Skip to content
RAZOOM is in active development right now — join the early access list
RAZOOM
RAZOOM

Menu

Request Early Access

The RAZOOM Method

Better answers begin with better ways of attacking the problem.
Illustration: a difficult problem is matched against eight methods — Reason, Research, Calculate, Model, Simulate, Optimize, Measure, Experiment. All eight stay under consideration as the signal moves between them; research, model and simulate are chosen and form the attack plan.
  1. 1Understand
  2. 2Clarify
  3. 3Formalize
  4. 4Decompose
  5. 5Select Method
  6. 6Investigate
  7. 7Challenge
  8. 8Synthesize
  9. 9Return outcome

A philosophy, not a prompt template.

The RAZOOM Method describes how difficult problems should be approached: understand the real question, ask only what matters, structure the problem, choose how each part should be attacked, and challenge the result before trusting it.

It is deliberately public. The value is not in knowing the steps — it is in applying them consistently, choosing well between methods, and measuring whether the choice was right.

Internally, new ideas are tested first in our research methodology, the Oracle Protocol, before they reach the product.

Cognitive strategies.

Different ways of attacking a problem, chosen to fit it. The point is cognitive diversity — not how many agents are involved.
  • Blind BrainstormIndependent approaches before models see each other's conclusions.

    Several approaches are developed in isolation, then compared. Agreement reached independently is evidence; agreement reached by imitation is not.

    Guards against: Anchoring and premature convergence.

  • Hypothesis EcologyMaintain competing hypotheses and allocate attention according to evidence.

    Several explanations stay alive at once. Each new piece of evidence shifts attention towards the hypotheses it supports, instead of being forced into the first story.

    Guards against: Locking onto the first plausible explanation.

  • Assumption AttackIdentify assumptions with the highest leverage and test them first.

    Assumptions are ranked by how much the conclusion depends on them. The most load-bearing ones are challenged before effort goes anywhere else.

    Guards against: Polishing conclusions that rest on untested premises.

  • Minority ReportProtect a competent dissenting analysis from majority suppression.

    When one well-reasoned analysis disagrees with the rest, it is preserved and examined on its merits rather than outvoted.

    Guards against: Consensus that is wrong in the same way.

  • Double-CruxIdentify what fact would cause competing analyses to change their conclusion.

    Disagreements are reduced to the specific fact each side would need to see to change its mind — which is usually the next thing worth finding out.

    Guards against: Disagreements that are argued instead of resolved.

  • Counterfactual World FarmTest recommendations across plausible alternative worlds.

    The recommendation is evaluated in a range of plausible futures. A good decision should survive most of them, not only the expected one.

    Guards against: Decisions optimized for a single forecast.

  • Destroy → Salvage → RebuildConstruct, attack, retain surviving parts and rebuild.

    An initial analysis is deliberately torn down. What survives the attack is kept; the rest is rebuilt on firmer ground.

    Guards against: Attachment to a first draft.

  • Semantic RepulsionDeliberately search for useful solutions distant from obvious ideas.

    Once the obvious options are on the table, the search is pushed away from them on purpose to find alternatives that would otherwise never be considered.

    Guards against: Option sets that are variations of one idea.

  • Cross-Domain TransplantTransfer structural solutions from unrelated domains.

    The problem's structure — not its subject — is matched against fields that have solved similar structures, and the solution is adapted.

    Guards against: Reinventing solutions that exist elsewhere.

  • Information AuctionEstimate which next piece of information has the highest decision value.

    Candidate questions, searches and measurements compete on expected decision value per unit of cost. The winner is pursued first.

    Guards against: Gathering information because it is available, not because it matters.

  • Strategy TournamentTest several reasoning approaches cheaply before allocating heavy compute.

    Candidate strategies are tried at small scale. Compute is then concentrated on the approach that shows the most promise for this problem.

    Guards against: Spending heavily on the wrong approach.

Some problems should be solved, not discussed.

When a formal tool can answer more reliably than a language model, RAZOOM should use the tool. These are the 36 methods it is being designed to route problems to.

Statistics and inference

What does the data actually support?

  • Statistics
  • Correlation and dependence
  • Regression
  • Bayesian analysis
  • Causal inference
  • Econometrics
  • Time-series analysis
  • Forecasting

Simulation

What could happen, and how likely is it?

  • Monte Carlo simulation
  • Discrete-event simulation
  • Agent-based modeling
  • System dynamics

Optimization

What is the best feasible allocation?

  • Linear optimization
  • Nonlinear optimization
  • Integer optimization
  • Robust optimization
  • Multi-objective optimization

Decision analysis

Which choice is best given uncertainty and values?

  • Decision theory
  • Expected utility
  • Regret
  • Value of Information
  • MCDA

Operations and reliability

Will the system hold, and where will it fail?

  • Operations research
  • Queueing theory
  • Capacity planning
  • Reliability analysis
  • FMEA
  • Stress testing

Systems and structure

How do the parts interact over time?

  • Network science
  • Dynamic systems

Formal reasoning

Is this provably true, consistent or satisfiable?

  • Constraint solving
  • SAT/SMT
  • Symbolic mathematics
  • Formal verification

Estimation

Is this even the right order of magnitude?

  • Fermi estimation
  • Dimensional reasoning

A language model is a poor calculator and an excellent interpreter. RAZOOM is designed to use each for what it does well.

What RAZOOM aims to learn.

For this type of problem, under these conditions, with these unknowns — which problem-solving method performs best?
  1. 1

    Problem structure

  2. 2

    Method selection

  3. 3

    Execution strategy

  4. 4

    Case outcome quality

The lasting asset is not a library of prompts. It is the evidence that links the structure of a problem to the method, and the method to the quality of the Case outcome — measured, not assumed, through RAZOOM Bench.

Test the method on a real problem.

Early-access participants will help us find where the method works — and where it does not.

Join Early AccessExplore RAZOOM Bench