Skip to content

SunrisesIllNeverSee/sigrank-hf-hackathon

Repository files navigation

title MO§ES SigRank
emoji 📡
colorFrom yellow
colorTo gray
sdk gradio
sdk_version 6.17.3
app_file app.py
hf_oauth true
pinned false
license mit
short_description Rank AI operators by architecture. Volume can't buy rank.
models
openbmb/MiniCPM4-0.5B
tags
thousand-token-wood
off-brand
tiny-titan
best-demo
minicpm
track:wood
sponsor:openbmb
sponsor:openai
achievement:offgrid
achievement:offbrand
achievement:sharing
achievement:fieldnotes

MO§ES™ SigRank — the diagnostic x-ray of the token economy

HF Space Model License Zenodo Patent Pending Benchmarks Track

A leaderboard that judges AI coding operators by architecture, not budget. Paste your token usage; get an operator profile with a tiny-model narration and your rank. The ranking metric Υ = (Cache·Output)/Input² penalizes raw-input padding quadratically — volume can't buy rank — but Υ is only the headline of a larger metric system whose mathematical thesis is the cascade decomposition.

GitHub: github.com/SunrisesIllNeverSee · Repo: moses-sigrank · Deck: mos2es.com/deck · Benchmarks: mos2es.com/benchmarks · Law: mos2es.com


SigRank is Layer 2 of the MO§ES™ stack — the human–AI operator intelligence layer. MO§ES™ is the substrate: compression, recursive execution, drift control, governed meaning at source. Built by Deric J. McHenry · Ello Cello LLC.

IP: 4 patent filings · IC 042 TM 99408355 · Conservation Law of Commitment — Zenodo DOI 10.5281/zenodo.18792459


Origin & Theory

SigRank is grounded in a published law: the Conservation Law of Commitment.

C(T(S)) ≈ C(S) — Commitment is conserved when enforcement holds. If the meaning of a signal is preserved across transformations, the architecture is sound. If it drifts, the architecture is leaking. Formal proof — Zenodo DOI 10.5281/zenodo.18792459 →

When token-level data became available via ccusage, the word-based commitment framework translated directly into token-domain measurement:

  • commitment → cache_create / output
  • reuse → cache_read / cache_create
  • transmission → output / input

The cascade identity (O/I) × (C_w/O) × (C_r/C_w) = C_r/I is the token-domain expression of the same conservation law. Operators who conserve commitment compound. Operators who don't, pay for it.

What it does

Paste ccusage claude --json (Claude Code), ccusage codex --json (Codex), or four numbers →

  • operator profile — a 0.5B MiniCPM model narrates your architecture, plus raw ledger, composition, full metrics, cascade breakdown
  • leaderboard placement vs real operators, ranked by Υ, with blended $/1M cost
  • trading card — species classification, cascade decomposition, composition bar

How to measure yourself

SigRank is local-first — the importer reads your usage on your own machine.

1 — Primary: the local importer. Clone it once, then run it (no install step — runs on system Python + Node.js):

git clone https://github.com/SunrisesIllNeverSee/moses-sigrank
cd moses-sigrank
./sigrank

./sigrank --codex for Codex · ./sigrank --all for both providers in one pass. It runs ccusage for you, computes your profile + Υ, and prints your board rank. Nothing leaves your machine.

2 — Backup: paste on the Space. No repo? Run ccusage yourself and paste the JSON into the Clock Your Signal box. Run one command per provider:

npx ccusage@latest claude --json
npx ccusage@latest codex --json

(ccusage installed globally? drop the prefix: ccusage claude --json.) Or type four numbers: input output cache_create cache_read.

⚠️ Don't use bare ccusage --json (no subcommand): it merges every agent into one total, which inflates input and distorts the architecture read.

3 — Saving (optional). Sign in with HuggingFace on the Space to earn one persistent board entry + session history (Greatest Hits). Without login, your read is a live snapshot only.

Codex note. Codex doesn't report a fresh-vs-cache input split, so its input is estimated: alone → AA-backed 2:1 baseline; with a Claude profile → your own Claude input:output ratio. Estimated rows flagged with *.

The model (Tiny Titan / Best MiniCPM)

openbmb/MiniCPM4-0.5B (0.5B params, well under the 4B cap) runs on ZeroGPU and narrates the operator read. Non-blocking: if unavailable, a deterministic template is used and the app still works. Everything quantitative is pure computation.

The full metric system

Υ is the ranking metric. The mathematical thesis is the cascade decomposition of leverage into three behavioral stages — Υ is what you sort by, the cascade is what explains why an operator lands where they do.

The cascade as a diagnostic

(O/I)  ×  (C_w/O)   ×  (C_r/C_w)   =   C_r/I
 │          │            │               │
transmission commitment  compounding   leverage
 generate    commit to   reuse what
 output      cache        was cached

10x DEV is the log₁₀ of that product — the amplification exponent. By telescoping, 10^(10x DEV) = leverage = C_r/I.

Operator species

species signature
Cascade Matrix velocity ≥ 1 and leverage ≥ 100 · recursive processing loop
Cache Architect leverage ≥ 10, velocity < 1 · persistent context layer
Converter Loop velocity ≥ 0.5, leverage < 2 · single-pass processing velocity
Throughput Pipe low on both axes · raw metric bandwidth
Non-Compounding no cache_create — cascade can't form

Species / quadrants

              high amplification
                     │
   CASCADE (stacks)  │   (— rare / empty —)
   low scale ◄───────┼───────► high scale
   CONVERTER (I→O)   │   THROUGHPUT (volume)
   CACHE ARCHITECT   │   (reuse)
                     │
              low amplification

MO§ES occupies the empty quadrant — low scale, high amplification. The claim is not "top of a ladder"; it's that this region of the token economy is structurally empty, and the geometry of (C·O)/I² is what makes it empty.

Full metric table

metric formula meaning
SNR O/(I+O) output share
10x DEV log₁₀(cascade) amplification exponent
Operating Ratio C:I:O, input=1 footprint vs Artificial Analysis 7:2:1
Velocity O/I output per input token
Leverage C/I cache reads per human token
Efficiency (C+O)/I ÷ 4.0 vs AA baseline
Avg $/1M blended cost ÷ total efficient architecture is also cheapest
Υ (Yield) (C·O)/I² un-gameable ranking metric

Benchmark convergence — two independent instruments

Instrument 1 — Artificial Analysis model benchmarks

AA benchmarks measure models at a 7:2:1 (cache : input : output) token mix. SigRank measures operators from their own token logs. Both converge: cache-dominant architecture is the most efficient AND cheapest per token. The AA 7:2:1 ratio is the source of the 4.0 efficiency baseline: (7+1)/2 = 4.0.

Instrument 2 — Five measured kernels (mos2es.com/benchmarks)

MO§ES operator measured against the AA Coding Agent Index field average — 7-day window, raw JSONL extraction across 98 session files, all subagents. #1 in all five categories:

kernel MO§ES™ field avg delta
I · Cache hit rate 94.66% 90.68% +3.98pp · #1
II · Output : Input ratio 17.9× 0.162× 110× field avg · #1
III · Tokens per task 810K 4.67M 5.8× fewer · #1
IV · Time per task 1.84 min 11.92 min 6.5× faster · #1
V · Cost per LOC $0.0007 $0.067 96× cheaper · #1

Raw token ledger (98 sessions · 1,123B total):

Input Output Cache Create Cache Read Total Sessions LOC
MO§ES™ 123K 3.90M 34.83M 1.084B 1.123B 98 / 1,465 tasks 35,242
Field avg (per model) 162.9M 17.2M 1.49B 1.67B 1 / 358 tasks 7,160

Field data: artificialanalysis.ai/agents/coding-agents · extracted 2026-05-14. MO§ES: sustained product build, 7-day window (2026-05-08 → 2026-05-14). The convergence of #1 leadership across both isolated benchmark runs and sustained product builds is the structural result.

This is the real-world grounding behind SigRank's thesis — the same operator that ranks #1 on Υ also ranks #1 on every external benchmark kernel.

IP & legal record

SigRank is Phase 2 of a fully filed, commercially structured IP portfolio:

filing serial date protected layer
PPA 01 · MOS²ES 63/877,177 Sep 7, 2025 · Conf. 7067 Constitutional governance, signal encoding, compression substrate
PPA 02 · SCS Engine 63/883,018 Sep 17, 2025 · Conf. 6401 Signal Compression System Engine, lineage, sovereign compression
Utility 03 · CIVITAS 19/426,028 Dec 18, 2025 · Conf. 2165 Frontend civic infrastructure: SIGRANK, SigEconomy, Agent City-State
PPA 04 · Commitment Conservation 63/991,282 Feb 26, 2026 · Conf. 6108 Semantic commitment conservation under recursive transformation
MO§ES™ Trademark 99408355 · IC 042 Sep 23, 2025 Software / governance protocol identity

Published law: Conservation Law of Commitment · Zenodo DOI 10.5281/zenodo.18792459

Full legal ledger: mos2es.com/legal

The MO§ES™ stack

SigRank sits at Layer 2 of a four-layer constitutional architecture:

layer product purpose
04 SIGNOMY Governed agent marketplace — execution-layer governance, participatory trust, agent provenance
03 SIGRANK Human–AI operator leaderboard — sync telemetry, resonance metrics ← this app
02 AQUA Application workflow — answer banks, reusable submission memory
01 MO§ES™ Substrate — compression, recursive execution, drift control, lineage

MO§ES™ IS THE ENGINE · AQUA + SIGRANK ARE THE WEDGES · SIGNOMY IS THE GOVERNED ECONOMY

Codex support

Codex never itemizes cache writes, so SigRank estimates the high-signal user input from output via two pathways (_codex_input_estimate in ingest.py):

  • Alpha — Codex alone: AA-backed 2:1 baseline — est_input = 2 × output
  • Beta — Codex + Claude profile: operator's own measured Claude input:output ratio — est_input = output × (claude_input / claude_output). CLI builds this automatically (./sigrank --codex reads Claude first).

Every Codex-derived row is flagged with * and names the exact pathway used.

Cost

For Claude Code, ccusage supplies real cost → exact $/1M. For manual/wild rows, $/1M is a list-price estimate (shown with ~). Either way: the cache-dominant operator at the top is also the cheapest per token, by an order of magnitude.

Demo video

Watch the demo on Loom →

Social post

View on X →

Notes

  • Persistence boundary. Pasted rows are scored live but not added to the persisted board — corpus is curated (owner-seeded). Backed by Supabase with hardcoded-seed fallback.
  • Cost provenance. Board's $/1M is a list-price recompute for every corpus row (~) so all rows compare apples-to-apples. MO§ES's verified ccusage cost is $0.527, reproduced exactly by the recompute.
  • Wild corpus (10 operators). Public ccusage footprints from tokscale.ai. MO§ES is verified ccusage data, not tokscale.
  • Υ is an engineered macro-efficiency index motivated by thermodynamics (Landauer, Ohmic dissipation); log Υ = X + log(Velocity). An analogy, not a microscopic-entropy derivation.

Built for the HF/Gradio Build Small Hackathon · Thousand Token Wood 🍄

Built by Deric J. McHenry · github.com/SunrisesIllNeverSee · Ello Cello LLC · Codex + Claude + Devin

Patent pending 19/426,028 · IC 042 TM 99408355 · DOI 10.5281/zenodo.18792459

About

SigRank — HuggingFace hackathon prototype (Thousand Token Wood track, MiniCPM4-0.5B). The concept proof that became signalaf.com. Live product is now github.com/SunrisesIllNeverSee/sigrank-app.

Topics

Resources

License

Stars

0 stars

Watchers

0 watching

Forks

Releases

No releases published

Packages

 
 
 

Contributors