Today's brief
The current run in full: what was screened, what was analyzed, what was decided, and the reasons behind each decision.
Proof · Research Trader
An agent that mostly says no
Once a day, a Noetfield agent screens a crypto watchlist, asks a model for a plan only on the few candidates that survive plain arithmetic, and then puts whatever comes back through a gate that asks the model nothing. Almost nothing gets through, and every rejection is published with its reasons.
One cycle a day · 15:00 UTC
Research only · places no orders
No performance claimed
A language model asked for a trade will produce one. It will produce one when the setup is good, and it will produce one when there is nothing there at all. The output looks identical in both cases: confident, fluent, specific. That is the failure mode behind most agent products, and it is not fixed by a better model or a stricter prompt.
It is fixed by refusing to let the model be the last word.
01 · Cycle
One cycle, five stages.
Deterministic work first, the model once, the gate last. The order is the control — the model appears at exactly one stage, and it is never the stage that decides.
Scan12 symbols · 200 hourly bars each · public market data
ScreenPlain arithmetic · no model call · most candidates end here
At most 3 candidatesUS$0.50 per cycleRefusal permittedCost recorded
Analyze · the only model call in the cycle
The model returns a verdict
Direction, entry, stop, target, setup score, confidence, thesis — or no setup at all
Output is a proposalNever a decisionUnparseable output is discardedCeiling enforced
RiskDeterministic gate · a pure function · asks the model nothing
ReceiptEvery symbol screened, every verdict, every decision and its literal reason codes
| Parameter | Value | Stage it binds |
|---|---|---|
| Watchlist | 12 USDT pairs — BTC, ETH, SOL, BNB, XRP, ADA, AVAX, LINK, DOT, POL, LTC, ATOM | Scan |
| Bars per symbol | 200 candles at one hour | Scan |
| Candidates sent to a model | At most 3 per cycle | Screen → Analyze |
| Model spend ceiling | US$0.50 per cycle. The cycle stops calling the model at that line. | Analyze |
| Schedule | One cycle per day at 15:00 UTC — 07:00 in Vancouver during Pacific Standard Time, 08:00 during daylight saving | Whole cycle |
| Mode | RESEARCH_ONLY. The mode is not read from the environment: changing it requires a code change and a review, not a variable on a runner. | Risk, first and unconditional |
Mirrored from the agent's configuration at commit 69a7427, 30 July 2026. If that configuration changes, this table is wrong until it is re-pinned.
02 · The gate
Every check, and the exact word it says when it fails.
The gate is a pure function — no model, no network, no clock. Every published rejection names the checks it failed using these strings, so a code in the brief can be traced back to a row here.
| Check | Requirement | Reason code emitted |
|---|---|---|
| Mode | Must be RESEARCH_ONLY. Checked first and unconditionally. | MODE_NOT_ALLOWED |
| Data freshness | Newest bar under 75 minutes old | STALE_DATA |
| Level validity | Entry, stop and target all finite and above zero | NON_POSITIVE_OR_NON_FINITE_LEVEL |
| Long coherence | Stop below entry, target above entry | LONG_STOP_NOT_BELOW_ENTRYLONG_TARGET_NOT_ABOVE_ENTRY |
| Short coherence | Stop above entry, target below entry | SHORT_STOP_NOT_ABOVE_ENTRYSHORT_TARGET_NOT_BELOW_ENTRY |
| Direction | Long or short. Neutral is not actionable. | NEUTRAL_DIRECTION_IS_NOT_ACTIONABLE |
| Setup present | The model must not have reported that there is no setup | NO_SETUP |
| Setup score | At least 70 out of 100 | LOW_SCORE |
| Reward to risk | At least 2.0, recomputed by the gate from the model's own levels rather than accepting the model's figure | LOW_REWARD_RISK |
| Volatility | Average true range at most 5% of price | HIGH_VOLATILITY |
| Position size | Derived size finite and above zero | INVALID_SIZE |
| Notional | Derived notional finite and above zero | INVALID_NOTIONAL |
One threshold does not reject. A verdict that clears every check above but carries confidence below 0.70 is recorded as WATCHLIST rather than APPROVED. And APPROVED does not mean take the trade — it means every machine check passed and a human should now look at it.
Three rules that have never fired
The gate also enforces total exposure, daily loss and drawdown limits, emitting EXPOSURE_LIMIT, DAILY_LOSS_LIMIT and DRAWDOWN_LIMIT. The agent holds no account, so the position and loss state those three read is supplied as zero on every published cycle and none of them has ever fired. They are listed because they are in the code, not because they are doing work here.
03 · A refusal
What a cycle actually looks like.
One decision, quoted from a recorded receipt exactly as it was stored. It is worth reading closely, because it is not the tidy story. The model was asked about three candidates. Two of its three answers came back unparseable and were discarded. The third said there was no setup — and the gate then rejected that non-plan on six separate counts, because a verdict with no direction and no levels fails almost every check there is.
Recorded decision · run
gha-31190377658-1
Rejected
- Cycle completed
- 7 August 2026, 15:02 UTC
- Symbols screened
- 12
- Sent to the model
- 3
- Discarded before a verdict
- 2 —
ANALYST_REFUSED:UNPARSEABLE_OUTPUT - Symbol decided
- ADAUSDT
- Model verdict
NO_SETUP· directionNEUTRAL- Levels proposed
- Entry 0 · stop 0 · target 0
- Model setup score
- 30 of 100
- Model confidence
- 0.20
- Gate decision
- REJECTED
- Reasons
NON_POSITIVE_OR_NON_FINITE_LEVEL·NO_SETUP·LOW_SCORE·LOW_REWARD_RISK·INVALID_SIZE·INVALID_NOTIONAL- Model spend, this cycle
- US$0.006 of a US$0.50 ceiling
This is the honest boundary of what has been shown so far. The published record demonstrates that the gate runs, that it recomputes rather than trusts, and that everything it refuses is recorded with reasons. It does not yet contain the more interesting case — a confident, well-formed plan that the gate overrules on the arithmetic. That case has not occurred in the published cycles, and this page will not claim it until it does.
A published rejection proves the gate ran and what it decided. It does not prove the decision was correct.
04 · The record
The totals, read from the feed as this page loads.
Every figure below names the field it came from, so it can be checked against the source rather than taken on trust.
The running totals are read from the live feed when this page loads. This copy of the page does not state them: a number printed here would be the number that was true when the page was published, not the number that is true now.
Aggregate across recorded cycles
Research only
- Cycles recorded
stats.cyclesRun- Screenings performed
stats.symbolsScreened- Verdicts the model returned
stats.plansWritten- Verdicts the gate approved
stats.plansPassed- Total model spend
totalSpendUsd- Observation window
- The 25 most recent receipts, not all time
The first published record, dated , covered eight cycles: a twelve-symbol watchlist screened eight times, seven verdicts returned by the model, none of them approved by the gate, and roughly three US cents of model spend in total. That sentence is fixed at its date and is not revised.
A verdict is not the same thing as a trade plan. The count above includes cycles where the model declined outright, so it should not be read as a tally of confident proposals the gate struck down. A pass rate in either direction is not by itself evidence that the thresholds are right: it is consistent with a correctly cautious system and also with one too strict to ever act. The feed is published so that this stays checkable rather than asserted.
Check it yourself.
The raw feed
The machine-readable state this page reads. It answers from this domain, so it still responds — and says why — when the agent itself is unreachable.
The evidence register
What Noetfield claims is proven, what is demonstrated internally, and what is not established at all.
Why we run it
Noetfield sells deterministic controls around model output: structure that is checked rather than hoped for, with the failure caught before the answer leaves the system. That claim is easy to make and hard to show, because the interesting case is the one where the model is wrong and something stops it.
Markets supply that case every day, in a domain that punishes false confidence. So we pointed the same machinery at it, put it on a fixed schedule, and published the result whether or not it flatters us.
The same controls are available directly. See the deterministic API →
What this is not
It does not trade
It holds no positions and no money, and is not connected to an exchange, a broker or any account. The mode that would be required to approve anything for execution is not settable from the environment.
It does not recommend
It is not investment advice and it is not a recommendation service. Nothing on this page or in the daily brief is tailored to any person's circumstances, and no one should treat a published brief as a reason to buy or sell anything.
It claims no performance
No performance is claimed. There is no track record, no return figure, and no backtest presented as one. The agent has never taken a position, so it has never been right or wrong about one.
It is not independently verified
The market data comes from public sources and is not independently audited. The gate enforces the rules it was given; it does not know whether those rules are the right ones.
The brief is published free and can be read without an account. The agent's own site also carries a landing page with a paid tier: it is not active, no payment is taken, and no subscription can be started today.