What KonomiStandard saves, in tokens
KonomiStandard is Thomas Frumkin's self-defining industrial standards compression format — its own goal line reads "COMPRESSION: max info, min tokens". This measures that on the format's own spec: the held-out layers (ISA-101, ISA-18.2, OPC-UA, MQTT/Sparkplug, Modbus, KPIs, crosswalks) as Thomas wrote them, against the same content as minified JSON — and, separately, what a shared dictionary adds when the receiver already holds the entries. Sealed before it was measured; reported whichever way it landed.
1 of 4 sealed rules held. With no shared dictionary, the held-out layers cost 5321 Claude tokens as KonomiStandard and 5756 as minified JSON — 1.08×. With a shared dictionary, naming the 47 held-out entries by address instead of sending them costs 855 tokens instead of 4905 — 5.74×, but only where the receiver already holds them.
| held-out layer | Claude tokens KonomiStandard | Claude tokens JSON | JSON ÷ KS | local qwen2.5 KS · JSON | characters KS · JSON |
|---|---|---|---|---|---|
| 🖥️ LAYER 4: ISA-101 (HMI Design) | 981 | 1095 | 1.12× | 616 · 626 | 1956 · 2134 |
| 🚨 LAYER 5: ISA-18.2 (Alarm Management) | 1066 | 1172 | 1.10× | 650 · 657 | 2139 · 2283 |
| 📡 LAYER 6: OPC-UA (Communication) | 658 | 711 | 1.08× | 376 · 382 | 1321 · 1375 |
| 📨 LAYER 7: MQTT/Sparkplug (Messaging) | 802 | 877 | 1.09× | 500 · 500 | 1414 · 1505 |
| 🔧 LAYER 8: Modbus (Field Protocol) | 655 | 758 | 1.16× | 445 · 451 | 1120 · 1295 |
| 📊 LAYER 9: KPIs (Performance Metrics) | 601 | 595 | 0.99× | 382 · 352 | 1259 · 1169 |
| 🔀 CROSSWALKS (δ maps) | 558 | 548 | 0.98× | 317 · 301 | 899 · 845 |
| all held-out layers | 5321 | 5756 | 1.08× | 3286 · 3269 | 10108 · 10606 |
The sealed rules
| rule | result | predicted | |
|---|---|---|---|
| the JSON twin of every held-out layer carries every word and number of the notation (coverage at least 0.999) | the JSON twin of every held-out layer carries every word and number of the notation | PASS | pass — the twin keeps every word on the development layers |
| with no shared dictionary, the held-out layers cost at least 1.2× more Claude tokens as JSON than as KonomiStandard | 1.08× — JSON 5756 tokens, KonomiStandard 5321 | FAIL | fail — the characters were within about 10% on the layers I built it on, and the ──── header rules cost tokens; I expect roughly 0.9× to 1.15× |
| and every held-out layer costs fewer Claude tokens as KonomiStandard | 2 of 7 layers cost as much or more as KonomiStandard | FAIL | fail — on some layers the JSON twin is already shorter in characters |
| with a shared dictionary, sending the addresses instead of the entries saves at least 10× in Claude tokens | 5.74× — 4905 tokens of entries against 855 tokens of addresses, when the receiver already holds them | FAIL | pass — an address is a handful of tokens and an entry is tens to hundreds; I expect 10× to 25× |
Claude tokens by the official CLI (claude-sonnet-5-5) on the estate's own login, each text measured against a one-character baseline; local tokens by qwen2.5:7b in Ollama. Sealed in d1e67b3; every count is in measure/data/run.json; CI rebuilds every JSON twin from Thomas's pinned file and re-grades the rules on every push.