nx_costest_lib.nx

nx_costest_lib.nx -- THE ONE COST ESTIMATOR (2026-09-16): three-point estimates (P10, P50, P90) in the boards' own unit, their aggregation over components, the CALIBRATION of that unit against RECORDED ACTUALS, and the compute-cost rules that turn a model size into hours at a measured rate. WHY. Operator 2026-09-16: "if we need 3rd party today then id like to know via the /compare capabilities what the lift would be to replace a 3rd party we are using thats open source etc with our own in cost", and "if the capability to estimate cost gets us to sota for our nishi analyst and other analysis systems like flipping cars or investing then im fine investing". The estate already prices a car flip by its P10 downside (nx_carflip_lib cf_score, nx_flip_score) and every board declares a cost per rung in u ("1 u = one measured session-leg", the plan's own unit row) -- but nothing CALIBRATED u against what the boards themselves recorded, and nothing aggregated a set of estimates honestly. Ranker rung PR7 (rk_cost_recal, "actuals write back") named this gap on 2026-08-20 and it stayed _ABSENT_. This lib is that primitive, shared by the foreign-asset ledger (nx_assets_lib), the ranker's cost model, the car-flip and investing lanes and the analyst. THE UNIT. Estimates are held as DECI-U (1.5 u is the integer 15), so the plan's "1.5" and "8.5" compare as integers; hours are held as CENTI-HOURS (12.34 h is 1234). A calibration rate is centi-hours PER U. AGGREGATION (ces_agg3). Components estimated independently combine as: P50 = sum of medians; the lower and upper half-widths (P50-P10, P90-P50) combine by root-sum-square, which is the three-point (PERT-class) rule for independent uncertainties. The straight sums of P10 and P90 are returned BESIDE the RSS figures as the fully-correlated bound, because a reader who sees only the RSS band will read it as the worst case and it is not: when every component slips together, the correlated sum is the honest ceiling. CALIBRATION (ces_census_plan, ces_cal_*). The boards journal `log|<epoch>|<rung>|<kind>|<text>` rows and every rung declares its cost in u. For a rung with at least one `land` row, the ACTUAL is the wall-clock from its FIRST journal row of any kind to its LAST land row -- a ceiling on effort, since it includes waiting -- and the ratio actual-hours per declared-u is one calibration sample. The quantiles of that sample (P10, P50, P90) are the calibration; a rung whose land row is its first row (zero elapsed) or whose cost is absent is EXCLUDED BY NAME and counted, never silently dropped, and the partition WITHLAND = CALIBRATED + EXCL_ZERO + EXCL_NOCOST is printed so a reader can check it sums. The ledger line (ces_cal_line) is written and parsed by ONE grammar in this file, so the writer (nx_costcal) and every reader (the compare emitter) cannot drift apart. QUANTILE RANK. index = permil * (n - 1) / 1000 (nearest rank, the same convention nx_bench_stats uses for its p95). ces_cal_min_n derives the smallest n at which P10, P50 and P90 index THREE DISTINCT RANKS -- below it a "P10" is the same sample as the median and the band is decoration. That bound is computed, not typed. COMPUTE RULES. Training costs about 6 * params * tokens FLOPs (forward and backward) and inference about 2 * params * tokens (the standard 6ND / 2ND rules of the scaling-law literature); ces_seconds divides by a MEASURED rate the caller passes in. The rate is data: the llm board measures the CPU engine at 819 ms per token for a 7B model, and a GPU rate is a bar until the estate measures one.