nx_eff_conv.nx source

1// nx_eff_conv.nx -- pure spend->money/energy conversion core (eff lane). SYSCALL-FREE integer math, 2// so it compiles anywhere (NAS buildroot incl older toolchains) -- decoupled from the getdents/fstatat 3// SCAN code in nx_eff_lib so the board (a plane READER) never drags scan syscalls into its compile. 4// Shared by nx_eff_lib (scan CLI) AND nx_eff_board_lib (ROI board) = DRY (rule 15). 5// Coefficients DECLARED + emitted transparently by every consumer (rule 11, operator "non-fake numbers"). 6// license_tier: ORIGINAL No hw writes (Rule 26). 7 8// weights x100 (basis points) = per-token cost relative to uncached input, Anthropic pricing shape. 9const EFL_W_IN: i64 = 100 10const EFL_W_OUT: i64 = 500 11const EFL_W_CR: i64 = 10 12const EFL_W_CW: i64 = 125 13const EFL_CENTS_PER_MTOK_IN: i64 = 1500 // uncached input $15.00/Mtok (Opus-class UPPER-BOUND default; editable) 14const EFL_MONEY_DIV: i64 = 100000000 // bp(x100) x per-Mtok(1e6) normalizer 15// ENERGY: compute-weighted (energy tracks FLOPs). We reuse the SAME class multipliers as cost (output 5x, 16// cache-read 0.1x, cache-write 1.25x vs input 1x) because both scale with compute -> cache reads (KV lookups) 17// are correctly cheap, not full-energy (the old flat all-tokens rate OVER-counted cache energy ~24x). 18// CALIBRATED to fetched research (2026-07-19): ~5e-4 Wh per OUTPUT token (Joule 2026 S2542-4351(26)00114-5; 19// TokenPowerBench arXiv 2512.03024; measured GPT-4o 3e-4, LLaMA-3.1-405B 8.6e-4 Wh/tok). At weight 500 (output), 20// energy_wh = weighted_bp / EFL_WH_ENERGY_DIV yields exactly 5e-4 Wh/output-tok. ESTIMATE, conservative-mid: 21// reasoning-heavy models (Opus-class) trend HIGHER (research: up to ~100x for long reasoning). 22const EFL_WH_ENERGY_DIV: i64 = 1000000 23 24func efl_weighted_bp(ti: i64, to_: i64, cr: i64, cw: i64) -> i64 { return ti*EFL_W_IN + to_*EFL_W_OUT + cr*EFL_W_CR + cw*EFL_W_CW } 25// money in cents = weighted-bp x cents/Mtok / normalizer (integer; no float ever) 26func efl_cost_cents(ti: i64, to_: i64, cr: i64, cw: i64) -> i64 { 27 let bp: i64 = efl_weighted_bp(ti, to_, cr, cw) 28 return bp * EFL_CENTS_PER_MTOK_IN / EFL_MONEY_DIV 29} 30// energy Wh estimate: compute-weighted (cache reads cheap, output expensive), calibrated to research ~5e-4 Wh/output-tok 31func efl_energy_wh(ti: i64, to_: i64, cr: i64, cw: i64) -> i64 { 32 let bp: i64 = efl_weighted_bp(ti, to_, cr, cw) 33 return bp / EFL_WH_ENERGY_DIV 34} 35 36// ---- PER-MODEL ATTRIBUTION (real cost, not the flat ceiling) ---- 37// Pricing FETCHED 2026-07-19 from platform.claude.com/docs/en/docs/about-claude/pricing (DECLARED, editable). 38// The output/cache multipliers are UNIVERSAL (output 5x, cache-write 1.25x, cache-read 0.1x of base input), 39// so per-model cost = weighted-bp(that model's tokens) x that MODEL's base-input rate. Only the input rate differs. 40const EFL_T_FABLE: i64 = 0 41const EFL_T_OPUS: i64 = 1 42const EFL_T_SONNET: i64 = 2 43const EFL_T_HAIKU: i64 = 3 44const EFL_T_OTHER: i64 = 4 45const EFL_NTIER: i64 = 5 46const EFL_RATE_FABLE: i64 = 1000 // Fable 5 / Mythos 5 base input $10.00/Mtok 47const EFL_RATE_OPUS: i64 = 500 // Opus 4.5-4.8 base input $5.00/Mtok 48const EFL_RATE_SONNET: i64 = 300 // Sonnet 4.5/4.6 $3 (Sonnet 5 intro $2 is lower -> conservative-high) 49const EFL_RATE_HAIKU: i64 = 100 // Haiku 4.5 base input $1.00/Mtok 50const EFL_RATE_OTHER: i64 = 500 // unknown model -> Opus-ish conservative 51func efl_tier_rate(tier: i64) -> i64 { 52 if tier == EFL_T_FABLE { return EFL_RATE_FABLE } 53 if tier == EFL_T_OPUS { return EFL_RATE_OPUS } 54 if tier == EFL_T_SONNET { return EFL_RATE_SONNET } 55 if tier == EFL_T_HAIKU { return EFL_RATE_HAIKU } 56 return EFL_RATE_OTHER 57} 58// cost cents for ONE tier's tokens at that tier's input rate (universal weighted-bp x per-model input rate) 59func efl_cost_tier(tier: i64, ti: i64, to_: i64, cr: i64, cw: i64) -> i64 { 60 let bp: i64 = efl_weighted_bp(ti, to_, cr, cw) 61 return bp * efl_tier_rate(tier) / EFL_MONEY_DIV 62}