Cross-Metric Analytics
Divergences & convergences across the 10 build archetypes · multi-metric combined charts · 1,586 HCM operators · 2026-08-10
Yield spread
4,694×
KINETIC vs INPUT-BOUND median Υ
Leverage spread
134×
KINETIC vs INPUT-BOUND median L
Velocity convergence
0.04–0.12
6 of 10 archetypes cluster here
Cache read dominance
88–96%
All archetypes are cache-read heavy
Key divergence
The field splits on leverage, not velocity. 6 of 10 archetypes have velocity between 0.04 and 0.12 — they all generate at similar rates. What separates them is how much they reuse: INPUT-BOUND reads 2.7× its input in cache; ARCHIVIST reads 30×; KINETIC reads 362×. Yield compounds this gap multiplicatively, producing a 4,694× spread from the same basic operating tempo.
1. Leverage vs Yield — the compounding scatter
Median leverage vs median yield per archetype (log-log)
Each dot is an archetype. Bubble size = operator count. Color = family. The relationship is near-linear on log-log — yield is leverage compounding on velocity. The outliers tell the story: KINETIC sits above the line (high velocity amplifies), INPUT-BOUND sits below (zero leverage kills yield even at low velocity).
2. Archetype radar — four workflow signatures
Normalized metric profiles: CONVERGENT vs KINETIC vs AMPLIFIER vs INPUT-BOUND
Six axes, each normalized to the field max. CONVERGENT and KINETIC are broad (everything elevated). AMPLIFIER is the "reuse specialist" — high leverage + construction but moderate velocity. INPUT-BOUND collapses to near-zero on every axis except input share — the "fresh input burner" signature.
3. Yield bars + Velocity line — where velocity diverges from yield
Median yield (bars, log scale) with median velocity overlaid (line)
Yield and velocity tell different stories. KINETIC has the highest yield AND the highest velocity — it's the "transmission monster." But RECURSIVE has 7.2 yield with only 0.19 velocity — it earns yield through leverage, not speed. And BUILDER has 1.1 yield with 0.077 velocity — average speed, below-average yield, because its construction ratio hasn't converted to reuse yet.
4. Construction vs Leverage — the build-to-reuse pipeline
Median construction ratio vs median leverage per archetype
Construction (cache_write/cache_read) measures how much new context you build per read. The Active Construction family (BUILDER, RECURSIVE, AMPLIFIER) sits in the upper-left-to-right diagonal — they build more AND reuse more. The Reuse Depth family (bottom-left cluster) barely builds. CONVERGENT is the only archetype that does both at extreme levels — the rare composition.
Key convergence
Cache-read percentage converges across the entire field: 88–96% of all tokens are cache reads for every archetype. The field is unanimously cache-read dominant. The divergence is in what feeds those reads — construction ratio. The Active Construction family writes 3–6% of their token budget as new cache. The Reuse Depth family writes 0.2–0.3%. Same read behavior, completely different write behavior. That's where the archetype signal lives.
5. Token composition (stacked) + SNR overlay
Four-pillar composition per archetype with SNR line
Bars show what fraction of each archetype's token budget goes to input, output, cache-write, cache-read. The SNR line shows signal-to-noise (output / total). The paradox: SNR is nearly flat across archetypes (0.02–0.60) despite massive yield differences. SNR doesn't separate the field — leverage does. The operators with the highest SNR (KINETIC, CONVERGENT) are also the ones with the smallest input share.
6. Metric heatmap — archetype z-scores across 6 dimensions
Standardized scores (z-score) per archetype per metric
Green = above field mean, red = below. This shows the full signature at a glance. CONVERGENT and KINETIC are green across the board. INPUT-BOUND is red everywhere except input share. The Reuse Depth family forms a gradient: PRIMING → CONTEXTUAL → DEEP READER → ARCHIVIST, each greener on leverage and yield as you go. BUILDER is the "average" archetype — near-zero on most axes.
7. Tokens/day vs Yield — the throughput paradox
Median tokens/day vs median yield per archetype (log-log)
If more tokens meant more yield, this would be a positive correlation. It's not. INPUT-BOUND burns 65M tokens/day for 0.1 yield. ARCHIVIST burns 20M for 3.5 yield. KINETIC burns 31M for 469 yield. Throughput doesn't predict yield — composition does. The operators moving the most tokens are often the least efficient at converting them to signal.
8. Family radar — the four workflow families
Convergence vs Generation family
CONVERGENT (all axes) vs KINETIC (velocity + leverage, low construction)
Reuse Depth vs Active Construction
ARCHIVIST (passive reuse) vs AMPLIFIER (active construction + reuse)
9. Population vs Yield — where the field concentrates
Operator count (bars) and median yield (line) per archetype, sorted by yield
The field's center of mass is in the low-yield archetypes. BUILDER (278 operators, yield 1.1) and CONTEXTUAL/ARCHIVIST (186 each, yield 0.8–3.5) hold the bulk. The high-yield archetypes are rare: KINETIC (113, yield 469) and CONVERGENT (104, yield 257) together are only 14% of the field but produce the top of the yield distribution. This is the power-law shape of operator skill.
The power law
The top 2 archetypes (KINETIC + CONVERGENT = 217 operators, 13.7% of the field) have a combined median yield of 363. The bottom 4 (INPUT-BOUND + PRIMING + CONTEXTUAL + BUILDER = 721 operators, 45.5% of the field) have a combined median yield of 0.6. That's a 605× gap between the top eighth and the bottom half. The field is not normally distributed — it's a power law, and the archetype system makes the structure visible.
Methodology
Charts: All charts use per-archetype medians from archetypes.json (10 deterministic build archetypes, 1,586 HCM operators). Radar charts normalize each metric to the field maximum (0–1 scale). The heatmap uses z-scores (standard deviations from the field mean). Scatter plots use log-log axes where indicated.
Metrics: Yield (Υ = cache_read × output / input²), Leverage (cache_read / input), Velocity (output / input), SNR (output / total), Construction (cache_write / cache_read), Input share (input / total tokens).
Population: human_center_of_mass (1,627 total, 1,586 HCM, outliers excluded). Input/total ratio analysis separates extreme humans from replay outliers.
Source: public/data/archetypes.json generated by scripts/gen-build-archetypes.mjs from public/data/field-analysis.json