Skip to content

AI_MODEL_ECONOMICS.md

Latest commit

 

History

History
153 lines (103 loc) · 6.5 KB

AI_MODEL_ECONOMICS.md

File metadata and controls

153 lines (103 loc) · 6.5 KB
Raw

Genetic System — AI model economics

Genetic tag: docs.public.genetic_system.economics.gen1

This document summarizes labor vs token economics for the Genetic System, synthesized from internal economics canvases (genetic-system-ai-model-economics, agentstack-genes-release-economics, agentstack-platform-sdk-leverage) and the public genetic-ai-starter harness.

Read this with caveats. Numbers are models and simulations unless labeled as measured in CI. They inform prioritization — they are not guaranteed savings on your calendar.

Two cost lines

Line What you pay Genetic System lever
Labor Engineer time, calendar, rework Fewer wrong-tree edits, faster onboarding, less grep archaeology
Tokens API spend per agent turn Shorter stable prefixes, less blind context, better cache hit rate

Primary ROI is labor, not shaving tokens alone. Token savings are a secondary benefit when prefixes stabilize (map → index → hot file).

Philosophy compression (measured)

Metric Value Source
Raw philosophy context ~103,222 tokens AgentStack gene-access bench (bench_gene_access.json)
Gene-indexed access path ~8,350 tokens same
Compression ratio 12.36× same

Scope: philosophy / gene access — not “compress your entire codebase.” The win is addressed reads instead of dumping all genes.

Platform inventory (measured snapshot)

Regenerate via genetic-ai-starter export-platform-stats.mjs (Jul 2026 export):

Metric Value
Genes 406
AI_INDEX.md (repo / platform) 186 / 162
Tier-1 map tags 421
Kit payload genes 27

Cross-cluster SYN ~16 is from the interactive genetic-system-site narrative, not the snapshot.

Navigation harness (synthetic)

From genetic-ai-starter shop-api fixture transcripts (weak baseline vs kit + indexes):

Scenario Score (0–10) Pass rate
Weak (no map) ~2.5 0%
Kit + indexes ~9 100%

Interpretation: effect order transfers to real repos; absolute scores depend on project shape. Run npm run harness in your clone after init.

EST gene harness (platform internal)

Metric Before After (indexed)
Navigation failure rate ~22% ~5%
Token factor (discovery) — ~0.62
Retry factor — ~0.85

These are engineering estimates used in release economics canvases — not production A/B.

Break-even and Monte Carlo

Model input Typical value Notes
FTE-week cost ~$3,500 Release economics canvas default
Break-even touches ~17 Feature-sized tasks touching the map
Monte Carlo P(save>0) 1.0 AgentStack Monte Carlo release-cost simulation with wide jitter

Caveat: Monte Carlo with generous uncertainty still shows positive expected value because wrong-tree rework dominates at scale. Your team rate and task mix may differ.

Release archetypes (weeks saved, order of magnitude)

From agentstack-genes-release-economics canvas — FTE-week = $3,500, turn-cost model:

Archetype Description Indicative week savings
A Greenfield app on AgentStack + kit Highest — map + SDK compound
B Brownfield add feature (500–2k files) Medium-high — discovery tax cut
C Large monorepo (5k+ files) High — blind grep fails without map
D SDK-only consumer Medium — leverage table below
E Docs / ops / KB only Medium — same invariant, non-code artifacts

Exact week ranges vary by team; see genetic-ai-starter VALUE_AND_ROI_BY_PROJECT_SIZE and DOC_CLAIMS_AUDIT.

Platform SDK leverage (calendar)

From agentstack-platform-sdk-leverage canvas — weeks not spent rebuilding:

Module Typical save Why
Auth + sessions 2–4 w Hosted identity, JWT, project scope
Payments / wallet 2–5 w agUSD, MCP commerce tools
8DNA / project data 1–3 w Genetic project records, not ad-hoc JSON
MCP + agentstack.execute 1–2 w Tool catalog vs bespoke integrations
Dual-shell SPA 3–6 w Audience, nav, view-as, pages map
RAG / neural cache 1–3 w Platform substrate vs DIY vector stack

Genetic navigation stacks on top: SDK removes build weeks; map removes find-and-fix weeks inside what you still own.

Model and agent landscape (2026)

Factors that increase the value of stable addresses:

Signal Implication for navigation OS
Context rot Long prompts hurt accuracy before window limits — prefer 2-file index reads
Prompt caching Stable map → index prefix caches; grep roulette does not
METR TH1.1 Autonomous horizons doubling ~every 89 days — errors compound over longer runs
Multi-agent fleets Shared tags prevent agent A and B patching different trees

What we do not claim

  • Genetic tags do not replace tests, code review, or security review.
  • Map maintenance has cost — threshold ~10+ integration points or non-obvious boundaries (see docs/AI_INDEXING_SYSTEM.md).
  • 12.36× is not “12× faster development” — it is philosophy access compression.
  • Harness 100% is a synthetic fixture — use as regression guard, not a sales guarantee.

Audit trail

Artifact Repo
metrics.snapshot.json genetic-ai-starter
platform-stats.snapshot.json genetic-ai-starter
DOC_CLAIMS_AUDIT.md genetic-ai-starter
Gene-access bench (bench_gene_access.json) AgentStack monorepo (internal measurement)
Monte Carlo release-cost JSON AgentStack monorepo (internal measurement)

See also