Source: c62 visibility-stream — pairs iteration: amplified metrics, pushed metaphors, 4 solid + 4 out-there
Same real data, new pairings. No new endpoints — the existing APIs were already holding unamplified metrics: session duration, cache-read context depth (the … figure, never yet charted), model mix, and per-commit adds vs deletes. Meaning-layer callouts only where earned (plots 1 and 8).
Metabolism
x = lines added (log) · y = lines deleted (log) · one dot = one commit. The diagonal is pure churn; below it the codebase grows, above it the codebase prunes. Healthy projects live on both sides — deletion is progress too, and most dashboards hide it.
1 commits with line changes · the ✓ callout is digest-derived (evidence-chip join) — the breaking cutover deleted more than it added, which is exactly what a schema unification should do
Call & response
One shared time axis, two stacked strips: AI output tokens per day above the spine, lines changed per day below. The pair reads as a duet — the call (tokens) and the response (code), same days, separate scales, each labeled.
days where the call has no response (tokens, no code) are research days; responses with no call are hand-work — both patterns are honest and visible
Context depth
x = cache-read tokens per session (log) · y = output tokens (log) · one dot = one session. The never-charted metric: how much context a session held vs what it produced. The question a governance buyer should ask: is deep context leverage, or is it sprawl?
0 sessions · the up-right drift says deep-context sessions produce more — context is leverage here, and that's measurable, not vibes
The shift log
x = hour of day · y = calendar day · one bar = one session's actual span (start → end) · color = model. Duration becomes visible — marathon sessions read as long bars; the model mix reads as color down the page.
0 session spans (UTC) · grey rows are weekends · hover for date, model, duration, tokens · sessions crossing midnight clamp at 24:00
The iceberg
One shared log scale, mirrored at a waterline: output tokens above, cache-read context below. The transparency metaphor made literal — …× more context is re-read than output is produced. What you see of AI work is the tip; visibility tooling exists because the mass is under the water.
per-day cache attributed to each session's start day (disclosed approximation) · both bergs share one log scale, so vertical distance is honestly comparable
The breath
x = day · y = net lines (adds − deletes), diverging around zero. Growth inhales above the line; pruning exhales below. The organism metaphor from the arcade thread, now carrying a real diverging metric — releases are marked as breath marks.
the deep exhale on Aug 6 is the schema unification pruning two command-specific pipelines · releases cluster after inhales — build, then ship
The orchestra
x = day · y = output tokens, stacked by model. The model portfolio as a score: who's playing, when, how loud. This is the "evaluated, discussed, efficiently used" claim as a picture — model choices are visible decisions, not defaults.
session tokens attributed to the session's dominant model and start day (disclosed approximation) · the fable-5 entrance in August is visible as a keychange
The exchange rate
x = day · y = tokens per line changed (log) — plot 2 of the last page was the relationship; this is its price history. A falling rate means a token buys more shipped code. Ticker framing on purpose: this is the number a CFO will eventually ask for.
grey dots = daily rate · violet line = 7-day rolling median · lower is better · the ✓ marks the cheapest shipped day in the window
The pairs catalog
Pairings not yet drawn, ranked by how much a partner or buyer would care — the backlog for iteration III:
the through-line of this iteration: the most persuasive metrics were already in the data, unlooked-at — context depth, deletions, duration, model mix. Visibility isn't collecting more; it's charting what you already have. Which is the pitch.