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.
28 lines (24 loc) • 1.84 kB
Markdown
No rendering format is ever blessed by taste. Any concrete choice of tape-row syntax,
glyph vocabulary, pane layout, or serialization style is a **hypothesis** that must
earn default status in a rendering tournament: cohorts of live agents run on
**identical Frames** over identical replay corpora, differing only in Rendering
variant, and are graded (ADR-0006) on decision quality, token cost, comprehension
probes, and attention efficiency — against each other and against the honest
baselines (raw JSON, ad-hoc prose narration). Cross-agent convergence produces
defaults; per-model divergence produces per-tokenizer defaults. The decision was
forced by direct experience: two informed designers held different confident mental
models of "the obviously right" tape row — so neither is data.
What is *not* up for grading (the invariants renderings compete inside, per the
platform-scope test): the renderer invents no value; the acting kernel is present;
streaming renderings are cache-monotone/append-only; the legibility floor (never
compress past interpretable units into opaque symbols); token budgets measured under
the declared tokenizer. Everything else — including ADR-0008's relative-candle row
format, which is hereby explicitly a **first candidate, not doctrine** — is a variant
in the pool.
Consequence: the Frame/Rendering split is load-bearing for experimentation itself —
because variants serialize the same typed Frame, a rendering tournament is cheap
(N renderers over one replay corpus, no new data, no new truth), and it is the
standard mechanism for evolving the screen, permanently. Adding a variant to the pool
is taste-free; winning is not. The live pool is seeded in
[`docs/rendering-variants.md`](../rendering-variants.md).