--- name: training-experiments description: Run and audit Blood Bowl RL experiments with immutable reward manifests, paired seeds, provenance/integrity gates, held-out transfer, checkpoint/build discipline, metric interpretation, opponent controls, and DECISIONS.md ledger updates. Use for any training run, reward ablation, A/B, checkpoint operation, evaluation, metric read, transfer study, or promotion verdict in this repo. --- # training-experiments — running the RL program You are continuing a long experiment program. Historical findings were paid for with substantial compute, but later controlled evidence can amend them. **Do not retune from aesthetics, training curves, or human-looking behavior. Change one declared factor through a frozen experiment, validate on match utility and held-out transfer, and write every accepted finding to `DECISIONS.md`.** State as of 2026-07-14 (reward/replay audit active, ledger through D186). Always read `AGENTS.md`, the tail of `DECISIONS.md`, and `docs/reward-and-replay-audit-2026-07-09.md` first. Newer ledger entries and immutable result artifacts supersede older sections of this skill. --- ## 0. July 2026 reward verdict and next experiment No reward has been promoted to production. - The repaired R0–R3 2×2 completed all eight paired arms at exact `249,954,304` native steps with two seeds, ≥10,001 final eval games/arm, identical frozen provenance, and zero integrity counters. Distance supplied nearly all scoring learnability: TD/game `+0.76365`, match score `+0.11130`, ball-forward `+4.45069`. It also increased block volume and 2D-red rate. - The R0/R2 transfer matrix completed 16 cells and 16,076 full games. Removing possession+gain (R2) reduced match score in all eight matched cells and raised losses and opponent TDs in all eight. R2 is not transfer-noninferior. - Keep R0 only as the next experimental baseline. Raw distance `ΔΦ` is a temporary scaffold, not exact PBRS at `gamma=0.995` and not final utility. - Next screen, with distance on in every arm: | Arm | Possession annuity | Ball-gain event | |---|---:|---:| | P0 | on | on | | P1 | on | off | | P2 | off | on | | P3 | off | off | Run `500M × 2 seeds`; then confirm R0 against the simplest transfer-noninferior survivor with learned opponents, roster macro, longer horizon, and a second ancestry. Do not change production defaults before all gates pass. - The July 13–14 attempt completed all arms but is rejected under D182. One training emission exceeded the PPO clamp by exactly `0.015`; the retained schema cannot identify its sign or terminal context. The audit separately confirmed an unsafe result-plus-incidental-shaping terminal path. It has no completion proof and cannot feed transfer. Correct and review terminal composition, add terminal/non-terminal recurrence telemetry, then rerun every arm under one uniform semantic contract. Canonical evidence: - `runs/reward-screens/reward-screen-20260709-v1/SCREEN_COMPLETE.json` - `runs/reward-transfer-20260713-v1/ANALYSIS.json` - `docs/reward-and-replay-audit-2026-07-09.md` ### Immutable experiment contract Use `tools/reward_manifest.py`, `tools/run_reward_screen.sh`, `tools/analyze_reward_screen.py`, and `tools/analyze_reward_transfer.py` for the audited workflow. For every causal arm: 1. Freeze source, installed env/config, imported module, Puffer patches, warm checkpoint, pool, backend, optimizer, seeds, execution order, and eval policy. 2. Supply a complete reward JSON; omitted and explicit-zero fields are different. 3. Record all hashes and persist immutable plan/status/result/completion records. 4. Require explicit phase telemetry, enough complete kickoff games, a final cumulative reprint, and zero clip/non-finite/engine-error/demo/fallback counts. 5. Reject integrity failures rather than averaging them into the comparison. 6. Report W/D/L, TD for/against, both sides, paired common-seed differences, and in-sample versus held-out results separately. 7. Treat two-seed/scripted-bot screens as descriptive, never as confidence intervals, tournament-strength claims, or production promotion. 8. On an episode-ending step, emit only explicit current-step objective reward (TD), result utility, and separately declared episode-terminal terms. Never let incidental action/board shaping co-stack invisibly with the result. --- ## 1. Compute targets **The table below is June historical context, not live inventory.** The paid Vast fleet was stopped after D176. Discover external state before any action; do not restart old instance IDs from documentation. The current free lightweight target is the Tailscale `wsl-ubuntu` RTX 2070. July audit work used the isolated checkout `/home/rache/bloodbowl-rl-audit`; `/home/rache/bloodbowl-rl` may host the production evaluator/stream and is out of scope unless the user explicitly authorizes changes. For any remote run, first record `tailscale status`, live processes, free disk, GPU/driver/CUDA, Python/Torch/compiler, checkout path, Git/worktree state, `_C.__file__`, `_C.env_name`, GPU flag, precision, and module SHA. Do not kill or replace an existing `server.py` process as part of training work. If Vast is intentionally reactivated, use `.claude/skills/fleet-ops/SKILL.md` and live `vastai show instances --raw`; do not copy old IDs or endpoints from the ledger. Use `ssh -n` plus timeouts in monitors, preserve cached replays and checkpoints, and remember that stopped instances can be reclaimed. --- ## 2. Current era: v4 - **obs-v4 = 2782 bytes** — probability planes A1/A2/B. Spec: `docs/obs-v4-spec.md`. - Anchor: **`training/bc_v4.bin`** (val exact 0.508, D53). **It lives ON THE BOXES only** — it is not in the local repo and `fleet.sh setup` excludes `training/*.bin`; copy it box-to-box for any new box. 2.09M pairs in `validation/pairs_v4` (`validation/pairs` is a symlink to it on v4 boxes). - **OBS SIZE SYNC POINTS — all three must agree:** 1. `BBE_OBS_SIZE` in `puffer/bloodbowl/bloodbowl.h` (~line 96) 2. `#define OBS_SIZE` in `puffer/bloodbowl/binding.c` line 8 3. `--obs-size` in `training/convert_checkpoint.py` (`DEFAULT_OBS_SIZE`, now 2782) A `_Static_assert` ties (1)↔(2) at build time; **(3) is the unguarded one** — a wrong `--obs-size` converts to silent garbage (this exact miss bit D54). - **Old obs-v3 lineage checkpoints are input-shape INCOMPATIBLE** with v4. Converting a v3 checkpoint requires explicit `--obs-size 1612`. Never mix lineages in a run or a tournament without converting deliberately. --- ## 3. Launch contracts ### Audited reward screens and transfer Use complete reward manifests under `puffer/config/rewards/` and the immutable workflow in §0. Do not reconstruct a July arm from a dashboard command and do not send it through bare `run_synthesis_c.sh`: historical launcher defaults are not a complete manifest and can silently inject old shaping. Preserve plan, per-arm status/result JSON, completion proof, final checkpoint, and analyzer output. ### Historical curriculum/lineage launcher `tools/run_synthesis_c.sh` has frozen v3 reward defaults, including the poisoned ball-loss profile. It is allowed only for exact historical continuation with a complete explicit override recovered from the original profile. Never use it for a new reward comparison. `ANCHOR`, `LOG`, and `STEPS` are historical env-var controls; choose a checkpoint by embedded step and hash, not mtime. **After ANY env code change, do ALL of, in order.** Source of truth is `puffer/bloodbowl/` (its `engine/`+`bb/` symlinks pull in `engine/src` + `engine/include`); the build compiles the INSTALLED snapshot under `vendor/PufferLib/pufferlib/ocean/bloodbowl/`, so an edit without re-install silently trains on stale rules: ```bash bash tools/install_puffer_env.sh # --check = drift guard, run it when unsure cd vendor/PufferLib && rm -rf build && ./build.sh bloodbowl --float ``` - `--float` is REQUIRED for the torch backend. Plain build = bf16, native CUDA only (that is the build `puffer match` needs — see §9). - **NEVER skip `rm -rf build`** — stale objects bite (footgun 4). - Sync `engine/` AND `puffer/bloodbowl/` **together**; a header without its `binding.c` builds a stale mixture (this hid the tier-distribution exports once, D49). After syncing to a box, `grep` the changed symbol on the box to prove the right code landed, then re-run install + rebuild ON THE BOX. Verify every launch from process state, fresh logs, manifest/status records, and the imported module. Do not trust a historical "LIVE" banner alone. --- ## 4. Reward economy (audited experimental baseline, not settled utility) The canonical R0 values live in `puffer/config/rewards/r0_full.json`; do not retype or partially override them. Important current values are terminal TD `0.4`, win/loss `±0.6`, draw `0`; possession `0.03`, gain `0.05`, loss `0`; distance-to-ball/endzone `0.02/0.04`; block terms `k_kd/k_value/k_ball/k_turnover/k_seq = 0.10/0.50/0.15/0.15/0.03`; rush cost `0.015`; other shaping zero. These define an **experimental control**, not a recommended final objective and not a bit-identical historical continuation. Interpretation after D177–D178: - Distance supplies most learnability. Keep it during decomposition, but call it a scaffold. Raw state delta is not exact discounted PBRS. - Possession+gain has a small screen effect but a consistent defensive transfer benefit. Keep the bundle only until the next screen isolates its components. - Rush and block terms remain scaffolds. Higher block count is not a promotion criterion; R2 blocked more and lost more in transfer. - Stalling is not an automatically accepted side effect. BB2025 explicitly penalizes a carrier who could score without dice; evaluate possession rewards against score/clock behavior and held-out match utility. **Invariants for every reward variation:** 1. Use a complete canonical manifest; never inherit an omitted field. 2. Keep `ball_loss = 0` while the possession annuity is on; no double fine. 3. Any term called a potential must use the exact discounted form `beta*(gamma*Phi(s') - Phi(s))`, or be labeled a non-invariant scaffold. 4. Rush-gate blitz exposure and include rush-failure turnover probability. 5. Record and reject reward clipping or non-finite values. 6. Do not reward realized dice outcomes as shaping. Event/EV terms can still redefine utility and require held-out match validation. 7. Do not promote from human-stat proximity, TDs alone, or training-opponent success. Require W/D/L and TD for/against on held-out opponents. 8. Keep terminal match utility authoritative: incidental shaping on the action that ends an episode must not share the terminal result emission. --- ## 5. Historical curriculum ladder The old Backplay ladder was `maxdist 6 → 9 → 12 → uniform → kickoff`, advancing in three-square increments with retention probes. It remains useful historical context, but July causal reward comparisons use full kickoff starts (`demo_reset_pct=0`). Never compare curriculum TD rates across start depths. --- ## 6. Standing doctrines — IMPERATIVE form These are laws, not suggestions. Each one was bought with a failed run. ### D46 — Discovery vs artifact (the 3-part grounding test) - **NEVER patch out an inhuman behavior because it "looks absurd."** Aesthetics is not a filter (TD-Gammon's doubles, AlphaStar's worker saturation). - Before patching ANY weird behavior, run all three: 1. **Grounding:** does it persist when wins/TDs vs a competent opponent are the binding currency, not shaped income? 2. **Routing:** is its profit causally routed THROUGH the objective, or around it? 3. **Human rates are evidence, not law.** A 10× human rate is a flag, never a verdict. - The verdict belongs to the **tournament**, not your taste. GFI-spam is the case study: presumed artifact (routed around the objective), still adjudicated by A/B — if the untaxed twin keeps high GFI AND wins, absurd is correct and the tax dies. - Treat `reward_rush_cost` as **anti-degeneracy scaffolding** (k-knob category, to be annealed away). NEVER describe or use it as a permanent value statement. - Remember the unifying diagnosis: ball-avoidance, scrum-hovering, and GFI-spam are ONE disease — positional value is denominated in points and there are no points. The cure is a working scoring economy, not taxes. (True GFI price is positional: prone players project no tackle zones; standing presence protects BOTH sides.) Testable prediction on file: GFI discipline emerges FIRST on the carrying team. ### D50 + D56 — Graduation rule - **Run the grounded win-rate exam (4096 full games from kickoff vs a frozen reference) ONLY as the FINAL-stage graduation test.** - **NEVER read a mid-curriculum tournament draw rate as failure.** Mid-curriculum tournaments draw 97–99% — that is EXPECTED. They are calibrations, not verdicts. - Training-tds ≠ winning until starts reach kickoff: endgame skill exists but is unreachable from the opening whistle until backward expansion connects them. - D56 nuance: draw rate RISES with prior strength (96.8% → 98.8% v3→v4) because stronger mirror priors mean fewer accident-decided games. The bar rises with the player — do not panic at a higher draw rate from a stronger anchor. ### D33/D34/D36 — Decision-time exposure pricing (Profile C) - Exposure-EV transfer pricing (`bb_blockev` closed forms; the k knobs; spec in `docs/reward-audit-decision-time.md` Addendum 3) is the **ONLY mechanism known to create contact-seeking**. Do not substitute shaped rewards: shaped/anchored lineages collapse contact to zero (PPO descends the "random contact is net self-harm" gradient — D33, scope includes contact-PLUS-movement risk); unshaped-from-scratch stays at random (D34 control). Only Profile C gives ~2× random volume at ~2× per-declaration conversion. - **ALWAYS rush-gate blitz exposure** (D36). - Exposure knob magnitude moves the brawl pole, NOT scoring (D40/D41). Don't crank it hoping for TDs. - Use `blocks_thrown` (one CHOOSE_DIE per resolved block) as the clean conversion denominator (D36/D34-AMENDED). ### D42/D43 — Anti-farming invariants See §4. Check the invariants before every reward-config launch. No exceptions. ### D28 + D26 — Gradual anneal - **NEVER cold-off a scaffold knob.** Halve per chained stage. Cold anneal decays scoring back to zero (D26: the D2 arm was dead by 5.4B steps). - Applies to exposure knobs too (D41 chain). But note D49: exposure-anneal alone never produced scoring — annealing is how you REMOVE scaffolding, not how you create capability. ### D29 — Every "may" is policy surface - **NEVER auto-resolve a rulebook optional choice.** Auto-policies delete tactics (the Wrestle counterexample reversed D9). All optional choices are USE_SKILL/DECLINE_SKILL or CHOOSE_OPTION windows. The probability engine resolves choice nodes owner-optimally; FUMBBL skillUse reports validate via lockstep AND feed BC pairs. ### D38 — Masked-sampler law - The torch backend **MUST** apply per-head action masks (`masked_fill(-inf)`) before rollout sampling, store masks in experience, and re-apply at train-time recompute. The viewer inherits the same patch. - **NEVER cite tds-emergence intuitions from pre-fix torch runs** (the unmasked era had `illegal_frac 1.000` and a decode-fallback shadow policy — all its measured dynamics are invalid). CUDA-backend results (D26–D28, D34) are unaffected. ### D174/D176 — Current BC and opponent verdicts - D174 supersedes the old "sequence context first" ranking: frame-stacked structural context was null because obs-v4 already encodes within-turn state; last-action history added only `+0.002`, inside noise. Do not fund recurrence without full-game lockstep data and a new falsifiable hypothesis. - D176 supersedes the old blanket "BC anchor prevents erosion" recommendation: full-strength persistent iid CE mechanically held BC accuracy but collapsed offense. Do not reintroduce it as the default and do not infer that a weaker coefficient must help. - Competent opponent quality produced the strongest defensive improvement, but a single deterministic bot is an overfit probe. Prioritize diverse opponents and held-out transfer over more corpus volume or stronger iid anchoring. ### D35 — Optimization-bound before capacity claims - **NEVER declare a BC model "capacity-bound" from val≈train alone.** Verify with extended cosine-decay steps first (bc_v3b gained +1.1pp val that the capacity-bound reading said couldn't exist). ### Misc standing rules - D6: `make goldens` is explicit; rules fixes are EXPECTED to break goldens. - D15: BB2025 team re-rolls are UNLIMITED per turn and replenish at half-time; skill re-rolls once per turn per player. - D25: setup is reserves-first, budget 24 (amends D1's 64+autofix). - D16-AMENDED (D36): check Claws BEFORE Mighty Blow; a Claws break leaves MB for the injury roll. - D42 (same entry as the laundering invariant): filter boxes by cores_effective × clock (766K SPS on 64c/3.4GHz = 7× box-2); cores_effective is the CONTAINER allocation, not the host's. - D44: the spectator event feed is POD-only cross-thread (see footgun 7). - D31: `bb_match.fan_factor[2]` changed `engine_fp` — regenerate the demo-state bank on ANY engine_fp change. --- ## 7. DECISIONS.md ledger discipline The ledger is at `DECISIONS.md` (through D186 as of 2026-07-14). It is the program's chronological memory. D177–D186 cover the reward screen, transfer, replay corpus, streaming loader, rejected partial-deployment run, corrected confirmation, and fail-closed vacation routes that supersede this skill's June snapshots. - **EVERY finding gets an atomic entry. No exceptions.** One decision/finding per entry, numbered D{n+1}, dated. If your write-up needs "and" between two findings, split into two entries. - Write the entry **when the finding lands**, not at end of session. Sessions die; the ledger survives. - Each entry records: what was run (boxes, configs, checkpoints, steps), what was measured (with numbers), the verdict, and what it amends or supersedes (cite the prior D-number explicitly, e.g. "amends D33"). - **NEVER act on an experimental verdict you haven't written down.** Ledger first, then change the default. - Amendments never delete: D9 was reversed by D29, D33 was scope-corrected — the history is the value. Append, cross-reference, never rewrite old entries. --- ## 8. Metric semantics — what the dashboard numbers mean | Metric | Reference values | How to read it | |---|---|---| | `tds` | stage-dependent | Per-episode touchdowns **FROM CURRICULUM STARTS**. NOT comparable across maxdist stages — a drop on stage advance is expected, not regression. | | `block_2dred_frac` | human 0.017; v3 agent plateau ~0.20; v4 0.169 falling | Fraction of blocks thrown at 2-dice-against. **Falling = probability planes working.** | | 2d : 2dred ratio | human 46:1 | Same axis as above. | | `possession_rate` | corrected BB2025 human genuine-turn rate 0.47394; per-game mean 0.47453 | Metric semantics changed historically. Never compare old ~0.15 dashboard values or synthetic turns directly to the corrected baseline; use it as a diagnostic, not a target. | | `gfi` (go-for-its) | human ~2–5/ep; agent ~25–35 | Presumed ARTIFACT per D46 unless it survives grounding. Track it; do not patch it outside the rush-tax A/B. | | `bc_acc` | v4 ceiling ~0.51 (runs hold 0.49–0.55) | Anchor agreement. Don't expect above ceiling; collapse far below it = erosion. | | `blocks_thrown` | — | Clean denominator for contact conversion (one CHOOSE_DIE per resolved block). | | `illegal_frac` | must be ~0 | If ≈1.000 the masked sampler is broken (D38). Stop and fix before trusting anything. | | Tournament draw rate | 96.8% (v3) → 98.8% (v4) mirrors | RISES with prior strength (D56). Mid-curriculum 97–99% is expected. | | SPS | torch ~0.6–0.77M; native 2.1M on 24c (D57); 766K = v3-era 64c reference | Box-throughput sanity check; box-2 is ~1/7 of the 64c boxes. | --- ## 9. Tournament procedure For July reward selection, first run the immutable final-policy evaluation and held-out transfer matrix described in §0. `puffer match` remains useful for checkpoint head-to-heads, but it is not a substitute for both sides, TD for/against, scripted/learned opponent diversity, roster macro, provenance, or integrity telemetry. Native↔Torch conversion zero-fills/drops biases; document and apply conversions symmetrically. 1. Ship both checkpoints to **box-1** (judge GPU). Box-to-box via `ssh -A`, never Mac→ssh4 for big files. 2. Convert: `python training/convert_checkpoint.py --to-cuda -o ` — **mind `--obs-size`** (2782 default = v4; v3 ckpts need `--obs-size 1612`). Wrong size = silent garbage; nothing checks it for you (§2). 3. Run on box-1 from `/root/bloodbowl-rl/vendor/PufferLib`: ```bash puffer match bloodbowl --load-model-path A_cuda.bin \ --load-enemy-model-path B_cuda.bin --num-games 4096 ``` `match()` **hard-errors unless the NATIVE CUDA backend is loaded** ("match() requires the native CUDA backend") — a `--float` torch build cannot serve it; box-1 keeps a native build for judging. A = slot 0, B = slot 1. 4. The result line is `games=4096/4096 A=x B=y draw=z`. If games < 4096, the run died — investigate, don't extrapolate. 5. Interpret per D50/D56 (§6): final-stage exam vs mid-curriculum calibration are different instruments. --- ## 10. Footguns (each one cost real hours — obey) 1. **Python `str.replace` silently no-ops on anchor mismatch.** ALWAYS `grep` the file after any scripted edit to prove the change landed. 2. **zsh does NOT word-split unquoted vars** — `set -- $hp` fails. Use arrays or explicit splitting. 3. **`pkill -f` matches its own watcher commands.** Use `[b]racket` patterns: `pkill -f '[r]un_synthesis'`. 4. **Partial code syncs build stale mixtures, and the build compiles the INSTALLED snapshot, not your edits.** Sync `engine/` AND `puffer/bloodbowl/` together; re-run `tools/install_puffer_env.sh` (use `--check` as the drift guard) + `rm -rf build` + rebuild on the box; verify with `grep` on the box. 5. **Vast stopped instances can be RECLAIMED.** Restart promptly or accept a replacement (new id, new ssh endpoint). Box-1 holds the SOLE replay-cache copy — treat it accordingly. 6. The balance ladder (`/tmp/vast_ladder.sh`) stops boxes in value order below $5/$3.5/$2 credit. Stop-only, reversible — but see footgun 5. It lives in Mac `/tmp`, which a reboot wipes. 7. **Memorial/feed render hooks: NEVER format strings or do I/O from env-stepping threads.** POD staging slot + render-thread consumer. Two SIGSEGVs taught this. 8. **raylib `InitWindow` segfaults when the Mac display is asleep.** `spectate.sh` gates on display-awake — keep that gate. 9. **Monitor/SSH loops need `ssh -n` and `ConnectTimeout`** or they hang forever. 10. (Launcher) Asymmetric step-cap overshoot ~1.5× is known and benign (§3). 11. **`vastai` CLI < 1.0 (brew 0.4) creates zombie instances.** Check the version before provisioning. --- ## 11. Open experiment queue — PRIORITY ORDER Work the queue top-down. Do not reorder without writing a ledger entry justifying it. 1. **Decompose R0's bundled possession family with distance fixed on:** both, possession-only, gain-only, neither; `500M × 2 seeds`, same frozen contract as D177. R2 is prior evidence for neither, not permission to skip a fresh matched cell. 2. **Confirm R0 against the simplest transfer-noninferior survivor:** `1B × 2 seeds`, learned-opponent transfer, both sides, roster-grid macro, TD for/against, and uncertainty. Reject any simplification that repeats R2's defense loss. 3. **Separate opponent curriculum from reward:** freeze the selected reward and compare diverse competent opponents. D176 shows opponent quality can dominate apparent defense improvements; a single deterministic bot is an overfit probe. 4. **Make distance policy-safer:** replace raw delta with exact discounted PBRS, or anneal it to zero and demonstrate non-inferiority against objective-only. 5. **Ablate rush and remaining block scaffolds one family at a time.** More blocks or more human-looking rates are not benefits unless held-out match utility improves. 6. **Reproduce the finalist from a second ancestry** (initial candidate: league9), then run the final long matched control. Only this stage can authorize a production default change. 7. **Improve replay sampling:** edition-exact, replay-disjoint, hierarchical depth/action/roster sampling with setup caps and grouped metrics; build replay-derived scenarios only after full-game lockstep coverage is verified. ### Background threads (not queue items, but don't lose them) - The strict BB2025 replay corpus is opening-censored and nearly lacks pass, handoff, foul, Jump, TTM, and Special targets. Do not claim rare-action or late-game learning from aggregate BC accuracy. - Recurrent sequence training remains contingent on full-game lockstep conversion; frame-stacked context was null in D174. - Asymmetric-stop overshoot remains low priority while bounded and manifested. - FUMBBL differential adjudications pending: D11 crowd-push dice order; D13 Secure-the-Ball bounce+turnover; D18 Break Tackle always-on + Secret Weapon approximation; D21 Saboteur auto-KO; D36 documented engine-parity divergences (Frenzy-vs-Steady-Footing missing second block ≤5.6% pool mass; Dump-Off dead once-latch); D9-class auto-policy declines generally. - Phase-3 rule TODOs: D4 kickoff events (7 of 11 no-ops; D31 landed Solid Defence/Quick Snap/Charge windows); D21 deferred Multiple Block + Pile Driver, Bloodlust/Animosity keyword data, Dodgy Snack drive-end restoration (needs procedural squads); D31 deferred Kick-off Return (BB2020 relic — stays a classified mapper divergence), Charge! up-front selection, TTM/KTM dice mapping; D19 Shadowing (needs per-turn MA-use counter); D12 Pitch-Invasion fan-factor modifiers partially landed in D31. - D23 decode low-x fallback bias: documented, not fixed (irrelevant under masked sampling). - The core July rules/reward corrections and experiment tooling are commit-backed. Production reward defaults remain unchanged: install/rebuild the corrected module only in an isolated experiment checkout for the next manifested run, then require the full promotion gates before any default. ### Unattended multi-day execution Do not encode the priority list as an adaptive agent that chooses its next arm from live metrics. Before an unattended window, resolve every branch that can be resolved, then freeze literal commands and expected artifacts in a schema-1 `tools/experiment_queue.py` plan. Every job needs a maximum runtime and success validator; long jobs need a progress artifact and short jobs need an explicit progress exemption reason (the exemption is capped at 30 minutes). Use typed command/validator/environment values: literal values cannot carry paths; immutable file and directory-tree inputs are pinned; output paths are declared mutable; generated inputs link to an earlier job's recorded artifact. Pin and recheck every executable and transitive input by byte size and SHA-256/tree identity, and use only the plan's allowlisted base environment. The queue root must be the isolated audit checkout, and the service must use `KillMode=control-group`. Every queue invocation is a pinned executable plus a reviewed, hash-pinned runner file. Literal arguments are limited to numbers, lowercase SHA-256 digests, and long flags; categorical/free-form strings belong in the pinned runner or config. `resume_safe` means the job's own runner validates a frozen manifest and every partial/completed artifact before continuing. It does not mean “probably okay to rerun.” A plan drift, failed gate, missing/stale progress file, disk limit, sustained thermal limit, or invalid success artifact halts the queue and leaves later work pending. An unattended queue may produce evidence but may never promote a reward or production default. Use `docs/vacation-autonomy-2026-07.md` as the operational checklist. A persisted halt is terminal. Do not restart it in place after editing state or artifacts. Preserve the halted evidence and create a newly reviewed queue ID/plan/state if human diagnosis authorizes follow-up work. For the July vacation contract, do not hand-author the typed plan. After the main `1B x 2` paired screen and both transfer strata pass, use `tools/freeze_vacation_queue.py`. Its literal order is: second-ancestry `1B x 2`, second scripted transfer, second learned-anchor transfer, two-lineage gate, main `paired-final`, second `paired-final`. `paired-final` is a fixed six-arm schedule (`both/candidate x seeds 42,43,44`) at `6B` requested steps per arm. `tools/run_reward_learned_transfer.py` loads each focal policy in both native backend roles against every pinned anchor. `tools/vacation_reward_gate.py` must revalidate both self-play screens and both transfer strata before producing the only artifact that unlocks the final screens. Never mark a PPO screen `resume_safe`; its partial optimization trajectory is not restart-equivalent. There are two reviewed no-candidate primary routes. If and only if the exact decomposition analysis recommends `both` and its eligible-candidate list is empty, freeze `candidate_arm=both`, null `anchor_config` and `main_learned_complete`, and `final_steps=12B`. The freezer then emits exactly two `control-final` jobs: R0-only seeds 42/43/44 from the main ancestry, then the same schedule from `league9`. This preserves the `72B` final budget without training a rejected objective. If any simplification remains eligible, the fallback must fail closed. The rejecting decomposition screen must also satisfy the full current config-tree, explicit default-config, compiled-module, exact nine-file runtime, patch, warm, and pool provenance contract. A legacy screen missing any of those identities is screening evidence only and must be rerun before it can authorize the control queue. The scripted matrix must also use `both` as reference and evaluate all three simplifications. An empty eligible list from a subset is not an all-candidates-rejected result. If an already-selected candidate instead fails the unchanged paired confirmation gate, the explicit `confirmation-rejected-baseline` route may emit the same two R0-only jobs only after independently regenerating the failure and matching the confirmation's embedded selection-transfer proof. It is not an all-candidates-rejected result and cannot switch candidates. D186 adds one optional, separate post-primary overflow because measured local throughput can leave the owned GPU idle before return. Use only `tools/freeze_vacation_overflow.py`: exact primary completion and its original validators unlock one unchanged R0 `control-final` screen from the static pool's exact netblock ancestry, `12B x seeds 42/43/44`. The delayed timer also requires primary-service inactivity, unchanged pins, no existing overflow state, and no GPU compute PID. It cannot run after primary failure/halt/drift, relaunch an interrupted PPO screen, choose a reward, or change production. Vacation screen jobs must invoke `tools/run_frozen_reward_screen.py`, which sets `ARM_DETACH=0`. The queue creates a new session for each job, and every screen, arm wrapper, trainer, and descendant must remain in that process group so its runtime/progress/capacity/thermal guards cannot strand GPU work. Do not add an inner `setsid`, `nohup`, daemonizer, or process supervisor to a queue-owned training path. --- ## 12. Session checklist Before doing anything: read `AGENTS.md`, D177–D186 and the ledger tail; discover live Tailscale/Vast state and processes; confirm the intended checkout is not the production checkout. Before launch: verify obs size, install drift, imported module/provenance, disk, complete reward manifest, opponent/data hashes, seeds, execution order, eval gate, and integrity rejection rules. After a run: analyze the immutable result, copy/hash the cap, separate in-sample from held-out evidence, append one atomic ledger finding, and only then consider a default. ## Run length / stopping rule (D168) — plateau, not a fixed budget Reward-economy / behavioral DIAGNOSTIC probes are NOT stopped at a fixed step budget. - **Extend-on-still-improving, stop-on-plateau.** Track the primary metric(s) the probe tests (block_2dred_frac, offassist_2d, etc.). At each cap, compare the metric over the last ~2B window to the prior ~2B. Still moving beyond noise → **chain a warm-restart from the cap** (same reward config + pool preseed; total-timesteps can't be raised mid-run) and continue until it flattens. Flat, or the verdict is already decisive → stop + record. - Judge "still improving" on the metric that MATTERS: behavioral curve for a probe; Elo-vs-frozen-anchors for a strength/final run — and there also stop on the EvalStop overoptimization downturn (k=2 consecutive Elo drops), not just plateau (a proxy can climb toward human while true strength regresses). - Probe = stop at behavioral plateau / clear verdict (cheap). Final policy = run to a real budget / Elo plateau for max strength. - To extend: warm-restart from the latest snapshot/cap, same config, same `--selfplay.league-preseed`.