He Was Socrates · Monthly Operating Cost Ledger

P04 The Cost-Conscious wireframe · 7 components · self-host · static-first · zero recurring SaaS

run r-20260507-010321 · spec lock e5dfadf2c8…314c5 · NO-CLOUD invariant #1

RECURRING TOTAL · $0.00 / month

Component Bill of Materials

$0 · local

01 · Surface

Bust UI (SwiftUI)

apps/macos/HeWasSocrates · 1-bit halftone PNG swap, 16 visemes, 30 fps tick. No paid design SaaS, no Figma seat at runtime.

src: Sources/HeWasSocratesApp.swift · ContentView.swift

$0 · one-time

02 · Model weights

Gemma 4 E4B 4-bit MLX (3.97 GB)

Pulled once via OS-mediated MLX cache from HuggingFace mirror (mlx-community/gemma-4-e4b-it-4bit). Zero monthly hosting fee; weights live on user's disk.

src: LLMRegistry.gemma4_e4b_it_4bit · sanctioned egress per CLAUDE.md inv#1

$0 · user GPU

03 · Inference

MLX runtime

mlx-swift-lm on Apple Silicon. TTFT 192 ms median (n=10, M1 Max, PR-Λ verify-2). User pays the electricity, not the API.

src: claudedocs/bench/2026-05-06-latency-bench.json

$0 · Apple SDK

04 · STT

Speech.framework

SFSpeechRecognizer with requiresOnDeviceRecognition = true. macOS 26 SpeechAnalyzer + AssetInventory (PR #33). No Whisper API metering.

src: AudioInputManager.swift · Preflight.swift

$0 · Apple SDK

05 · TTS

AVSpeechSynthesizer

Yuna (ko) / Samantha (en), premium → enhanced → default chain. No ElevenLabs subscription, no per-character billing.

src: TTSManager.swift

$0 · local DB

06 · Persistence

Wondering Log (Core Data on SQLite)

SHA-256 fingerprint dedup, deterministic JSON export. No managed Postgres, no Pinecone vector DB seat, no Supabase project. The disk is the database.

src: WonderingLog.swift · Core Data store

$0 · GitHub free

07 · Distribution

GitHub Pages + Releases

Two-Weeks-Team org, public repo. Gallery + hero preview at /docs. Zero CDN bill, zero Vercel Pro seat, zero domain renewal if .github.io is used.

src: idea.json §3 General · public Pages URL

Monthly Bill · He Was Socrates (self-host)

Line item

Unit

Monthly

Bust UI hosting (compiled into .app)

—

$0.00

Gemma 4 weights storage (user's disk)

3.97 GB

$0.00

Inference compute (user's GPU)

~192 ms TTFT

$0.00

STT minutes (Apple Speech.framework)

unmetered

$0.00

TTS characters (AVSpeech, Yuna ko)

unmetered

$0.00

Wondering log storage (Core Data)

local

$0.00

Static gallery hosting (GitHub Pages)

100 GB/mo soft

$0.00

Recurring monthly total

—

$0.00

Scales with users at the same rate it scales with zero users: $0 · ∂cost/∂user = 0.

Cloud-Equivalent Counterfactual · 5,000-seat school district · 9-month year

He Was Socrates (this stack)

$0 / mo

Recurring SaaS spend: zero. One-time per device: 3.97 GB download, ~10 min on home broadband. Per-student marginal cost: $0.00.

Cloud-equivalent build

~$5,403 / mo

Estimated for 5,000 students × ~30 turns/week × 4 weeks. See itemization below.

Cloud line item

Unit price

Monthly

GPT-4o inference (Socratic dialogue, ~600M tok/mo)

$0.0050 / 1K in

~$3,000.00

ElevenLabs TTS (Korean voice, Creator tier × 9 seats)

$22 / seat

~$198.00

Whisper API STT (audio minutes)

$0.006 / min

~$1,800.00

Pinecone vector DB (wondering log embeddings, Standard)

$70 base + usage

~$285.00

Vercel Pro (preview + admin)

$20 / seat × 6

~$120.00

Cloud-equivalent recurring total

—

~$5,403.00

∂cost/∂user ≈ $1.08 / student / month · linear with usage · indefinite recurrence

Delta · why this stack wins on the ledger

−$5,403

Δ Monthly

Recurring savings

Cloud build − this build = $5,403 − $0 = $5,403 / mo avoided. Over a 9-month school year: ~$48,627 per district per year not spent on SaaS.

$0 vendor lock-in

Δ Risk

Zero vendor exposure

No API keys to rotate. No price-hike letters. No deprecation notices. The bill cannot go up because there is no bill.

$0 egress

Δ Privacy bill

No data-egress fees, no DPA

NO-CLOUD invariant means student utterances never leave the Mac, so the procurement officer also avoids the unbudgeted line item: legal review of a third-party data-processing agreement.

Sources · CLAUDE.md absolute invariant #1 (no network.client / network.server) · HeWasSocrates.entitlements · claudedocs/bench/2026-05-06-latency-bench.json (PR-Λ TTFT 192 ms n=10) · function_call_contract.yaml · idea.json §3 General (GitHub Pages publication target). Cloud unit prices are public list rates as of 2026-05-07; treat the ~$5,403 figure as a representative counterfactual, not a binding quote. All metrics on this page resolve to a commit, bench file, or vendor pricing page — zero unsourced numbers per spec success criterion #2.