Merge pull request #57 from cybertronai/docs/token-math-correction

0bserver07 · web-flow · commit fcb11ec991f3 · 2026-05-08T15:37:29.000-04:00
docs: add measured token math (closes #56)
diff --git a/BUILD_NOTES.md b/BUILD_NOTES.md
@@ -140,6 +140,26 @@ That's the entire DSL. Every stub had to fit.
 - **6 PRs merged via `gh pr merge`** in-session (the rest were merged separately by Yad)
 - **24 git pushes**
 
+### Token consumption — measured from JSONL session logs
+
+The harness display the lead session was showing during the build (something like `~k/1M (% used)`) is **the current context window utilisation, not cumulative tokens consumed**. It answers "how much room is left in the 1M-token window?", not "how much did the build cost?". The honest cost number requires aggregating the JSONL files for the lead + every subagent.
+
+Counted across the 63 JSONL session files in `~/.claude/projects/-Users-yadkonrad-dev-dev-year26-feb26-SutroYaro/` within the build window (2026-05-01T21:00 → 2026-05-04T23:30 UTC):
+
+| Bucket | Tokens | % of total |
+|---|---:|---:|
+| Input (uncached, fresh content sent to the model) | 381,505 | 0.06% |
+| Output (model generations) | 8,248,370 | 1.25% |
+| Cache creation (first-time write of a prefix into the cache) | 34,376,850 | 5.20% |
+| **Cache read** (re-loading already-cached prefix on subsequent turns) | **617,889,626** | **93.49%** |
+| **Total tokens touched** | **660,896,351** | 100% |
+
+About **661 million tokens** crossed the model boundary during this build. Why cache reads dominate: 1,069 lead-session assistant turns × growing conversation history × Anthropic's prompt caching means each turn re-reads the system prompt + tool definitions + prior turns out of cache (heavy discount) instead of paying full input rate.
+
+63 distinct sessions worth of work participated: lead + 62 subagent dispatches (54 builders + 7 Explore auditors + 1 claude-code-guide). Claude Code spawns each subagent dispatch in its own session; the lead's JSONL only records the dispatch call and the subagent's final return, not the subagent's internal turns.
+
+The full explainer of how to read these numbers (and how the harness UI display ≠ build cost) is in [issue #56](https://github.com/cybertronai/hinton-problems/issues/56). Companion to [schmidhuber-problems #19](https://github.com/cybertronai/schmidhuber-problems/issues/19) — same correction, same machinery.
+
 ### Skills
 
 - **One skill call.** `sutro-sync`, used once at the very start to pull Telegram + Google Docs + GitHub context.
@@ -211,7 +231,7 @@ The session also has frustrated moments. They are part of an honest report: when
 - **53 / 53** Hinton-paper stubs implemented
 - **27 reproduce** paper claims, **25 partial** (gap documented), **1 honest non-replication**
 - **~30 wall hours**, with overnight idle gaps
-- **~800,000 tokens** on Claude Opus 4.7 (1M context)
+- **63 distinct sessions** (lead + 62 subagent dispatches) consuming **~661 million tokens total**, of which **93.49% is cache_read** (re-loaded prefix from prior turns). Harness "~800k" display was current context-window utilisation, not cumulative cost. Full breakdown in [issue #56](https://github.com/cybertronai/hinton-problems/issues/56).
 - **1 GitHub issue** as the SPEC
 - **1 `TeamCreate`**, **53 named teammates**, **11 waves**
 - **18 issues + 15 PRs** filed
diff --git a/README.md b/README.md
@@ -4,6 +4,8 @@ A reproducible-baseline catalog of the synthetic learning problems that appear i
 
 **Site**: https://cybertronai.github.io/hinton-problems/ • **Catalog**: [RESULTS.md](RESULTS.md) • **55 of 55 stubs implemented** (PRs #32–#41 + DBN + DBM add-ons)
 
+**Build cost / token math**: ~661M tokens across 63 sessions (lead + 62 subagent dispatches), 93.5% cache_read. The harness "~800k" was context-window utilisation, not cumulative consumption. Breakdown: [BUILD_NOTES.md § Token consumption](BUILD_NOTES.md) + [issue #56](https://github.com/cybertronai/hinton-problems/issues/56).
+
 ## Introduction
 
 > The field has standardized on backprop by the end of the '80s, and Hinton gives a sample of problems that were used at the time. In the last 20 years, we have transitioned to GPUs, and the math has changed considerably. Instead of being bottlenecked by arithmetic, the shrinking of transistors means that arithmetic is essentially free, and all of the work comes from data movement. **Backprop is inefficient in terms of "commute to compute ratio"** because it requires fetching all of the activations for each gradient add.
