[fdrocha]

FreeCiv v3 runs: each run is one configuration (players, barbarians, fork turn, horizons, replays, anchor seeds,
natcond, pod machines, network volume) in data/freeciv/fbsim_v3/runs/<run>/, driven end to end by nine numbered
step scripts. The draw, elicitation and scoring scripts take the fork turn, horizons and player count from the run;
without a run they keep the paper's settings, rules and outputs.

• Run folders and run.env for the fleet: fleet/1_new_run.sh creates a run (run.env, shell.sh, empty
  fleet_state/), refusing overlapping seeds and a dirty checkout; paths.sh, the pod scripts and runcfg.py read
  run.env; pods are named fbsim3-<run>-… and archive to /vol/fbsim_v3/<run>/.
• Draw settings from the run: questions, criteria, reports and the natural-conditional preamble use the run's
  fork turn and horizons. draw_v1.py skips empty draw cells and stops with a clear message when the bank, tails
  or mirrors set has no candidates (too few replays per anchor).
• Bank cells take turns; one-player subjects recur per band: with a run, the bank's (q band × horizon) cells take
  one question each in turn instead of filling one after another, so low bands no longer use up the subjects the
  higher bands also have. With one player, a world has one subject per family and horizon, so the one-per-subject cap
  applies per band. On p1_t60 the bank goes from 276 questions (149/91/15/12/9 per band) to 676 (149/140/130/123/134).
• Prompts and folders from the run: the background sentence, no-news turns and report names come from the run;
  sets, reports, panels, --out and scores default to the run folder. score_v2.py scores a draw without natural
  conditionals, as a one-player run has.
• Unparsed answers left out of the scores for runs: an unparsed binary answer is no longer scored as 0.5; its row
  stays unscored so the parse rate still counts it. Without a run the paper's 0.5 stays the default.
• norm_constants_v1.py: writes continuous_norm_constants.json by the paper's rule (reproduces the paper's file
  byte for byte).
• finalize.sh stops at consolidation, and on errors: the unused sweep is no longer run; a failed pull, QC or
  consolidation now fails the step, and a stale consol_v4_SUMMARY.json is removed first.
• Retry failed lane launches: fixes a race on rp.sh's address cache that failed parallel launches; a launch
  that ends without lanes is reported with an ERROR line and retried up to --max-retries times (default 5).
• pod_setup survives a dropped ssh session: pod_setup.sh runs detached and bringup.sh polls its exit code,
  so a dropped connection no longer fails a healthy pod.
• worker.sh restarts lanes left as zombies: a lane counts as alive only if its pid still runs that lane; kill -0
  also succeeded on exited lanes left as zombies, so a top-up after a rebalance started no new lanes and workers sat idle.
• Runs shortcut in fleet/: fleet/runs links to the runs folder (made by setup_worktree.sh).
• Numbered step scripts: 2_anchors.sh … 9_score.sh, one step each, rerunnable after an interruption; pods
  are deleted only after a y/N prompt.
• Docs: RUNBOOK.md (one run end to end, the run-only draw and scoring rules, existing runs), README, AGENTS.md,
  PROVENANCE.json, CLAUDE.md.

Testing

• Full run on RunPod: p1_t60, one player, fork at turn 60, anchors 9301–9308 × 1,000 replays, 16 workers
  (October 29–30). Steps 1–9 all ran: 8,000 replays with no failures or truncations, corpus QC and consolidation
  clean, about $77 of compute. Step 8 elicited 22 models for $16.78 (Qwen3 235B A22B and o4 Mini are no longer on
  OpenRouter); rerunning it with the same label resent only the calls whose stream had dropped. This run found the
  zombie-lane bug, the scorer's missing-natcond failure and the skewed one-player bank fixed here.
• Smoke test: steps 1–6 for p1_pilot2 (3 anchors × 10 replays, 4 workers, October 28), which found the
  address-cache race and the dropped-ssh bringup failure. On it, step 7 stops with the new tails message.
• Without a run, consolidation, bank, draws, criteria, reports and scores are byte-identical to before: on the
  1-player pilot, on two of the paper's 5-player anchors, and for draw_v1.py and score_v2.py on p1_t60's data
  after the draw and scoring changes. A T120 copy of the pilot gives the same sets with every turn moved by 60.
  norm_constants_v1.py reproduces the paper's file.
• Locally, with stand-ins for ssh/rsync/setsid: bringup.sh's detached pod_setup (exit 0, exit 3, timeout),
  pull2.sh/finalize.sh failure paths, and launch_pending.sh retries up to the limit.
• Not tested: any T120 run on RunPod; any run with two or more players (natural conditionals with a run); the
  zombie-lane fix as shipped by a normal bringup (on p1_t60 it was copied to the running pods by hand, and the lanes
  it restarted finished the run).

🤖 Generated with Claude Code

https://claude.ai/code/session_01Tze4TwpX5r8wVBbqZaQpZZ