kestrel.markets
Version:
A typed, token-efficient language + runtime for agentic trading: agents author bounded plans, the runtime fires them at the tick. CLI + typed library + MCP server.
35 lines (30 loc) • 2.4 kB
Markdown
# Grade: everything authored is gradable, and names are data
> **Cross-reference (2026-07-16).** The **contamination fence** in honesty rule (4) below
> — days before an LLM author's training cutoff are practice-tier only; honest grades come
> from post-cutoff, date-blinded replay — is **formalized and overtaken** by ADR-0040
> (*clock-honest wakes*, which adds the latency-blind/latency-claim leg to the same
> season-honesty machinery) together with the platform's season-governance ADRs (which
> home the practice/served/sealed season taxonomy and the season-validity set-law). This
> ADR states the fence as a principle; those ADRs are where its enforcement now lives.
Grade's scope is a single principle: **if you can author it, you can grade it** — a
Plan (realized EV through fills), a Wake or Scan (attention efficiency: tokens per good
decision), a View (do agents decide better through it than a baseline), a regime tag
(gated-beats-pooled), a Pod node (the PM's allocation decisions, distinct from its
leaves' trading). This replaces "backtesting a strategy" with grading judgment at every
level of the org tree, because in Kestrel the agent — not a parameterized algorithm —
is the thing being evaluated.
Four honesty rules, all learned the hard way in live trading: (1) **every
grade stamps its judge** — fill-model name+version, corpus, blinding; EVs across
fill-model versions refuse naive comparison; (2) **counterfactuals are syntax, not
bespoke scripts** — `VS ungated`, `VS null` (the structural null crosses the SAME fill
model), `BRACKET` (floor/expected/ceiling fills); (3) **cells, not pools** — `BY`
stratification is first-class because pooled averages wash out exactly the cells where
edges live; (4) **the contamination fence** — when the graded author is an LLM, days
before its training cutoff are practice-tier only; honest grades come from post-cutoff,
date-blinded replay (the weights leak; no code fence stops them).
**Names are data.** A recurring free-text name is a lineage (one strategy, many
authored instances) and the natural aggregation key: `BY name` / `BY lineage` are
first-class grade dimensions, and clustering names with their graded outcomes is the
emergent-taxonomy mechanism — discovered strategy families and sub-regimes graduate
into the Armory with provenance. Emergence through authorship; discipline through
grading.