Verify the numbers yourself

By Yad Konrad — @0bserver07

The promise: every number on these pages traces back to analysis/data/sessions.jsonl — the raw transcript of the orchestrator session that drove this build, plus the 58 worker session JSONLs it spawned. The scripts that extracted those numbers are in analysis/scripts/ and the data is in analysis/data/. You can re-run, spot-check, or audit any specific claim.

The pipeline (3 steps, ~3 minutes total)

# 1. Parse Claude Code session JSONLs into structured metrics + dispatch tables.
#    Reads ~/.claude/projects/-Users-yadkonrad-dev-dev-year26-feb26-SutroYaro/*.jsonl
python3 analysis/scripts/analyze_sessions.py

# 2. Scrub foul language and local-path leaks from the dumped data (idempotent).
python3 analysis/scripts/redact_hops.py

# 3. Emit the markdown pages under BUILD_INTERNALS/ from the data.
python3 analysis/scripts/build_artifact.py

Re-running the pipeline regenerates this site’s data-driven pages (sessions.md, cost-rollup.md, orchestration-map.md, all 13 wave pages, next-phase.md) from the raw JSONL. Hand-written narrative pages (Overview, FAQ, What worked / didn’t, How to reproduce, Worker prompt anatomy, Patterns observed, Human in the loop, Pivot moments, this page) are not auto-generated — they cite numbers from the data.

Spot-check any claim

Five quick verifications you can paste into a terminal and run:

“40 Yad-typed prompts to the orchestrator”

python3 -c "
import json
recs = [json.loads(l) for l in open('analysis/data/sessions.jsonl')]
orch = next(r for r in recs if r['session_id'].startswith('63285119'))
yad = teammate = 0
for h in orch['all_hops']:
    t = (h.get('text') or '').strip()
    if t.startswith('<teammate-message'): teammate += 1
    elif t and not t.startswith(('<local-command','<command-name','Base directory','[REDACTED','[Image')):
        yad += 1
print(f'Yad-typed: {yad}  teammate-replies: {teammate}')
"

Expected: Yad-typed: 40 teammate-replies: 142

“1.13 billion tokens total / $3,879”

python3 -c "
import json
recs = [json.loads(l) for l in open('analysis/data/sessions.jsonl')]
total = 0; cost = 0
PRICE = {'input':15/1e6,'output':75/1e6,'cache_read':1.5/1e6,'cache_write_5m':18.75/1e6,'cache_write_1h':30/1e6}
for r in recs:
    if '63285119' in r['session_id'] or 'schmidhuber-impl' in (r.get('first_hop_text') or ''):
        for k, v in r['tokens'].items():
            total += v; cost += v * PRICE.get(k, 0)
print(f'Tokens: {total/1e9:.2f}B  Estimated cost: \${cost:,.0f}')
"

Expected: Tokens: 1.13B Estimated cost: $3,879

“73 Agent dispatches from the orchestrator (58 builders + 15 audits)”

awk -F'\t' '$1 == "63285119-154e-42ab-9555-7a42471b0309" {print $3}' analysis/data/agent_dispatches.tsv | sort | uniq -c

Expected: 58 general-purpose and 15 Explore.

“58 worker sessions, all linked to schmidhuber-impl”

grep -c 'schmidhuber-impl' analysis/data/sessions.jsonl

Expected: at least 59 (orchestrator’s TeamCreate description + 58 worker first-prompts). The Sessions page enumerates them.

“8 direction-changing prompts out of 40”

The 8 are listed verbatim with timestamps in Pivot moments. Cross-check against the orchestrator’s hop list:

python3 -c "
import json
recs = [json.loads(l) for l in open('analysis/data/sessions.jsonl')]
orch = next(r for r in recs if r['session_id'].startswith('63285119'))
# The 8 highest-leverage hops (by timestamp + content match to Pivot moments page)
keys = ['THIS IS WRONG PRACTICE','not rely on me anymore','have we verified','dispathc multiple agents','review it/audit','BUT FIRST FIRST FINISH','its mdBook','why are there teams']
hits = [h for h in orch['all_hops'] for k in keys if k in (h.get('text') or '')]
print(f'Found {len(hits)} direction-changing prompts in transcript.')
"

Expected: Found 8 direction-changing prompts in transcript.

Corrections in this build’s history

Two numbers in earlier drafts were wrong and were corrected in dedicated PRs. Listed here so the audit history is transparent — and so future readers can see the same kind of correction is welcomed:

If you find a number that looks wrong, open an issue or a PR with the verification command and the expected value. Numerical-audit PRs against this catalog are encouraged.

What’s traceable, what isn’t

One last note

The session JSONLs themselves contain Yad’s raw typed prompts. Foul language was redacted in 11 hops (10 frustrated venting, 1 image-path leak) before the data was published — see analysis/scripts/redact_hops.py for the exact redaction policy. Otherwise the transcript is unmodified.