▮ AUTONOMOUS TRADING FIRM · 12 DESKS · ANCHORED ON NEAR · TEE-SIGNED LLM CALLS
Twelve specialist desks — regime, smart-money, Allora macro, on-chain, vol-forecasting — converge into a PM whose every call is anchored on NEAR, its LLM answers signed inside a NEAR AI Cloud TEE, and the whole firm can be hired per job on NEAR Agent Market. Orders fill on Bitget Demo. It fires ENTER only on a validated edge; when there isn't one the gearbox holds N. The discipline to hold is the product.
every figure below is real — the measurements, the failed edges, the abstentions
regime test accuracy · NEAR
78.9%
registry empty → holds N
▮ THE LOOP · IT RUNS ITSELF · 24/7
A whole firm on one closed cycle.No off switch.
A steward (or a hire on NEAR Agent Market) wakes the firm, telemetry fans to all twelve desks, the PM rules, the risk gate sizes, a rare ENTER fills on Bitget Demo and anchors on NEAR through an OutLayer custody wallet — then the edge-lab learns, memory stores it, and the wire loops back. Hover any node to see what it does.
the real firm, self-running · steward or Agent Market job → 12 desks (LLM calls in a NEAR AI Cloud TEE) → PM → risk → Bitget Demo → NEAR anchor → edge-lab learns → memory → loop · mostly ABSTAINs (discipline)
▮ THE FIRM · ONE CYCLE · EVERY READ
A four-strokedecision engine
Scan, argue, decide, seal — the same cycle on every read. Today there is no validated edge on the floor, so combustion stays dark. That no-spark is the product.
stroke 01 / 04
INTAKE
no spark
gearbox
registry empty → gear holds N
- 01
Intake
Ten desks pull live telemetry — regime, positioning, funding, whales, TVL, macro, volatility.
- 02
Compression
Bull argues bear. Twelve voices compress into one pressurized read.
- 03
Combustion
The PM is the spark plug. ENTER ignites only on a validated edge.
- 04
Exhaust
The decision — even the restraint — is sealed on-chain on NEAR.
▮ THE EDGE LAB · COST-AWARE OOS · WALK-FORWARD · FDR
The dyno doesn't lie
dyno run · NEARUSDT perp · out-of-sample · net of fees
0 of 2 passed
FUNDING FADE
32 trades · 37.5% win · -91.1 bps per trade
PRICE MOMENTUM
80 trades · 48.8% win · -36.7 bps per trade
No edge
registry stays empty
Gate: out-of-sample ROI above zero and above buy-and-hold, more than 20 bps per trade (the round-trip cost), at least 20 trades, FDR control and a consistent walk-forward. Nothing was tuned until it passed — so the firm abstains on NEAR.
2
hypotheses tested
0
passed the gate
2
scrapped
the scrap shelf · cause of death
- FUNDING FADE−28.5% OOS · below buy-and-hold
- PRICE MOMENTUM−35.0% OOS · buy-and-hold +62.6%
- OI-CHANGE SIGNALSnot testable — no OI history
gates: cost-aware OOS · anchored walk-forward · outlier-robust · Benjamini-Hochberg FDR
▮ HONESTY LAB · FINDING #10
We crash-testour own execution
the demo order book · NEAR · 10 round trips
+2.21%
median buy fill vs live price
Bitget Demo quotes a wide bid-ask, so the buy lands above the market while the sell lands on it.
impact report · realistic execution
−2.23 %
per round trip · median
A demo account would make every trade look worse than it is. So the paper ledger at live prices is the official PnL and the demo fill is only proof that the order path works.
Disclosed
a firm that publishes its crashes is a firm whose passes mean something
● QC INSPECTION SHEET · 9FINDINGS · WE PUBLISH WHAT DOESN'T WORK
The shop report
Every claim hits the bench as a part. Some we keep; most we scrap. Here is the log — tinted by verdict, stamped, and linked in full.
bench (shown)
5 kept · 4 scrapped
- REAL
Accuracy ≠ profit
NEAR regime classifier: 78.9% test accuracy vs 76.1% persistence — and still no edge that clears costs.
- REJECTED
NEAR funding-fade fails the gate
−28.5% OOS over 32 trades (−91 bps each); buy-and-hold −22.7%.
- REJECTED
NEAR price-momentum fails the gate
−35.0% OOS over 80 trades while buy-and-hold made +62.6%.
- ABSTAIN
The registry stays empty
No hypothesis cleared the bar, so the firm abstains. We don't lower the gate.
- ABSTAIN
OI edges are untestable on NEAR
The venue gives a live OI snapshot only, and funding covers 24% of the bars. We say so.
- REAL
Direction from tokens, not `kind`
Reading the sign from GeckoTerminal's `kind` flips it in USDC/wNEAR pools. Tokens matched 10 of 10 on-chain checks.
- REAL
An honest security veto
No token-security API exists for NEAR, so the check reads the chain and prints `honeypot=unverified` instead of faking a verdict.
- REAL
A rejected hire refunds in full
The contract returns the deposit and emits `job_rejected`; the app reads the event, not just 'success'.
- REAL
Demo fills are not our PnL
Bitget Demo buys fill +2.21% above live (median). The paper ledger at live prices is the PnL.
every row traced to a run — a firm that logs its scraps is one whose keeps mean something
▮ VERIFIABILITY · NEAR · PUBLIC READ
The black box is public
Flight recorders survive the crash — that's the point. Every cycle's decision, even the restraint, lands in a public table and anchors to NEAR.
first anchor · proof-stone
BLOCK
276,890,553
HWo1NgDKkJ…UfBAhgep
core.anchor · deploy smoke test · sealed on NEAR
lap log · live decision log
no data yet
live feed unreachable here — the anchor plaque is the documented record
▮ THE MACHINERY · ALWAYS-ON · RAILWAY + NEAR
Runs while you sleep
THE ORG BRAIN
12-desk firm + PM debate loop · LLM calls on NEAR AI Cloud (TEE) · Groq as fallback · runs every cycle in the cloud
THE EDGE LAB
cost-aware OOS · walk-forward · outlier-robust · FDR — the dyno itself
THE EXECUTOR
OutLayer Agent Custody signs `trade` and `anchor` from a TEE-held key, under the owner's policy · orders fill on Bitget Demo · Kelly-capped tickets
THE CONTRACTS
identity · reputation · validation · jobs · vault · anchor — one core contract on NEAR
THE RECORDER
every decision sealed: Supabase public table + NEAR anchor
THE FEEDERS
optional: POST /ingest takes positioning CSVs; scripts/88 refreshes the carry & whale caches in Supabase
THE PROOF
every LLM call runs on NEAR AI Cloud (TEE) and is signed; proofs are sealed with each desk record