← index week of 2026‑07‑17 · TTSO · Framework · briefing·actor·sandbox·engine·fold·trigger·judge·miner·park·store trace viewer →

framework · briefing → actor → sandbox → engine → fold → trigger → judge → miner → park → store → (again) briefing

Click a node — real round excerpts and full skill text open.

10 nodes run a single loop. actor checks its work for free in sandbox (simulate), then commits a live action to the engine; the residual fills fold's single (α,β) gate, and trigger (SURPRISE/GAME_OVER) wakes sleep — judge scores it, miner writes skills, park holds unpromoted candidates, and store promotes/demotes them forward into the next briefing. All numbers and quotes are verbatim from data/{turns,skills,runs,variations}.json (2026‑07‑16/17 c0716h9/c0717h10 runs + ttso/ source) · hover=preview · click=pin · tabs=Source·Live·Variations · #anchor deep links.

01

7 Runs — read the table first

runs.json · c0716h4~h9 · c0717h10
runroundsactionslevelswinsleepnotable

Only h9 shows a measured L0→L1 grammar transition + r9 GAME_OVER→r10 RESET; h10 goes through two GAME_OVERs (r9/r19) + an r8 wincond demotion (credence 0.18).

02

Pipeline — click = real data

10 nodes · 1 loop
store ↻ briefing Promoted skills appear in the next round's briefing ordered by credence — one unified loop.
hover = preview in right panel · click = pin full content · on mobile the panel sits below the chart.
03

Full Raw Text — 4 embedded data sets

verbatim · exactly as this page's <script type="application/json"> holds it
data/turns.json — 5 live round excerpts (c0716h9_thin_s0 ×4, c0717h10_thin_s0 ×1)

Each record: source, round, tag, hypothesis, expect, why, think, reasoning (full THINK+code block), action, feedback (full engine OBSERVED+briefing), state, levels. This is the same JSON the pipeline panel above quotes from — readable as-is in the #turnsData script tag via the browser's "view page source".

data/skills.json — 4 full skill texts (r10 promoted mechanic · r8 demoted wincond · r77 parked wincond · r78 metathinking REASONING LOG)

Each record: tag, id, kind, level, when/predict_next/clears/target/propose (full Python code), doc (claim/quote/when_to_use/when_not), scope, parents, alpha, beta, evidence, probation, born_sleep, status, future_f. Readable as-is in the #skillsData script tag.

data/runs.json — 7 run summaries (h4~h10)

Each record: run, file, rounds, actions, levels_reached, state, win, sleep_count, notable. Readable as-is in the #runsData script tag. Same source as the §01 table above.

data/variations.json — 3 variation contrasts (trigger 1-strike vs 2-strike · fold GAME_OVER present/absent β · actor PROMPT vs PROMPT_THIN)

top-level: trigger_1strike_vs_2strike (rule_code + measured h9_L1_r6_r12 + h10_r1_r8 sleep rounds), gameover_learning_yes_no (h9 frozen predictor r65 vs h10 empirical accumulation r9/r10 + constants), prompt_vs_prompt_thin (PROMPT/PROMPT_THIN full text + selection logic). Each node panel's Variations tab reads directly from here — readable as-is in the #variationsData script tag.

provenance
turns.json/skills.json/runs.json/variations.json are static snapshots pre-extracted into presentation/codenative/2026-07-17/data/ — this page reads them not via fetch but via 4 inline <script type="application/json"> tags at the bottom of the file with the values baked in directly (works even over file://), parsed with JSON.parse. Original sources: logs of the two runs under runs/ (c0716h9_thin_s0, c0717h10_thin_s0) + level_0/1/2_reasoning_log.md under runs/s5h9/g2-on-ft09-s0-s5h9/·runs/s5h10/g2-on-ft09-s0-s5h10/ + bundle.jsonl (the persist directory of the same two runs, the source variations.json quotes) + ttso/ source (control.py, evaluate.py, prompt.py, thin_run.py, run_game.py, g2_run.py).
GO

Onward