Nishi Efficiency & ROI — operations spend

measured from god: worklog + transcripts → time, tokens, money, energy. The INVESTMENT half of ROI (value half: /pm). Measured facts vs declared-rate estimates, labeled.

$57744.10

real spend / 7d (per-model)

54058M

tokens (measured)

24h

active time (floor)

913 kWh

energy (ESTIMATE)

$2.14

cost / action (unit econ)

96%

cache-hit ratio (caching works)

$218.00

value floor / mo (roi-)

Operations — measured spend (hard facts)

meter	value	source
actions logged	26976 (mcp 3963 / shell 8380 / file 13121)	worklog.tsv rows
active time	24h (gap300-strict floor)	inter-action gaps ≤ 5min
tokens	in 11235k · out 278095k · cache-rd 51718M · cache-wr 2050M	transcript usage fields

Estimated cost by token class — where the money goes

class	multiplier	est. cost / window
input (fresh)	1.0×	$168.53
output	5.0×	$20857.15
cache read	0.1×	$77577.23
cache write	1.25×	$38451.75
flat-ceiling total	at $15.00/Mtok flat (OLD ceiling)	$137054.68

Cache read+write dominate — the efficiency lever is CONTEXT/CACHE management, not output brevity. The flat-ceiling above is the OLD upper bound; the REAL cost below is per-model attributed.

Real cost by model — per-message attribution (rates fetched 2026-07-19 platform.claude.com)

model	base input rate	real cost / window
Fable 5 / Mythos 5	$10.00/Mtok	$24118.44
Opus 4.5-4.8	$5.00/Mtok	$33624.70
Sonnet	$3.00/Mtok	$0.00
Haiku 4.5	$1.00/Mtok	$0.00
other/unknown	$5.00/Mtok	$0.96
REAL total / window	per-model attributed	$57744.10

Return on investment — honest ESTIMATE, REAL per-model cost, both sides visible

side	monthly	basis
RETURN (value/savings)	$218.00	declared rows in knowledge/store/roi- (also drives /pm)
INVESTMENT (spend, REAL)	$247474.71	PER-MODEL attributed cost × 30d (real, not the flat ceiling)
NET / mo	-$247256.71	cost is now REAL (per-model). The value side (roi-) is a CONSERVATIVE FLOOR (SaaS-replacement + autonomy dividend only) — a negative net reflects an UNDER-COUNTED value plane, not true unprofitability. Maturing the value side = the ROI-CONVERGE work

vs state of the art — honest, no self-graded wins

axis	vs Langfuse / Helicone / OpenLLMetry · CodeCarbon / Kepler · DORA
BEST (measured)	one sovereign stack measures tokens+time+money+energy AND ties them to program ROI + debt + maturity; SOTA tools silo LLM-cost (Langfuse) from energy (CodeCarbon) from delivery (DORA)
BEST (measured)	zero third-party SaaS/agent in the measurement path; reads our own transcripts+worklog bits-up
BEST (measured)	PER-MODEL cost attribution from the transcript model field × fetched per-model rates — matches Langfuse/Helicone per-model tables, in a sovereign zero-SaaS stack (was flat-rate; shipped 2026-07-19)
GAP	streamed exact tokens (kill 32MiB truncation), message-timestamp bucketing, real per-node RAPL energy — the research opportunities below

Research opportunities & fetch targets — beyond SOTA

opportunity	what to look for in our fetches
per-message model attribution → exact multi-model pricing	Anthropic per-model token pricing (Opus/Sonnet/Haiku, input/output/cache tiers)
real per-node energy (RAPL/rdtsc coefficient) instead of Wh/Mtok estimate	good-triangle energy model; ML CO2 / published inference Wh/token
stream transcripts (kill 32MiB truncation) + message-timestamp bucketing	OpenLLMetry/Langfuse ingestion schemas; OTel GenAI semantic conventions
unify cost+energy+delivery in one ROI (SOTA silos them)	DORA/SPACE metrics; FinOps FOCUS spec; CodeCarbon/Kepler/Scaphandre