Genesis v0.3.0 -- Token economics as first-class design dimension
Markdown that steers an LLM is code -- now it has a cost vocabulary too.
This release adds token economics as a structural concern that
the architect treats the same way it already treats threading,
persistence, attention, and grounding: substrate vocabulary, named
patterns, process touchpoints at steps 1 / 3.2 / 6 / 8, and a
falsifiable handoff-packet artifact.
Highlights
- Two orthogonal operator knobs.
stance(frugal | balanced |
quality | unbounded) shapes the SHAPE of the design. Cap (dollar /
token / premium-request hard ceiling) shapes the SIZE. - Seven new catalogue entries. A12 GRADIENT WORKFLOW;
B12 MODEL ROUTER, B13 CACHE-AWARE PREFIX, B14 PROMPT THRIFT,
B15 TOOL SUBSET, B16 EFFORT GOVERNOR; R5 COST PRUNE. - Substrate vocabulary. Cacheable prefix, variable suffix,
output tax, cache breakpoint, cache invalidator, TTL tradeoff,
per-harness billing unit. Plus role-class taxonomy (planner /
implementer / reviewer / trivial / long-context-retriever). - Step 6 cost-projection artifact. Six parts -- bands are
CONTRACT (auditable at step 8), dollar ranges are PREDICTION
(sourced from per-harness adapter pricing footnote, date-stamped).
Empirical proof
Same operator prompt as the existing examples/04-pr-review-advisory,
run conceptually against both v0.2.0 (BEFORE) and v0.3.0 with
stance: frugal, cap = $1/PR (AFTER):
| Scenario | BEFORE (v0.2.0) | AFTER (frugal) |
Factor |
|---|---|---|---|
| M, warm cache | $2.145 / PR | $0.143 / PR | ~15x cheaper |
| M, cold cache (first run of TTL) | $2.145 / PR | $0.247 / PR | ~8.7x cheaper |
| L (50K-token diff), warm cache | $6.850 / PR | $0.615 / PR | ~11x cheaper |
Same A1 PANEL structure preserved -- cost discipline does not break
threading discipline. Pricing dated 2025-11-14, verified within 90
days. Full per-component arithmetic and reproduction steps in
dev/empirical-proof/scenario-pr-review-panel.md.
Docs
New GitHub Pages category Reference > Token Economics with four
pages (Overview, Operator stance & cap, Cost projection artifact,
Cost patterns). New worked example 06. Cost-aware panel under
Resources > Examples.
Browse: https://danielmeppiel.github.io/genesis/reference/token-economics/
Validation
Four cold sub-agent audience panels (CTO / Engineer / DevUX /
FinOps) reviewed the design. CTO: SHIP -- mandate it. Engineer:
SHIP WITH FIXES (3 BLOCKERs, all addressed in convergence iter 1).
FinOps: DEFENSIBLE -> RIGOROUS (arithmetic blocker, fixed). DevUX:
FRICTIVE-but-fixable (P0/P1 deferred cosmetic).
Five of six new positive eval scenarios pass; the one miss
(p-006) is a capitalization-only substring miss with the pattern
referenced correctly throughout.
Related work
The Rosetta-stone mapping between genesis patterns and classical
GoF / architecture vocabulary was extended in parallel for the
agentic-sdlc-handbook -- four new patterns mapped, three rejected
with reasons: danielmeppiel/agentic-sdlc-handbook#11.
Breaking changes
None. The corpus is additive. Existing designs continue to work;
the new fields default to absent. stance defaults to balanced,
which corresponds to genesis's prior behavior on most decision
points.
Trade-offs noted
- Pattern renumbering. Issue #9 proposed B11-B15 and A10; we
shipped B12-B16 and A12 because A10 / B11 were already claimed
during the design window. No semantic change; references
internally consistent. - SKILL.md +45 lines over the 500-line nominal budget. Existing
tendency; new step 3.2 + step 6 fields are the marginal addition. - DevUX P0/P1 deferred to v0.4 (vocabulary load, 4-stance enum
collapse possibility, cap-halt timing, dollar precision). Cosmetic,
documented as forward-looks in the docs themselves.
Pull request
See #10 for the
full design report, audience-panel transcripts, eval runs, and
file-by-file rationale.
Closes
- #9 -- Integrate token economics as first-class design dimension