Kalkasautonomous decisions

Domain modules

Three modules, one decision path.

Each module answers the same kernel contract with different material: what counts as evidence, which gates must pass, and what is never requested. The loop itself does not change between domains.

Walk the loopOpen the consoleDomain reference

The three modules

Sport and racing, quant research, financial analysis.

Three modules today, one decision path underneath: the same six stages, the same mandate rules, and the same receipt shape for every run.

Full-field, pre-off, source-bound

Sport and racing

Ranked beliefs across a whole field, priced against the quotes that were reachable before the off.

Runs bind a sealed point-in-time racing corpus, produce a ranked belief over the full field, and gate every wager through stake bounds and ruin policy before the decision is emitted. Settlement is the published result; learning records whether the belief or the price was wrong.

supplies

  • Chronological, pre-off history sealed with source authority and digest
  • Quote snapshots per runner, with availability times
  • Published results used only for settlement, never for inference

gates

  • Stake bounds and maximum-loss policy per run
  • Ruin policy evaluated before any stake is requested
  • Out-of-sample holdout required before a candidate can act

never

  • No account, wager, or funds reach a sportsbook or exchange
  • No claim of profit, edge, or strike rate
  • No outcome data available before the off
Read the sport and racing module

Out-of-sample, cost-aware, merge-gated

Quant research

Research decisions where promotion depends on held-out evidence, not on in-sample fit.

Distributed training merges candidate parameters across shards, then evaluation compares the candidate against the baseline on the same sealed out-of-sample cases. Promotion requires beating the floor after deflation; in-sample fit alone can never promote a candidate.

supplies

  • Immutable point-in-time market and research corpus
  • Artifact and parameter lineage per shard
  • Development and out-of-sample case identities kept disjoint

gates

  • Out-of-sample coverage floor with abstentions counted
  • Deflated performance floor after selection cost
  • Capital-risk admission before any position intent

never

  • No broker, exchange, or capital path is contacted
  • No promotion from in-sample fit alone
  • No synthetic digest standing in for real lineage
Read the quant research module

Knowledge-dated, restatement-aware

Financial analysis

Analytical stances that respect knowledge dates, restatements, and survivorship.

Facts are admitted at the date the market could have known them, not the date the report was printed. Restatements and delistings are modelled rather than papered over, and the run ends in an analysis stance with a report projection — never an order.

supplies

  • Knowledge-date admission for every reported figure
  • Restatement and survivorship treatment recorded per fact
  • Provenance chain from filing to belief

gates

  • Mandate-safe stance under declared analysis bounds
  • Settlement and evaluation bound to the same run
  • Corrections append rather than rewrite history

never

  • No order path and no execution adapter
  • No recommendation presented as advice
  • No figure admitted before it was knowable
Read the financial analysis module

Comparison

What changes between modules.

Read the three modules side by side: the evidence each one supplies, the gates it enforces, and the effects it never requests.

Same loop, different material: what each module supplies, gates, and never does.
ModuleEvidenceGatesNever
Sport and racing
  • Chronological, pre-off history sealed with source authority and digest
  • Quote snapshots per runner, with availability times
  • Published results used only for settlement, never for inference
  • Stake bounds and maximum-loss policy per run
  • Ruin policy evaluated before any stake is requested
  • Out-of-sample holdout required before a candidate can act
  • No account, wager, or funds reach a sportsbook or exchange
  • No claim of profit, edge, or strike rate
  • No outcome data available before the off
Quant research
  • Immutable point-in-time market and research corpus
  • Artifact and parameter lineage per shard
  • Development and out-of-sample case identities kept disjoint
  • Out-of-sample coverage floor with abstentions counted
  • Deflated performance floor after selection cost
  • Capital-risk admission before any position intent
  • No broker, exchange, or capital path is contacted
  • No promotion from in-sample fit alone
  • No synthetic digest standing in for real lineage
Financial analysis
  • Knowledge-date admission for every reported figure
  • Restatement and survivorship treatment recorded per fact
  • Provenance chain from filing to belief
  • Mandate-safe stance under declared analysis bounds
  • Settlement and evaluation bound to the same run
  • Corrections append rather than rewrite history
  • No order path and no execution adapter
  • No recommendation presented as advice
  • No figure admitted before it was knowable

Also on the kernel

Forecast packs for probability questions.

These packs answer graded and labelled questions where the outcome is a probability rather than a price. The lifecycle, the mandate, and the receipt are the same as for any other module.

Threshold

Binary events against a declared threshold, with formula or statistical belief.

Weather

Multiclass forecasts where the outcome is one of several graded conditions.

Sports event

Two-sided event markets with their own quote snapshot, stake gate, and settlement.

Analysis

Reusable binary forecasts for elections, weather, and other labelled questions: belief, recommendation, and abstention without execution.

Read the module reference, then run one.

The reference carries the full contract for every pack. The quickstart runs a module locally against sealed corpora, so you can watch the gates decide before anything is requested.

Domain referenceQuickstart