Second Order
Book a call

SecondOrder

Stochastic search.
by the AI model
Deterministic validation.
by your code, on your criteria
Researcher ownership.
you, at every step
Keep your edge.
open weights, full privacy
§1The research loop

One loop. Nothing taken on trust.

The AI model proposes. Everything else is executed, judged and recorded.

Algorithm 1
  1. 1:loop
  2. 2:h ← Hypothesis(Memory)The AI model proposes, informed by memory.stochastic
  3. 3:r ← Experiment(h, your code)Your code runs it. Traced and hashed.deterministic
  4. 4:v ← Verdict(r, your criteria)Your criteria decide what good research is.deterministic
  5. 5:i ← Interpretation(h, r, v)The model says what it learned, in plain words.readable
  6. 6:Memory ← Memory ∪ {h, r, v, i}Everything is kept, failures included.persistent
  7. 7:end loopMemory returns to the next hypothesis.
§2Domain expertise

The AI model becomes an expert in your R&D.

A research memory we built recalls every experiment before the next proposal. The model stops being generic and starts knowing your field. The memory stays yours.

(a)First question(b)100 experiments(c)One year
Figure 1.How expertise compounds. (a) A general model. (b) It knows your dead ends. (c) An expert in your field.

Table 1.What it remembers.

Experimentscode, verdict, interpretation
Failureskept, so nothing is tried twice
Claimsfrom your papers, cited
Lineagewhat builds on what
Figure 2.Research graph: 8 nodes, 8 relations. Illustrative.HypothesisDatasetExperimentResultFailed approach
§3On your terms

Your files. Your functions. Your models. Your servers.

Exactly what you provide is what runs. Every experiment proves it.

  1. (i)

    Your data

    Exactly your files.

    Internal panels, vendor feeds, papers. No substitutes, no copies.

  2. (ii)

    Your functions

    Your code, executed.

    Backtester, cost model, risk: real pipeline stages, never imitated by the model.

  3. (iii)

    Your models

    Own the weights.

    Keep your alpha and control your cost: open-weight models you own, run inside your perimeter.

  4. (iv)

    Your servers

    Dedicated or on-premise.

    One tenant, one machine. Nothing leaves it.

3.1Proof of use

What you provide is what runs.

Every experiment carries a trace: each stage, whose code, the file’s hash, the output. Change a file, the hash changes.

Table 2.Experiment trace. Illustrative.

#stagesourcefilehashoutput
1featuresyoursfactors_lib.pya41f9cfeature panel
2signalmodelhypothesis spec7d02e1positions
3backtestyoursbacktester.pyc9b3a0P&L series
4riskyoursrisk_model.pyf15e77risk report
5judgeyour criteriacriteria.yaml0be4d2✓ accepted

3.2Reinforcement learning

Your own AI model, trained by your own research.

Not a bigger model. An open one, post-trained inside the loop, that learns what works in your field and what does not.

  1. (a)Trained on your experiments

    Every experiment is a labelled episode. Your own research becomes the training set.

  2. (b)Rewarded by your code

    The deterministic judge is the reward. No model opinion, no way to argue with it.

  3. (c)Kept in your perimeter

    An open model, post-trained on your servers. Your expertise never leaves.

§4First domain

Quantitative finance, first.1

High stakes, heavy R&D, results that rarely replicate. The place to be creative without being wrong.

  • Market data
  • Signal research
  • Statistical analysis
  • Backtesting
  • Portfolio research
  • Systematic strategies
in-sampleout-of-sample
Figure 3.Experiment 0142, walk-forward validation; cumulative, normalised. Candidate signal (solid) against the reference (dashed); the hatched band is out-of-sample. Illustrative.

1Second Order is research infrastructure. It is not an investment product, does not provide investment advice, and makes no claim about financial performance.

Private preview

See the loop run.

Book a call. We run the loop live on a question from your field.

Book a demo callOpen the demo