Skip to content
TVIJO / RESEARCH ATLAS 2.0.0-candidate

A research atlas for three projects

Choose a testable configuration for a task. Connect projects, acceptance criteria, benchmarks, and a plan for your own evaluation.

3 projects8 evaluation lenses0 own snapshotsCatalog: 2026-11-09
Tasks and decision criteria

Project × evaluation lens

Select a cell: task → acceptance → metric → configuration.

Candidate · plannedUnverifiedReady to testMeasurement existsColor denotes readiness, not a quality score.
Project tasks by evaluation lens. Buttons open details below.
Project / lensProbability & calibrationHow uncertain is the decision, and what is the cost of an error?Numerics & optimizationAre the formulas, constraints and numerical answers correct?Data science & evidenceAre the data and conclusions supported by traceable evidence?Marketing & causal experimentsDoes the product improve qualified conversion and customer value?Unit economics & macro scenariosWhat are contribution, cash flow and sensitivity to external conditions?Management & deliveryAre tasks accepted on time with clear ownership and dependencies?Bounded autonomyWhich authorized actions can finish without intervention?Controlled self-improvementDoes a versioned candidate improve results on protected evidence?
Tvijo AIOS
Hotido.com
FinEdg

Matching relationships: 24. On narrow screens, scroll the matrix horizontally.

Tvijo AIOS / Probability & calibration

Calibrated task routing

Candidate · planned

Concrete task

Estimate uncertainty before choosing a fast route, escalation or review. Choose the authorized action with the lowest expected loss; validate confidence against held-out outcomes.

Acceptance rule

Proposed gates, to freeze before running: Upper bound of the paired 95% confidence interval for ΔBrier (Chosen-answer correctness Brier; candidate − FROZEN baseline) ≤ 0; Critical violation count observed = 0. Freeze the baseline identity, version/hash and comparison protocol before any run; without them the comparative gate is not assessed. Insufficient data: not assessed.

What to measure

  • Chosen-answer correctness Briersquared_probability · lower is better
  • Reliability diagrambin_probabilities · descriptive metric
  • Accepted task ratefraction · higher is better
  • Critical violation countcount · lower is better
Expected project impact — hypothesis

Hypothesis: improve cost_per_accepted while monitoring accepted_task_rate. No measured impact yet.

Proposed pilot scope: Isolated evaluation environment

Planned, not an executed run: 200 independent task/incident · 3 repetitions per unit. Repetitions do not increase independent n.

Proposed pilot: 200 independent units × 3 repetitions. A design starting point, not a power claim or completed work. No measured model winner; scope sandbox_only.

Prepare an experiment
Evidence status matters.

A documented feature, a test candidate, and a measured result are different states. Session cost does not measure success, and coding-agent performance does not establish trading profitability.

Turn the matrix into an evidenced decision

  1. Pin the model, provider, harness/runtime versions, effort, tools, memory, scope and complete budget.
  2. Predeclare the observation unit, acceptance criteria and protected dataset. Repetitions of one task remain dependent.
  3. Evaluate candidates on the same cases. Preserve failures, latency, intermediate calls and cost records.
  4. Check mandatory constraints first. Then compare quality with uncertainty, complete cost and time among eligible candidates.
  5. Use randomization and mature conversion for marketing; chronological windows, data vintages and PAPER for FinEdg.
  6. Move a candidate through shadow, sandbox and a bounded pilot. Preserve a protected holdout, experiment budget and verified rollback.

The 24 project protocols in this version are proposals. No measured Atlas snapshot has been imported for these protocols; local fixture trials are reported separately.