1  #+title: Pi-Tokenomics
 2  #+author: Xavier Brinon
 3  #+date: [2026-04-19]
 4  #+startup: indent
 5  #+options: toc:2 num:nil ^:{}
 6  
 7  * What This Is
 8  
 9  A generalized framework for reasoning about the /token economics/ of
10  AI workflows — the tradeoffs between /intelligence per token/, /tokens
11  per dollar/, and /tokens per second/ as they play out across the
12  stages of any production AI pipeline.
13  
14  The framework lifts the "production function" idea from AISLE's
15  [[https://aisle.com/blog/ai-cybersecurity-after-mythos-the-jagged-frontier][AI Cybersecurity After Mythos — the Jagged Frontier]] out of its
16  cybersec-specific container and asks: /what survives?/
17  
18  Short answer: the three metrics, the Pareto-frontier methodology, the
19  "no single model fits" claim, and two architectural patterns —
20  /Cheap fan-out → expensive fan-in/ and /Expensive plan → cheap
21  execution/. What does /not/ survive: named pipeline stages, cybersec
22  intelligence floors, domain-specific evals.
23  
24  The framework replaces named stages with eight /stage-classifying
25  properties/ that determine which metric dominates for any stage of
26  any workflow. The result is prescriptive: for each stage, pick the
27  cheapest model on the Pareto frontier that clears the intelligence
28  floor /as implied by the property mix/.
29  
30  * Why It Exists
31  
32  Most production AI apps hardcode one model across every stage of a
33  pipeline. The framework argues this is doubly wasteful: frontier
34  prices on the bulk stages, and frontier-quality penalties on the
35  deep stages a single model cannot quite handle. The correct unit of
36  optimisation is the /stage/, not the /app/.
37  
38  This repo is where that argument is being built, stress-tested, and
39  (eventually) turned into something publishable and/or operational.
40  
41  * Contents
42  
43  | File                       | Role                                                                                     |
44  |----------------------------+------------------------------------------------------------------------------------------|
45  | [[file:token_economics.org][=token_economics.org=]]      | The framework. Metrics, properties, architectural patterns, four insights, open questions, directions for further exploration, downstream forks and research roadmap. |
46  | [[file:UBIQUITOUS_LANGUAGE.org][=UBIQUITOUS_LANGUAGE.org=]]  | DDD-style glossary. Canonical terms, aliases to avoid, flagged ambiguities, example dialogue. |
47  
48  * How to Read It
49  
50  Depends on intent.
51  
52  - First encounter, you want the thesis :: Read
53    [[file:token_economics.org][=token_economics.org=]] top to bottom. The =Origin= and =Three
54    Metrics= sections frame the argument; =Property Axes= is the core
55    move; =Architectural Patterns= and =Four Insights= are where the
56    framework becomes prescriptive.
57  - You are applying it to a real pipeline :: Start with the
58    [[file:UBIQUITOUS_LANGUAGE.org::*Example dialogue][example dialogue]] in the glossary (it shows the framework applied
59    to a TanStack loader audit), then jump to =Property Axes= and
60    =Architectural Patterns= in the main doc.
61  - You are editing or extending the framework :: Read the glossary
62    first — especially =Flagged ambiguities=. Several terms
63    (/orchestration/ vs /routing/, /accuracy/ vs /recall+precision/,
64    Cobb-Douglas vs Leontief) have open disagreements that new prose
65    should not re-introduce.
66  - You are picking up contributor work :: Start with =Downstream
67    Forks and Research Roadmap= in the main doc for the currently
68    prioritised work; fall back to =Directions for Further
69    Exploration= for the full idea well. Items are named, not
70    numbered, so they can be referred to by title.
71  
72  * Status
73  
74  Working document. The framework is coherent but not yet validated
75  against a real pipeline. /Open Questions/ in =token_economics.org=
76  flag the framework's most vulnerable spots — non-verifiable tasks,
77  Cobb-Douglas vs Leontief framing, agent-tool-level applicability,
78  maintainer trust as an external constraint.
79  
80  No code yet. A research roadmap has been documented in =Downstream
81  Forks and Research Roadmap= of the main doc: three planned downstream
82  projects — a power-bar extension, a Pi-CrewLens plugin, and a
83  nano-analyzer per OWASP × tech stack — and a seven-step research
84  sequence that unblocks them. Current focus: /Operational metric
85  definitions/ (item 1); work not yet started.
86  
87  * References
88  
89  - AISLE article :: [[https://aisle.com/blog/ai-cybersecurity-after-mythos-the-jagged-frontier][AI Cybersecurity After Mythos — the Jagged Frontier]]
90  - Companion AISLE piece :: /System Over Model: Zero-Day Discovery at
91    the Jagged Frontier/ (referenced throughout the glossary)