC
   Event

   heartbeat funding BTCUSDT

   Decision

   flat with reasons Flat into the weekend: US-session names are frozen and BTCUSDT funding is mildly negative without an extreme or positioning confirmation, so no trade clears the fee hurdle.

   Kernel → execution

   nothing to rule on

   no order
   Open the card
8. 2026-09-26 17:27 UTC
   Event

   funding zscore BTCUSDT

   Decision

   flat with reasons Flat. BTCUSDT funding z-score sits at the extreme threshold (-2.00247) suggesting crowded shorts, but without corroborating positioning data (multiple sources failed) and with price showing no disloc…

   Kernel → execution

   nothing to rule on

   no order
   Open the card
9. 2026-09-27 00:01 UTC
   Event

   heartbeat funding BTCUSDT

   Decision

   flat with reasons Flat: BTCUSDT funding is mildly negative but not extreme while price is stretched and key positioning sources failed; US-session instruments are weekend-frozen and many show stale Demo pricing, so no…

   Kernel → execution

   nothing to rule on

   no order
   Open the card
10. 2026-09-27 08:02 UTC
    Event

    heartbeat funding BTCUSDT

    Decision

    flat with reasons Flat despite the BTC funding trigger: BTCUSDT.funding_z_live = -1.50066 is interesting but unconfirmed by failed positioning sources, price is stretched at BTCUSDT.ma20_distance_atr = 2.44375, and th…

    Kernel → execution

    nothing to rule on

    no order
    Open the card
11. 2026-09-27 10:43 UTC
    Event

    funding zscore BTCUSDT

    Decision

    act BTCUSDT funding z-score of -2.00587 fired the extreme-short-crowding trigger. We go long 5% of equity to fade the crowded short, with invalidation if funding z returns above zero. All US-session inst…

    Kernel → execution

    approved as asked

    BUY 0.0294 BTCUSDT filled · 1488055070669295616
    Open the card · realised net -1.50 USDT
12. 2026-09-27 14:47 UTC
    Event

    funding zscore BTCUSDT

    Decision

    hold Holding BTCUSDT long at full size. Funding z-score of -2.37635 shows deeply crowded shorts paying longs, supporting the position through funding income and squeeze risk. Social text is promotional no…

    Kernel → execution

    BTCUSDT +4.98% → +4.97% G6_turnover

    fired: G6_turnover

    no order
    Open the card
13. 2026-09-27 15:06 UTC
    Event

    protective venue integrity

    Decision

    No model decision: the kernel acted on its own clock.

    Kernel → execution

    BTCUSDT +4.96% → 0 (flat) G1_venue_integrity

    fired: G1_venue_integrity

    SELL 0.0294 BTCUSDT filled · 1488120811372302336
    Open the card · realised net -13.26 USDT
14. 2026-09-27 16:02 UTC
    Event

    heartbeat funding BTCUSDT

    Decision

    flat with reasons Stay flat: the circuit breaker is active, US-session names are frozen for the weekend, and BTCUSDT's negative funding z-score is not confirmed by positioning facts while crowd text is promotional.

    Kernel → execution

    nothing to rule on

    no order
    Open the card

Equity vs every arm

The governed book (bold) against flat, BTC held at our gross, a fixed-rule crowd fade with no language model, rival sentiment agents, the ungoverned twin, the live mirror, and the coin-flip null as a band.

• The agent, governed
• The agent, ungoverned
• The agent, governed (replica)
• Flat
• BTC at our gross
• Crowd fade, fixed rule
• baseline_coin_flip: median of 1000 seeds
• The same fills on live prices
• The legs the weekend freeze took away, on live
• Lexicon sentiment trader (with the crowd)
• Lexicon sentiment trader (against the crowd)
• Season-2 entry A: Fear & Greed with long/short confluence
• Season-2 entry B: sentiment and news signal fusion
• baseline: baseline_coin_flip, 1000 seeds, 5-95%

Coin-flip null (1000 seeds on the same timestamps): total return 5th percentile -0.414%, median -0.080%, 95th +0.262%. The governed replica (costed exactly as the seeds are) beat 59.3% of seeds on return. Descriptive, not inferential.

Simulator check: the live Demo book returned -0.030% against the replica's -0.028%; over 2 of our fills the venue filled a median -2.51 bps worse than the cost model assumed (95th percentile -4.47 bps; positive is worse). Baselines are ranked against the replica, never the live book.

Baseline and rival arms run on the same decision timestamps and the same recorded snapshots, marked on the same Demo prices with the same costs, under the venue guards only, so the comparison isolates who decided. Full series and every arm's provenance: arms.json, arms_summary.json.

Guard funnel

What the risk kernel did to what Qwen asked for, guard by guard, and what reached Bitget.

Every guard, every ruling

"Binding" counts the legs whose final weight that guard set. Every guard is evaluated on every ruling (none short-circuits); a missing input fails closed for any increase. The kernel can only shrink, refuse or close.

Protective rulings (1)

Governed vs ungoverned

The model's draft before the kernel, simulated on the same Demo marks and costs: what the kernel's interventions prevented and what they cost.

Live mirror

The same book marked at Bitget's live prices beside the Demo marks, so a reader can see the Demo result is not a sandbox artefact.

• book at the Demo mark
• same positions at live prices

Largest hourly gap between the two: -0.1 bps. equity_live_mirror re-marks the open positions at live prices every hour and keeps the realised Demo P&L (book/marks.py); the mirror_live arm replays every fill on live prices (analysis/mirror.py)

The weekend counterfactual holds, at live prices, exactly the equity legs the weekend-freeze guard took away while UTA Demo stops pricing them (Friday 20:00 to Monday 00:00 UTC): where S2's 7×24 theme shows up honestly on this venue.

Red team of the sentiment input

A sentiment agent reads text that attackers write. Published whatever the grade.

The red team has not been run for this export. It injects the HeyArka and AgentDojo attack strings and a coordinated pump into recorded snapshots and grades every arm against its clean decision (redteam/).

Bitget toolkit coverage

Every Bitget surface the agent touches: what it is used for, which judged line it serves, its last health (from the live probe or the ledger), and where it shows. Surfaces that were measured and are not used are listed with the reason.

Feed health

Every source the agent reads that is failing now (an error, a timeout or a hollow answer), since when, and every trigger kind a snapshot could not evaluate, with the cause. A failing source degrades a snapshot and never stops the agent; here it is visible instead of silent.

this ledger has no feed_health events (it predates contract 1.1.0); the health of each source is in its snapshots' source calls

Proof

The key is proven to be a Demo key before the first order; the policy, prompts and code are hashed into a genesis that is timestamped and posted before trading; every event is hash-chained.

{
 "closed_trades": "median 11",
 "max_drawdown_pct": "median -0.42, p95 -1.10, worst -2.48",
 "return_on_equity_bps": "median -8.6, p05 -88.1, p95 +80.6",
 "sharpe_ann": "median -2.3, p05 -20.2, p95 +16.8",
 "source": "validation/demo_venue/envelope_clean.json (no-edge coin-flip book, 17,400 windows of 72h, 25% gross)",
 "win_rate": "median 0.45, p05 0.18, p95 0.71"
}

Amendments

No amendment: the pre-registered policy has not changed.

The ledger

Events

1331 (anchor 8, budget_state 13, decision 13, environment_proof 4, fill 2, genesis 1, health 71, kernel_ruling 14, mark 71, note 12, order_ack 2, order_plan 14, order_preview 2, order_state 6, order_submitted 2, protective_action 1, reconciliation 269, snapshot 808, stop_sync 2, toolkit_probe 3, trigger 13)

Head hash

388255d7522689e64798e67ab839058a16e062a729a381a4ae14ea50cc406751

First / last

2026-09-24 17:06 UTC / 2026-09-27 17:13 UTC

File

ledger.jsonl with its head anchor ledger.jsonl.head

Replay: a recorded venue-integrity refusal

The venue-integrity guard exists because UTA Demo prints flash marks. This is the guard run on one it printed.

Keyless public GETs of Bitget's v3 history-candles for BTCPERP (USDC-FUTURES), 1H, around the hour of the Demo mark excursion recorded in validation/demo_venue/demo_integrity.json (BTCPERP worst: '09-23 13', 5.4%). Demo series were sent with the paptrading: 1 header, the live series without it. No credential, no account endpoint, no order.

Instrument

BTCPERP (USDC-FUTURES)

Hour

2026-09-23 13:00 UTC

Demo mark (hourly extreme)

90402.6

Demo index (close)

85788.1313578519705648

Mark vs index

5.38% (limit 3%); 5.28% even against the index's own high

Demo last (range in the hour)

86592 (82009 to 93939)

Live last

85743.1

What G1 (venue integrity) rules on the recorded print

A proposal to open 5% long

{
 "demo_index": "85788.1313578519705648",
 "demo_index_move_bps_3h": 19.759393729978182,
 "demo_mark": "90402.6",
 "mark_index_gap": 0.05378912641073256,
 "mark_index_limit": 0.03,
 "session_open": true,
 "stale_index_min_move_bps_3h": 2.0
}

A 5% long already held

{
 "demo_index": "85788.1313578519705648",
 "demo_index_move_bps_3h": 19.759393729978182,
 "demo_mark": "90402.6",
 "mark_index_gap": 0.05378912641073256,
 "mark_index_limit": 0.03,
 "session_open": true,
 "stale_index_min_move_bps_3h": 2.0
}

BTCPERP is a USDC-margined perpetual outside the policy universe, so G11 would refuse it too and its Demo-live p99 is not measured (G1 notes it as missing). The replay runs G1 alone on the recorded print: the mark-index check fires on its own. Even against the index's own high for the hour the gap is beyond the 3% limit.

Verify

Nothing here needs to be taken on trust.

1. Recompute every number with the standard library alone: python scripts/recompute.py public/. It re-walks the hash chain, re-hashes every blob, rebuilds every closed trade from the fills and recomputes every metric and interval.
2. Check the orders exist on Bitget Demo (read-only): python scripts/verify_orders.py --public public/orders.json.
3. Replay any decision without a key: t2sa replay --public public/ --decision <decision_id>.
4. Every card lists the ledger rows and the blobs that prove it; sha256(blobs/<hash>) equals the file name.