Join the waitlist

A researcher that shows its sources.

The agent plans experiments and reads the results, and every figure it quotes links back to a deterministic engine run.

Join the waitlist How findings work
app.evrantlabs.com/research
You

Why does the ORB lose money on Fridays?

AIAgent
Ran decomposition by weekdayStudy
Read resultsBacktest run

Friday is the only weekday with negative expectancy, per trade over trades, against overall. The losses sit in entries late in the morning. I have not measured how many reverse through the range: no source for fri.reversal_share.

Experiment proposalawaiting you

Run a study of late Friday entries so the reversal share has a source

ApproveDismissNothing changes until you approve.

Illustrative session. Figures are examples, not performance claims.

The agent researches. The engine decides.

Anything shown as a result comes from a deterministic engine job. The agent's writing is stored beside it as interpretation.

AreaThe agentThe engine
IdeasStructures your idea into a thesis, theory and specification.Stores each version immutably, with its parent and the reason it changed.
CodeWrites and repairs C# against the Strategy SDK.Compiles, verifies and executes it in a sandbox.
ExperimentsProposes the experiment and states its compute cost.Runs every backtest, study, optimisation and Monte Carlo path.
NumbersQuotes results as references to artefact fields.Produces every metric, and renders each reference from its artefact.
DataAsks for development data within your budgets.Enforces the split plan, exposure counts and leakage checks.
TradesNever takes part in an individual trading decision.Runs strategy logic and model inference deterministically.

Every claim has a type and a source.

The type decides what a finding must cite, and numbers inside it are references, not text.

  • Observations and hypotheses can stand without citations
  • Evidence and conclusions must cite engine results
  • Hover a number to see the run field behind it
  • Findings are never edited. Revisions supersede them
Findingsbase-orb v45 findings · 1 superseded
  1. ConclusionTime of dayfnd_01J9ORB038

    The Friday weakness sits late in the morning and appears in both halves of the development range.

    Citessupportsrun_01J9ORBDEMO07supportsrun_01J9ORBDEMO09

  2. EvidenceExperiment resultfnd_01J9ORB035

    Skipping Friday entries after 11:00 moves expectancy from to .

    Citessupportsrun_01J9ORBDEMO07interpretsrun_01J9ORBDEMO01

  3. HypothesisParticipationfnd_01J9ORB032

    Late-morning Friday breakouts fail because participation thins before the weekend.

    Tested byexp_01J9ORB019supersedesfnd_01J9ORB029

  4. Hypothesisfnd_01J9ORB029Superseded

    Friday losses come from lower volatility at the open.

  5. ObservationWeekday splitfnd_01J9ORB031

    Friday entries have the lowest expectancy of any weekday: .

base-orb
Supports the thesis
9
Contradicts it
5
  • Findingfnd_01J9ORB027Open

    Volume filter improves expectancy

    contradictsrun_01J9ORBDEMO12

    Run on 2020 alone: the filter removes the strongest month

  • Findingfnd_01J9ORB032Open

    Friday participation thins after 11:00

    contradictsrun_01J9ORBDEMO11

    Friday volume after 11:00 is within the range of other weekdays

  • Findingfnd_01J9ORB018Superseded

    Edge is stronger on volatile days

    contradictsrun_01J9ORBDEMO05

    The top volatility decile has a negative mean and the widest interval

Bad news gets the same billing.

Contradicting results stay open in the evidence graph, and every new version is checked against your original thesis.

  • Contradictions stay open until someone deals with them
  • Evidence for and against is shown at equal weight
  • Each version is classified from fix to departure

Autonomy on a budget you set.

A campaign runs the research loop without you, inside limits on experiments, compute, agent credits and what the agent may see.

  • Stops when the hypothesis is falsified or the evidence is strong enough
  • Stops when a budget runs out or experiments stop adding findings
  • Ends with a summary finding that links every experiment
  • Never touches sealed data, candidates or paper sessions
Bounded researchcmp_01J9ORB02Active
Inspect evidencePropose experimentRun analysisInspect resultUpdate hypothesisPropose next

Is the Friday effect real?

Experiment 1

Budget used

Backtests0 / 12
Compute0 / 3,600 s
Agent credits0 / 60

    Permitted methods

    • Backtest
    • Optimisation
    • Monte Carlo
    • Cost stress
    • ML comparison

    Structural changes, extra scope and validation are outside this authorisation.

    The same tools in any MCP client.

    Your own MCP client runs the same commands as the in-app agent, under your authorisation, budgets and split enforcement. Approvals and sealed access stay with you.

    Your MCP clientevrant · 101 tools
    1. whoamireadmember · workspace research · 6 scopes
    2. get_strategy_workflowreadbaseline accepted · implementation review next
    3. submit_strategy_backtestcomputequeued · job_01J9ORB07
      { "versionId": "28d795ee-…", "timeframe": "PT5M",
  "rangeStartUtc": "2021-12-06", "baseline": false }
    4. wait_for_jobreadaccepted · run_01J9ORB07
    5. get_run_resultsreadaccepted metrics
      { "runId": "run_01J9ORB07", "resultHash": "sha256:9f2c41e7…",
  "metrics": { "expectancyR": 0.14, "trades": 1284, … },
  "artifacts": ["trades:7c41d8", "equity:7c41d8"] }
    6. request_mandate_authorizationmandaterefused · INVALID_ARGUMENTS

      Every mandate range must be development history. Validation and sealed history is reserved.

    7. request_human_decisionresearchmandate_authorization · awaiting you

      Recorded in Reviews → Agent requests: authorize mandate

    Results come back as artefact ids and hashes, never pasted data.

    Illustrative session; tool names and scopes are from catalogue 2026-10-21.5. ML training and studies run inside a mandate you authorise.

    Early access

    Work with a researcher that shows its sources.

    Join free for an early-access invitation. No account or payment required. We’ll confirm market coverage before you choose a plan.

    For your early-access invitation only. Request removal anytime at [email protected]. See our privacy notice.