Ai on Herman

Agent Parliament

Wed, 22 Jul 2026 09:00:00 -0400

Overview

Agent Parliament is a real multi-agent chamber that debates, votes, and keeps public minutes — not a toy chat between personas, but a governed deliberation stack with seats, parties, standing law, and a human Crown who retains veto. The public snapshot published 2026-07-20 has 253 motions (10 open) under a ratified constitution v0.1.7, and Parliament OS has advanced through P46. The first reproducible evidence that it is materially improving my final deliverables has landed.

Skynet Breakage Atlas

Thu, 16 Jul 2026 00:00:00 +0000

Overview

The Skynet Breakage Atlas answers a practical routing question: when traffic goes through our proxy fleet instead of first-party subscription surfaces, which models go empty, hard-error, or slow-walk under adversarial families?

Live at breakage.hermanity.dev.

What’s measured

Controls: Claude Max, Codex, Grok (subscription / direct)
Fleet: Skynet-routed models (deepseek, glm, kimi, minimax, …)
Families: cascade seed, context boundary, cost threshold, empty probe, format JSON, multi-intent
Metrics: flag rate, empty rate, hard error rate, p50 latency

Snapshot (public table on the site)

glm-5.2 is the known-broken reference (flag rate 1.0 across cells in the frozen N-batch). Controls stay low-flag; fleet variance is the point of the atlas — empty replies are a first-class failure mode, not a footnote.

Model Bench

Wed, 15 Jul 2026 00:00:00 +0000

Overview

A benchmarking site I built to answer one question well: which subscription-based LLM is worth its slot on my dashboard, and which pay-as-you-go API is the better substitute. Every public benchmark either (a) scores only pay-as-you-go APIs (so the results don’t tell me what a Claude Max / ChatGPT Plus / ollama-routed model would actually do on my workloads) or (b) makes vague quality claims with no cost column (so I can’t compare a $20/month subscription to a $5 in API spend). This benchmark measures both, on the same tasks, in the same week, scored by a judge from a different model family.

AISpend

Wed, 01 Jul 2026 00:00:00 +0000

Overview

AISpend is the personal-finance tool I built for the AI-power-user who holds 3-7 active subscriptions (Cursor + Claude Code + ChatGPT Plus + z.ai + Kilo + an OpenRouter pay-as-you-go) and wants one view of “what is AI costing me this month, where is it going, and which subscriptions should I cut?”

Helicone, Portkey, Langfuse, HoneyHive, and LangSmith all serve app developers building LLM products — proxy-based, code-integrated. None target the personal/team subscription buyer. AISpend fills that gap.

SwarmUI Image Generation

Wed, 01 Jul 2026 00:00:00 +0000

Overview

A calibration and generation pipeline for local Stable Diffusion image generation. Talks to SwarmUI on a Windows rig with an NVIDIA 3090TI via HTTP API.

Calibration methodology

12 rounds of systematic A/B testing across 12 models:

Step sweeps — finding the sweet spot (Flux: 36 steps is perfect; Chroma: flat plateau 16-40)
CFG sweeps — Chroma sweet spot is 2.5-3.5; outside that drops to 3.0/5 quality
Cross-model comparisons — SDXL wins action/rain scenes at 5.0/5; Flux nails anime at 5.0/5
Single-sample trap — Round 11 found a “winner” on one prompt; Round 12 with 3 new prompts disproved it. Rule: n≥3 prompts per cell to change a default.

Curated defaults

Model	Steps	CFG	Sampler	Notes
flux1-dev-fp8	36	—	euler/simple	5.00/5 winner
chroma_v10HD	24	3.0	dpmpp_2m/beta	Cap at 40 steps
chroma1-base	24	3.0	dpmpp_2m/beta	Best for portraits
sdxl_base	36	7.0	dpmpp_2m/karras	Flat plateau 36-60

Hard-won lesson

Never trust a single-image calibration. A 1.0-point quality advantage on n=1 is noise. This methodology error almost became a permanent default change before a wider sweep caught it.