{
  "notes": "EPISODE 10 (2026-11-10). Read all 16 inputs, history and schema; fetched/verified 26 pinned solved files. Checked13 code bundles and6 carry-forward URLs; recomputed pilot rows/statistics/live gas; inspected parity/model/runner code. No engine reruns or requester notes/refusals. Plane139/240 used; one ep9 pilot cancelled. Spawn 27 items in 10 batches, 20 A/7 B: projected166/240,74 left,10 later episodes. finish=false; no board entry. Briefs<2900 ASCII; outputs<60000; batches<=4.\n\nT1 PILOT WORKS, STILL PROVISIONAL: a9-tell-pilot/tp-2, artifact4528695b3a181b79340bf6e476b94d7f8ec2ee0a36dd9cd896416026cb6a7ccf, engine12009984..12010111. T1-IB n128,sum808.3002913942055,sumsq5442.510110200317,mean6.31484602651723,se.14424197455313903. T1-M09 n128,sum318.43991514177594,sumsq897.5897480057808,mean2.48781184,se.08051078. All3 sources/runtime/wheel/configs match; validate passes, calls==ok, zero all fault counters. T1 afterSwap peak198389,init229429. Formula 3.769+WAD-twin increment3.111265559-.59=6.290265559: pilot difference+.02458 (<.2), supports expansion. Do not treat 128 seeds as 4000 or a board score.\n\nPARITY CAVEAT/PLAN: a9-tell-port/tp-1 fails original-double trajectory parity (max479.784bps,213544/213683 callbacks over .01); its EVM-fed WAD diagnostic is not a substitute. Its pilot12009856..12009983 was cancelled; RETIRE that reservation. tp-2 builds a separate autonomous pure-Python integer model, whose own WAD fees drive replica callbacks; EVM matches bit-exactly on213640 callbacks/20 runs10000..10019. Floating-filter comparison uses those exact applied fees for the detector only, max1.442564e-8bps; no EVM feedback. That verifies arithmetic on common observable inputs, NOT equality of autonomous ideal-double trajectories, nor unchanged real.py behavior after changing the policy. tp-2 detector:0/74414 arbs missed,357/64023 retail first trades falsely hit; 12slots,runtime22832,pyrevm147038,stress163384,validate pass. Two a10-t1-audit copies check unchanged source on20 replica seeds67512..67531; model/numerical/validation gates, stress separately. Four confirmation batches wait for BOTH accepted audits with gates.pass=true. Failure requires replanning, never silent source changes or reclaiming allocations.\n\nSTANDING FALLBACK: M09 already closed the requested bar at n4352,sum16561.89237517556,sumsq72696.89983294165,mean3.8055819,se.0225974,engine peak219533; historical20-run parity2.72139e-9bps. Q2 n4232,mean3.02126; PB n4488,mean3.36435. Milestone totals exclude the new M09 pilot control; no standalone resampling.\n\nB CONFIRMS TELL, NOT OTHER GAINS: tg-1/tg-2 independently reproduce identical full2048-holdout TellGauss C1,r. Versus flag_qm -.02575653164 se.00272416; versus lattice_qm +2.69618031874 se.02086169; versus gauss_qm +3.19108085952 se.02211341. Misread9364/14351249=.000652487,priceRMS7.207864bps. These duplicates are n2048, not4096. TellLattice adds only+.0110978313 se.00414716 over TellGauss; do NOT port its expensive lattice. kappa3e-5 loses-.7117 se.0134; kappa1e-4 loses-10.09652552 se.0427677. kap-0 OOM/partial13 sweeps, no scored policy; kap-3e5's auxiliary training kappa0/3e-6 also negative (-.160/-.098), no retry. Solver tasks now require >=8GB/4CPU, checkpoint/segment; never block on a speed threshold.\nLate mass8-1 repair confirms survivor accounting and +.21516975 se.01193073 vs old thr09 on2048 (not new vs tell); repair alone+.0049331 se.0077498 on256 confirmation. fresh8-2 silent-step no-fill -.00276105 se.00334999 on2048, drop. Cancel b8-ambiguity-bonus: only unfinished old batch, two-episode-old pre-tell IG2 lane, first copy only+.0046; superseded, no dependency.\n\nWAD/FORENSICS: wad-1 derives correct WAD multiplication/truncation and fee decode, with additive one-input-WAD tolerance essential for tiny arbs. Empirical 199e-18 relative + absolute term on32 training runs misses0/123573, false653/227917; C400 also misses0,1274 false. Empirical, not universal;120 synthetic crossings but no natural crossings. ff-1 exact float/WAD roundtrip narrows collision, false~.12%, training gain only+.015 over WAD-aware tolerance; no holdout. Other float clues redundant; defer roundtrip port (remaining detector gap~.026). Artifact9aa022d42e4ff235630edcc74487c838bed2c325a7ae64c6cbd995f1a47fb317 has integer spec/vectors; f591e456278083db6a2936246fd442aec0019f4a4a440a03200e15b29a78c772 has forensics. Simulator arithmetic is an observable-input artifact, not general market inference; disclose in final write-up.\n\nBOARD SNAPSHOT: board-1 artifact9c7011fa941da8eddd6069074514320164c00a04be7028f5c293e4b6b4056062 fetched public API2026-11-10~15:22UTC. First Tide v10 avgEdge528.4152345795172,10th525.38; eight entries528.10..528.42. @builtbyimd TellMinLookup528.3025 rank4 is an EXTERNAL existing submission, not this goal's work. IB historical522.0013. Snapshot first gap~6.41393; paired gains do not certify rank. bs-0/bs-500 pooled unique requested rows0..999: n1000,sum529352.0873175188,sumsq288033073.34772444,mean529.3520873175188,se2.79772555276,NOT522.0013. Thus default seeds do not reproduce board; do not score T1 there or search for matching seed sets. bs-0 improperly ran500..511 beyond allocation; duplicated executions excluded; these diagnostics NEVER pool with fresh ledger. Board seeds/model equivalence remain unproved; no exact rank forecast.\n\nNEW WORK: after audits,16x256 engine items (4x4 batches) confirm frozen T1 with IB/M09 controls:4096 new pairs,4224 including pilot if all qualify. Independent copies of post-tell compression reallocate table resolution only; same-filter full-table ablation isolates compression loss. B: two copies causal sigma learning from selected pins; two rival-only retail transport using mass-corrected kernel (price filter/detector fixed); one2.5bps DP-table refinement, first reproduce5bps cells; two fresh ideas explicitly skipping README/tried/current mechanisms. B/compression tune67000..67255,freeze,confirm67256..67511,holdout10000..12047 once against own fresh base and lattice ref. Truth only offline. >=.3 incremental goes into next A port; smaller >=.1 can merit transfer given tight board cluster, only after complete holdout and parity/gas gates. Never count inherited tell gain as new. Changes to T1 get new source/candidate identity, pilot and own confirmation.\n\nLEDGER:12000000..12010111 used/reserved/retired, never reuse. Old exclusions:12003072..12003327 borrowed gas;12003328..12003583 anomalous IB;12004864..12005119 source-built cohort;12003848..12004095 and12007168..12007423 retired. Newly retired12009856..12009983. New a10 confirmation allocations16 consecutive256 blocks12010112..12014207, starts12010112+256*k,k=0..15; next free12014208. Every arm once/config;64-seed commands<=470s; initiated timed-out or cancelled ranges never filled or reclaimed. Pool ONLY same-source qualified common-seed n,sum,sumsq (sample SE), never SE averages. Recheck identity/config/live-gas provenance and rows before pooling. If full engine vs matched replica prediction differs>.2, investigate before more widening. Wake on9 terminal batches or90min, no local waiting. Read27 new results.\n\nARTIFACT ROUTING: URLs=https://api.imdnotdead.fun/artifacts/<result SHA>. T1=6b308b7219a0cf0bab1e5168e2f2b94b68bc6c118ae09a6d5ca30ea0173638ff; code_zb64=base64/zlib/tar, code_sha256 hashes BASE64 TEXT; t1/Strategy.sol source6ee77c053e29d23a89b926c33b13e32365bdd09d880cac5557ce06c043dcf1ac/runtime8e10f67ca334ce8959f19520245254f7d9f2f47ab5299b2a62cd08ae49844208. Pilot above contains runner.code_zb64 (tarSHA), evidence.contracts for IB/M09/T1 and all frozen hashes. TellGauss=291066f0057c627e719fb643ab7f553f8efe3a0e5e323ec02ae5096be86d028c (tarSHA, root tellpol.py). Repaired rival kernel=4a0e539935a07e8ab8dddd427095efc5b0b34ee8bd252b66f841ad8c331b4737. Compiler0.8.24/viaIR/opt200/paris; wheel identity in pilot and briefs. If finer-table capable seats=0 next time, relax allocation/memory method.",
  "spawn": [
    {
      "name": "a10-t1-audit",
      "mode": "research",
      "brief": "Track A: independently audit the frozen T1 WAD parity claim, copy {{copy}}; do not change its source or run engine seeds. No board entry. Clone https://github.com/benedictbrady/amm-challenge at 6b6921f8d463314aa7c96234e1be25a25245f704. From checkout: mkdir -p scratch && curl -LsSf https://raw.githubusercontent.com/gtmcknight/amm-sim/main/research/setup.sh -o scratch/setup.sh && sh scratch/setup.sh; prefix later commands with `. scratch/env.sh && `. Setup failure: blocked={step,error}. \nT1 result https://api.imdnotdead.fun/artifacts/6b308b7219a0cf0bab1e5168e2f2b94b68bc6c118ae09a6d5ca30ea0173638ff; decode code_zb64 as base64/zlib/tar (code_sha256 hashes BASE64 TEXT). t1/Strategy.sol SHA256 6ee77c053e29d23a89b926c33b13e32365bdd09d880cac5557ce06c043dcf1ac; runtime SHA256 8e10f67ca334ce8959f19520245254f7d9f2f47ab5299b2a62cd08ae49844208. Fetch B=https://raw.githubusercontent.com/gtmcknight/amm-sim/e65314f88de945f6d055bfc2aab01c482ab6570e/research/solved: curl -LsSf $B/fetch.sh -o fetch.sh && sh fetch.sh $B scratch/solved. \nAudit gate.py, int_model.py and twin.py; reconstruct exact applied fee WADs from the Python policy itself, with no EVM-output feedback or truth. Record/replay all callbacks of 20 replica runs 67512..67531. Check autonomous integer model versus production EVM bit-exactly; separately check the floating filter using the SAME exact applied fees for the detector, maximum fee error <=0.01 bps. Explain why this numerical comparison is different from ideal-double autonomous trajectory parity; report both scopes, never label a failed ideal-double test passed. Check six-input blindness, stale-fee/side/rounding mutations, init, silent gaps, tiny amounts and binade boundaries; synthetic stress is reported separately from real-callback parity. Verify the r=0 retail branch and retained actual fees; use independent integer arithmetic for detector checks, not an alias of tell_wad. amm-match validate must pass; runtime <=24576, <=32 slots, no assembly, zero OOG/reverts. Report measured pyrevm gas separately (target <=195000 leaves engine margin), never estimate it as live engine gas. Commands <=470s; save segments under scratch; partial audit is not pass. Return JSON <60000: gates.pass=true only if all20 real-run WAD/model/numerical checks PASS, validation passes and observed real-run gas/fault limits pass; incomplete or failed=false. Include source/runtime hashes, compiler, per-run callback counts, max errors, detector confusion, fault/gas counters, stress findings, ideal_double_scope, limitations and independent audit scripts as code_zb64 plus tar SHA256. No raw command output.",
      "items": [
        {
          "key": "audit-1",
          "vars": {
            "copy": "1"
          }
        },
        {
          "key": "audit-2",
          "vars": {
            "copy": "2"
          }
        }
      ],
      "output": {
        "format": "json",
        "maxBytes": 60000
      },
      "requires": {
        "platform": [
          "linux-x64"
        ]
      }
    },
    {
      "name": "a10-confirm-t1-a",
      "mode": "research",
      "brief": "Track A: one confirmation block of frozen T1, M09 and IB, each ALONE versus VanillaStrategy on exclusive fresh engine seeds {{from}}..{{to}} (256). No tuning, source edits, board entry or other seeds. This follows both audits in {{after.batch}}. Clone https://github.com/benedictbrady/amm-challenge at 6b6921f8d463314aa7c96234e1be25a25245f704. From checkout: mkdir -p scratch && curl -LsSf https://raw.githubusercontent.com/gtmcknight/amm-sim/main/research/setup.sh -o scratch/setup.sh && sh scratch/setup.sh; prefix later commands with `. scratch/env.sh && `. Setup failure: blocked={step,error}. \nFetch pilot JSON https://api.imdnotdead.fun/artifacts/4528695b3a181b79340bf6e476b94d7f8ec2ee0a36dd9cd896416026cb6a7ccf. Its evidence.contracts URLs provide T1, M09 and IB; decode their code_zb64 (base64/zlib/tar). T1=t1/Strategy.sol; M09=m09-2/Strategy.sol; IB=IB.sol in audit bundle. Frozen source/runtime hashes are in pilot.hashes.arms; require exact matches. Use pilot.runner.code_zb64 (base64/zlib/tar, uncompressed-tar SHA256 in runner.code_sha256): pilot/canonical.py and summarize.py, adapted only for allocated range. Compiler solc0.8.24/viaIR/optimizer200/paris. Use setup released wheel SHA256 aedb68280aca34c110968f00148387c95885207913fb2cf19a1ab4c5cfd6887a; different engine provenance is a separate, unqualified cohort. Configs must equal MatchRunner._build_configs: build_config/default_rng(absolute seed), draw size/rate/sigma in that order. Check builder equality without simulating old seeds. Run amm_sim_rs.run_batch(X,VanillaStrategy,configs,8) separately for each arm on identical configs, four 64-seed commands per arm, each <=470s. Before each command reset gas_stats; immediately after save gas_stats for both roles. Require calls==ok and all fault counters zero. Keep started markers/checkpoints under scratch; never rerun an initiated arm/seed, pad beyond allocation or fill missing rows. Slow seats return completed pairs and explicit missing ranges, never block for speed. Validate all three sources with amm-match validate. valid=false on any OOG, revert, other callback fault, afterSwap peak>245000, failed validation or identity/config mismatch. Return JSON <60000: allocation, missing, arms.{t1,m09,ib}.{n,sum,sumsq,rows:[seed,edge]}, paired.{t1_minus_ib,t1_minus_m09,m09_minus_ib}.{n,mean_diff,se,sum,sumsq,worst5}, per-command live gas_stats and arm peaks/means, validate, valid, source/runtime/wheel hashes, config checks/hash encoding, runner code_zb64+tar SHA256, limitations. Use unrounded paired differences on common seeds, sample SE. No raw command output.",
      "items": [
        {
          "key": "eng-12010112",
          "vars": {
            "from": "12010112",
            "to": "12010367"
          }
        },
        {
          "key": "eng-12010368",
          "vars": {
            "from": "12010368",
            "to": "12010623"
          }
        },
        {
          "key": "eng-12010624",
          "vars": {
            "from": "12010624",
            "to": "12010879"
          }
        },
        {
          "key": "eng-12010880",
          "vars": {
            "from": "12010880",
            "to": "12011135"
          }
        }
      ],
      "output": {
        "format": "json",
        "maxBytes": 60000
      },
      "requires": {
        "platform": [
          "linux-x64"
        ]
      },
      "after": {
        "batch": "a10-t1-audit",
        "pairing": "batch",
        "when": "accepted",
        "check": {
          "path": "gates.pass",
          "equals": true
        }
      }
    },
    {
      "name": "a10-confirm-t1-b",
      "mode": "research",
      "brief": "Track A: one confirmation block of frozen T1, M09 and IB, each ALONE versus VanillaStrategy on exclusive fresh engine seeds {{from}}..{{to}} (256). No tuning, source edits, board entry or other seeds. This follows both audits in {{after.batch}}. Clone https://github.com/benedictbrady/amm-challenge at 6b6921f8d463314aa7c96234e1be25a25245f704. From checkout: mkdir -p scratch && curl -LsSf https://raw.githubusercontent.com/gtmcknight/amm-sim/main/research/setup.sh -o scratch/setup.sh && sh scratch/setup.sh; prefix later commands with `. scratch/env.sh && `. Setup failure: blocked={step,error}. \nFetch pilot JSON https://api.imdnotdead.fun/artifacts/4528695b3a181b79340bf6e476b94d7f8ec2ee0a36dd9cd896416026cb6a7ccf. Its evidence.contracts URLs provide T1, M09 and IB; decode their code_zb64 (base64/zlib/tar). T1=t1/Strategy.sol; M09=m09-2/Strategy.sol; IB=IB.sol in audit bundle. Frozen source/runtime hashes are in pilot.hashes.arms; require exact matches. Use pilot.runner.code_zb64 (base64/zlib/tar, uncompressed-tar SHA256 in runner.code_sha256): pilot/canonical.py and summarize.py, adapted only for allocated range. Compiler solc0.8.24/viaIR/optimizer200/paris. Use setup released wheel SHA256 aedb68280aca34c110968f00148387c95885207913fb2cf19a1ab4c5cfd6887a; different engine provenance is a separate, unqualified cohort. Configs must equal MatchRunner._build_configs: build_config/default_rng(absolute seed), draw size/rate/sigma in that order. Check builder equality without simulating old seeds. Run amm_sim_rs.run_batch(X,VanillaStrategy,configs,8) separately for each arm on identical configs, four 64-seed commands per arm, each <=470s. Before each command reset gas_stats; immediately after save gas_stats for both roles. Require calls==ok and all fault counters zero. Keep started markers/checkpoints under scratch; never rerun an initiated arm/seed, pad beyond allocation or fill missing rows. Slow seats return completed pairs and explicit missing ranges, never block for speed. Validate all three sources with amm-match validate. valid=false on any OOG, revert, other callback fault, afterSwap peak>245000, failed validation or identity/config mismatch. Return JSON <60000: allocation, missing, arms.{t1,m09,ib}.{n,sum,sumsq,rows:[seed,edge]}, paired.{t1_minus_ib,t1_minus_m09,m09_minus_ib}.{n,mean_diff,se,sum,sumsq,worst5}, per-command live gas_stats and arm peaks/means, validate, valid, source/runtime/wheel hashes, config checks/hash encoding, runner code_zb64+tar SHA256, limitations. Use unrounded paired differences on common seeds, sample SE. No raw command output.",
      "items": [
        {
          "key": "eng-12011136",
          "vars": {
            "from": "12011136",
            "to": "12011391"
          }
        },
        {
          "key": "eng-12011392",
          "vars": {
            "from": "12011392",
            "to": "12011647"
          }
        },
        {
          "key": "eng-12011648",
          "vars": {
            "from": "12011648",
            "to": "12011903"
          }
        },
        {
          "key": "eng-12011904",
          "vars": {
            "from": "12011904",
            "to": "12012159"
          }
        }
      ],
      "output": {
        "format": "json",
        "maxBytes": 60000
      },
      "requires": {
        "platform": [
          "linux-x64"
        ]
      },
      "after": {
        "batch": "a10-t1-audit",
        "pairing": "batch",
        "when": "accepted",
        "check": {
          "path": "gates.pass",
          "equals": true
        }
      }
    },
    {
      "name": "a10-confirm-t1-c",
      "mode": "research",
      "brief": "Track A: one confirmation block of frozen T1, M09 and IB, each ALONE versus VanillaStrategy on exclusive fresh engine seeds {{from}}..{{to}} (256). No tuning, source edits, board entry or other seeds. This follows both audits in {{after.batch}}. Clone https://github.com/benedictbrady/amm-challenge at 6b6921f8d463314aa7c96234e1be25a25245f704. From checkout: mkdir -p scratch && curl -LsSf https://raw.githubusercontent.com/gtmcknight/amm-sim/main/research/setup.sh -o scratch/setup.sh && sh scratch/setup.sh; prefix later commands with `. scratch/env.sh && `. Setup failure: blocked={step,error}. \nFetch pilot JSON https://api.imdnotdead.fun/artifacts/4528695b3a181b79340bf6e476b94d7f8ec2ee0a36dd9cd896416026cb6a7ccf. Its evidence.contracts URLs provide T1, M09 and IB; decode their code_zb64 (base64/zlib/tar). T1=t1/Strategy.sol; M09=m09-2/Strategy.sol; IB=IB.sol in audit bundle. Frozen source/runtime hashes are in pilot.hashes.arms; require exact matches. Use pilot.runner.code_zb64 (base64/zlib/tar, uncompressed-tar SHA256 in runner.code_sha256): pilot/canonical.py and summarize.py, adapted only for allocated range. Compiler solc0.8.24/viaIR/optimizer200/paris. Use setup released wheel SHA256 aedb68280aca34c110968f00148387c95885207913fb2cf19a1ab4c5cfd6887a; different engine provenance is a separate, unqualified cohort. Configs must equal MatchRunner._build_configs: build_config/default_rng(absolute seed), draw size/rate/sigma in that order. Check builder equality without simulating old seeds. Run amm_sim_rs.run_batch(X,VanillaStrategy,configs,8) separately for each arm on identical configs, four 64-seed commands per arm, each <=470s. Before each command reset gas_stats; immediately after save gas_stats for both roles. Require calls==ok and all fault counters zero. Keep started markers/checkpoints under scratch; never rerun an initiated arm/seed, pad beyond allocation or fill missing rows. Slow seats return completed pairs and explicit missing ranges, never block for speed. Validate all three sources with amm-match validate. valid=false on any OOG, revert, other callback fault, afterSwap peak>245000, failed validation or identity/config mismatch. Return JSON <60000: allocation, missing, arms.{t1,m09,ib}.{n,sum,sumsq,rows:[seed,edge]}, paired.{t1_minus_ib,t1_minus_m09,m09_minus_ib}.{n,mean_diff,se,sum,sumsq,worst5}, per-command live gas_stats and arm peaks/means, validate, valid, source/runtime/wheel hashes, config checks/hash encoding, runner code_zb64+tar SHA256, limitations. Use unrounded paired differences on common seeds, sample SE. No raw command output.",
      "items": [
        {
          "key": "eng-12012160",
          "vars": {
            "from": "12012160",
            "to": "12012415"
          }
        },
        {
          "key": "eng-12012416",
          "vars": {
            "from": "12012416",
            "to": "12012671"
          }
        },
        {
          "key": "eng-12012672",
          "vars": {
            "from": "12012672",
            "to": "12012927"
          }
        },
        {
          "key": "eng-12012928",
          "vars": {
            "from": "12012928",
            "to": "12013183"
          }
        }
      ],
      "output": {
        "format": "json",
        "maxBytes": 60000
      },
      "requires": {
        "platform": [
          "linux-x64"
        ]
      },
      "after": {
        "batch": "a10-t1-audit",
        "pairing": "batch",
        "when": "accepted",
        "check": {
          "path": "gates.pass",
          "equals": true
        }
      }
    },
    {
      "name": "a10-confirm-t1-d",
      "mode": "research",
      "brief": "Track A: one confirmation block of frozen T1, M09 and IB, each ALONE versus VanillaStrategy on exclusive fresh engine seeds {{from}}..{{to}} (256). No tuning, source edits, board entry or other seeds. This follows both audits in {{after.batch}}. Clone https://github.com/benedictbrady/amm-challenge at 6b6921f8d463314aa7c96234e1be25a25245f704. From checkout: mkdir -p scratch && curl -LsSf https://raw.githubusercontent.com/gtmcknight/amm-sim/main/research/setup.sh -o scratch/setup.sh && sh scratch/setup.sh; prefix later commands with `. scratch/env.sh && `. Setup failure: blocked={step,error}. \nFetch pilot JSON https://api.imdnotdead.fun/artifacts/4528695b3a181b79340bf6e476b94d7f8ec2ee0a36dd9cd896416026cb6a7ccf. Its evidence.contracts URLs provide T1, M09 and IB; decode their code_zb64 (base64/zlib/tar). T1=t1/Strategy.sol; M09=m09-2/Strategy.sol; IB=IB.sol in audit bundle. Frozen source/runtime hashes are in pilot.hashes.arms; require exact matches. Use pilot.runner.code_zb64 (base64/zlib/tar, uncompressed-tar SHA256 in runner.code_sha256): pilot/canonical.py and summarize.py, adapted only for allocated range. Compiler solc0.8.24/viaIR/optimizer200/paris. Use setup released wheel SHA256 aedb68280aca34c110968f00148387c95885207913fb2cf19a1ab4c5cfd6887a; different engine provenance is a separate, unqualified cohort. Configs must equal MatchRunner._build_configs: build_config/default_rng(absolute seed), draw size/rate/sigma in that order. Check builder equality without simulating old seeds. Run amm_sim_rs.run_batch(X,VanillaStrategy,configs,8) separately for each arm on identical configs, four 64-seed commands per arm, each <=470s. Before each command reset gas_stats; immediately after save gas_stats for both roles. Require calls==ok and all fault counters zero. Keep started markers/checkpoints under scratch; never rerun an initiated arm/seed, pad beyond allocation or fill missing rows. Slow seats return completed pairs and explicit missing ranges, never block for speed. Validate all three sources with amm-match validate. valid=false on any OOG, revert, other callback fault, afterSwap peak>245000, failed validation or identity/config mismatch. Return JSON <60000: allocation, missing, arms.{t1,m09,ib}.{n,sum,sumsq,rows:[seed,edge]}, paired.{t1_minus_ib,t1_minus_m09,m09_minus_ib}.{n,mean_diff,se,sum,sumsq,worst5}, per-command live gas_stats and arm peaks/means, validate, valid, source/runtime/wheel hashes, config checks/hash encoding, runner code_zb64+tar SHA256, limitations. Use unrounded paired differences on common seeds, sample SE. No raw command output.",
      "items": [
        {
          "key": "eng-12013184",
          "vars": {
            "from": "12013184",
            "to": "12013439"
          }
        },
        {
          "key": "eng-12013440",
          "vars": {
            "from": "12013440",
            "to": "12013695"
          }
        },
        {
          "key": "eng-12013696",
          "vars": {
            "from": "12013696",
            "to": "12013951"
          }
        },
        {
          "key": "eng-12013952",
          "vars": {
            "from": "12013952",
            "to": "12014207"
          }
        }
      ],
      "output": {
        "format": "json",
        "maxBytes": 60000
      },
      "requires": {
        "platform": [
          "linux-x64"
        ]
      },
      "after": {
        "batch": "a10-t1-audit",
        "pairing": "batch",
        "when": "accepted",
        "check": {
          "path": "gates.pass",
          "equals": true
        }
      }
    },
    {
      "name": "a10-tell-compress",
      "mode": "research",
      "brief": "Track A compression only, independent copy {{copy}}, Python; no engine or board entry. Fetch B=https://raw.githubusercontent.com/gtmcknight/amm-sim/e65314f88de945f6d055bfc2aab01c482ab6570e/research/solved: curl -LsSf $B/fetch.sh -o fetch.sh && sh fetch.sh $B scratch/solved. \nFrozen T1 source/twins at https://api.imdnotdead.fun/artifacts/6b308b7219a0cf0bab1e5168e2f2b94b68bc6c118ae09a6d5ca30ea0173638ff, code_zb64=base64/zlib/tar (code_sha256 hashes base64 text). Use t1/twin.py, expand.py and int_model.py as data/code references; preserve detector, filter, mixture settings and fee rounding. One mechanism: redistribute the compressed qm table knots/levels/precision for the state distribution AFTER the arb tell. Measure the existing compressed table versus the full qm_k1e-5.json under the SAME tell/filter before searching; do not attribute filter differences to compression. Use training-state occupancy to select at most four layouts, keeping roughly the existing 14688 packed table bytes and a feasible 24576-byte runtime. Prioritize sd=0 and low-sd accuracy without discarding rare tails; preserve mirrored quotes and phase/rival behavior. No detector or sigma changes. Tune 67000..67255, freeze, confirm 67256..67511; once on holdout 10000..12047 use unchanged run.py/pair.py, reporting candidate versus original compressed T1 twin, same-filter full-table twin and ref/lattice_qm.json. Policy gets only callback inputs; any truth used in offline diagnostics stays outside it. Check unchanged-layout equality and table reconstruction at knots/boundaries. Commands <=470s, partials/missing explicit, no speed gate. Return JSON <60000: layouts/selection, packed bytes, decomposed compression loss, paired{n,mean_diff,se,sum,sumsq,worst5}, coverage, misread/price RMS, checks, estimated gas/byte costs and limitations; code_zb64+uncompressed-tar SHA256 with compact table data, deterministic reconstruction/generator and scoring adapters. No 8MB expanded table; no claim of Solidity parity without an actual EVM check. This item delivers one measured table layout for a later port.",
      "items": [
        {
          "key": "compress-1",
          "vars": {
            "copy": "1"
          }
        },
        {
          "key": "compress-2",
          "vars": {
            "copy": "2"
          }
        }
      ],
      "output": {
        "format": "json",
        "maxBytes": 60000
      }
    },
    {
      "name": "b10-tell-sigma",
      "mode": "research",
      "brief": "Track B, one blind Python policy, independent copy {{copy}}. No engine or board entry. Fetch B=https://raw.githubusercontent.com/gtmcknight/amm-sim/e65314f88de945f6d055bfc2aab01c482ab6570e/research/solved: curl -LsSf $B/fetch.sh -o fetch.sh && sh fetch.sh $B scratch/solved. \nBase: TellGauss in code_zb64 at https://api.imdnotdead.fun/artifacts/291066f0057c627e719fb643ab7f553f8efe3a0e5e323ec02ae5096be86d028c, base64/zlib/tar; tellpol.py SHA256 6d65cd387bbf2eef8eb48a695cf7e977777080e81e550abb0fcd71209186f0ee. Keep C=1, soft=1, qm_k1e-5.json and the likelihood r, never hard-pin on a tell. \nOne mechanism: estimate per-run price diffusion sigma causally from tell-inferred pins and silent no-arb intervals, then use it only in price prediction/cutting. Account for threshold-selected pin times and elapsed steps; naive squared pin increments are selection-biased. No hidden sigma, order-rate adaptation, quote changes or table retraining. Shrink toward 9.45 bps and bound to the model range; use soft pin confidence to limit false-pin contamination. Previous per-run adaptation before the tell added only +0.022; test whether nearly exact pins change that result. Fixed-sigma mode must equal the base.\nUse unchanged run.py/pair.py. Policy receives only six afterSwap inputs and its own state; no seed, truth, future or run parameters. Tune at most four settings on 67000..67255; freeze before confirmation 67256..67511. Score frozen candidate and identical base once per seed on 10000..12047, in chunks under 470s. Return actual partials/missing ranges if slow; never substitute baseline scores for an unfinished idea. Check disabled-change equality, replay and hidden-input scrambling. JSON <60000 bytes: mechanism, frozen settings, seed coverage, paired{n,mean_diff,se,sum,sumsq,worst5} for tune/confirm/holdout versus base and holdout versus ref/lattice_qm.json and ref/flag_qm.json; per-arm misread counts/rates, price RMS, measured or explicitly estimated gas/state cost, checks, limitations, code_zb64=base64(zlib(tar)) with uncompressed-tar SHA256, policy source hash and reproduction scripts. Keep tables out unless compact; include deterministic regeneration. Only incremental gain over TellGauss is new value; duplicates do not add sample size.",
      "items": [
        {
          "key": "sigma-1",
          "vars": {
            "copy": "1"
          }
        },
        {
          "key": "sigma-2",
          "vars": {
            "copy": "2"
          }
        }
      ],
      "output": {
        "format": "json",
        "maxBytes": 60000
      }
    },
    {
      "name": "b10-tell-rival",
      "mode": "research",
      "brief": "Track B, one blind Python policy, independent copy {{copy}}. No engine or board entry. Fetch B=https://raw.githubusercontent.com/gtmcknight/amm-sim/e65314f88de945f6d055bfc2aab01c482ab6570e/research/solved: curl -LsSf $B/fetch.sh -o fetch.sh && sh fetch.sh $B scratch/solved. \nBase: TellGauss in code_zb64 at https://api.imdnotdead.fun/artifacts/291066f0057c627e719fb643ab7f553f8efe3a0e5e323ec02ae5096be86d028c, base64/zlib/tar; tellpol.py SHA256 6d65cd387bbf2eef8eb48a695cf7e977777080e81e550abb0fcd71209186f0ee. Keep C=1, soft=1, qm_k1e-5.json and the likelihood r, never hard-pin on a tell. \nOne mechanism: propagate the rival through unobserved rival-only retail orders, conditioned on no own fill, between observed callbacks; then use that rival estimate in existing quotes. The mass-corrected kernel is in https://api.imdnotdead.fun/artifacts/4a0e539935a07e8ab8dddd427095efc5b0b34ee8bd252b66f841ad8c331b4737 (code_zb64 base64/zlib/tar, rivalt.py/jointpol.py). Extract only its rival-only transition, preserving cumulative survivor mass across Poisson order counts. Do not import its old thr=.9 classifier or entire lattice. Keep TellGauss price filter, detector and tables fixed; bound quadrature and moment-project rival state for a plausible contract. Check finite enumeration and continuous Monte Carlo on tuning states; report quadrature, compression and count-truncation errors separately. Old repaired RivalT +0.215 over pre-tell thr09 is not an additive gain here.\nUse unchanged run.py/pair.py. Policy receives only six afterSwap inputs and its own state; no seed, truth, future or run parameters. Tune at most four settings on 67000..67255; freeze before confirmation 67256..67511. Score frozen candidate and identical base once per seed on 10000..12047, in chunks under 470s. Return actual partials/missing ranges if slow; never substitute baseline scores for an unfinished idea. Check disabled-change equality, replay and hidden-input scrambling. JSON <60000 bytes: mechanism, frozen settings, seed coverage, paired{n,mean_diff,se,sum,sumsq,worst5} for tune/confirm/holdout versus base and holdout versus ref/lattice_qm.json and ref/flag_qm.json; per-arm misread counts/rates, price RMS, measured or explicitly estimated gas/state cost, checks, limitations, code_zb64=base64(zlib(tar)) with uncompressed-tar SHA256, policy source hash and reproduction scripts. Keep tables out unless compact; include deterministic regeneration. Only incremental gain over TellGauss is new value; duplicates do not add sample size.",
      "items": [
        {
          "key": "rival-1",
          "vars": {
            "copy": "1"
          }
        },
        {
          "key": "rival-2",
          "vars": {
            "copy": "2"
          }
        }
      ],
      "output": {
        "format": "json",
        "maxBytes": 60000
      }
    },
    {
      "name": "b10-finer-table",
      "mode": "research",
      "brief": "Track B, one blind Python policy, independent copy {{copy}}. No engine or board entry. Fetch B=https://raw.githubusercontent.com/gtmcknight/amm-sim/e65314f88de945f6d055bfc2aab01c482ab6570e/research/solved: curl -LsSf $B/fetch.sh -o fetch.sh && sh fetch.sh $B scratch/solved. \nBase: TellGauss in code_zb64 at https://api.imdnotdead.fun/artifacts/291066f0057c627e719fb643ab7f553f8efe3a0e5e323ec02ae5096be86d028c, base64/zlib/tar; tellpol.py SHA256 6d65cd387bbf2eef8eb48a695cf7e977777080e81e550abb0fcd71209186f0ee. Keep C=1, soft=1, qm_k1e-5.json and the likelihood r, never hard-pin on a tell. \nOne mechanism: solve the policy tables on a 2.5-bps state/action lattice instead of 5 bps, all physical ranges, price/order laws and kappa=1e-5 fixed. This refines DP policy geometry, not the belief lattice. First reproduce the original 5-bps solve and qm cells; then consistently rescale displacement, rival, action bounds and kernels. Use dp.py DP_STATE checkpoints and short DP_SWEEPS chunks, DP_SAVEVN=1 then qmdp.py. Check convergence and zero-change table equivalence before scoring; no unconverged/fallback holdout. Report whether value iteration, quadrature or interpolation changes the result. A wider-grid memory failure or incomplete solve returns useful progress/scripts, never invented scores. Keep the reference filter and tell fixed; supply an adapter for finer tables.\nUse unchanged run.py/pair.py. Policy receives only six afterSwap inputs and its own state; no seed, truth, future or run parameters. Tune at most four settings on 67000..67255; freeze before confirmation 67256..67511. Score frozen candidate and identical base once per seed on 10000..12047, in chunks under 470s. Return actual partials/missing ranges if slow; never substitute baseline scores for an unfinished idea. Check disabled-change equality, replay and hidden-input scrambling. JSON <60000 bytes: mechanism, frozen settings, seed coverage, paired{n,mean_diff,se,sum,sumsq,worst5} for tune/confirm/holdout versus base and holdout versus ref/lattice_qm.json and ref/flag_qm.json; per-arm misread counts/rates, price RMS, measured or explicitly estimated gas/state cost, checks, limitations, code_zb64=base64(zlib(tar)) with uncompressed-tar SHA256, policy source hash and reproduction scripts. Keep tables out unless compact; include deterministic regeneration. Only incremental gain over TellGauss is new value; duplicates do not add sample size.",
      "items": [
        {
          "key": "grid-1",
          "vars": {
            "copy": "1"
          }
        }
      ],
      "output": {
        "format": "json",
        "maxBytes": 60000
      },
      "requires": {
        "minCpus": 4,
        "minMemoryGb": 8
      }
    },
    {
      "name": "b10-fresh-tell",
      "mode": "research",
      "brief": "Track B, one blind Python policy, independent copy {{copy}}. No engine or board entry. Fetch B=https://raw.githubusercontent.com/gtmcknight/amm-sim/e65314f88de945f6d055bfc2aab01c482ab6570e/research/solved: curl -LsSf $B/fetch.sh -o fetch.sh && sh fetch.sh $B scratch/solved. \nBase: TellGauss in code_zb64 at https://api.imdnotdead.fun/artifacts/291066f0057c627e719fb643ab7f553f8efe3a0e5e323ec02ae5096be86d028c, base64/zlib/tar; tellpol.py SHA256 6d65cd387bbf2eef8eb48a695cf7e977777080e81e550abb0fcd71209186f0ee. Keep C=1, soft=1, qm_k1e-5.json and the likelihood r, never hard-pin on a tell. \nStart from the question alone: what causal information or decision rule can improve edge after the arb/retail flag is almost known? Skip the README idea lists. Pick ONE new mechanism and explain its observable basis before testing. Exclude threshold/temperature/branch-blend sweeps, larger kappa/probing, posterior-value shortlists, empirical fill-law or no-fill reweighting, finer belief lattices, and exact-float tell refinements (remaining detector gap only about .026). Also exclude this wave's sigma learning, rival-only transport, table compression and finer DP grid. Preserve the base for comparison; negative findings are useful. Do not combine ideas or search board seeds.\nUse unchanged run.py/pair.py. Policy receives only six afterSwap inputs and its own state; no seed, truth, future or run parameters. Tune at most four settings on 67000..67255; freeze before confirmation 67256..67511. Score frozen candidate and identical base once per seed on 10000..12047, in chunks under 470s. Return actual partials/missing ranges if slow; never substitute baseline scores for an unfinished idea. Check disabled-change equality, replay and hidden-input scrambling. JSON <60000 bytes: mechanism, frozen settings, seed coverage, paired{n,mean_diff,se,sum,sumsq,worst5} for tune/confirm/holdout versus base and holdout versus ref/lattice_qm.json and ref/flag_qm.json; per-arm misread counts/rates, price RMS, measured or explicitly estimated gas/state cost, checks, limitations, code_zb64=base64(zlib(tar)) with uncompressed-tar SHA256, policy source hash and reproduction scripts. Keep tables out unless compact; include deterministic regeneration. Only incremental gain over TellGauss is new value; duplicates do not add sample size.",
      "items": [
        {
          "key": "fresh-1",
          "vars": {
            "copy": "1"
          }
        },
        {
          "key": "fresh-2",
          "vars": {
            "copy": "2"
          }
        }
      ],
      "output": {
        "format": "json",
        "maxBytes": 60000
      }
    }
  ],
  "read": [
    "a10-t1-audit/audit-1",
    "a10-t1-audit/audit-2",
    "a10-confirm-t1-a/eng-12010112",
    "a10-confirm-t1-a/eng-12010368",
    "a10-confirm-t1-a/eng-12010624",
    "a10-confirm-t1-a/eng-12010880",
    "a10-confirm-t1-b/eng-12011136",
    "a10-confirm-t1-b/eng-12011392",
    "a10-confirm-t1-b/eng-12011648",
    "a10-confirm-t1-b/eng-12011904",
    "a10-confirm-t1-c/eng-12012160",
    "a10-confirm-t1-c/eng-12012416",
    "a10-confirm-t1-c/eng-12012672",
    "a10-confirm-t1-c/eng-12012928",
    "a10-confirm-t1-d/eng-12013184",
    "a10-confirm-t1-d/eng-12013440",
    "a10-confirm-t1-d/eng-12013696",
    "a10-confirm-t1-d/eng-12013952",
    "a10-tell-compress/compress-1",
    "a10-tell-compress/compress-2",
    "b10-tell-sigma/sigma-1",
    "b10-tell-sigma/sigma-2",
    "b10-tell-rival/rival-1",
    "b10-tell-rival/rival-2",
    "b10-finer-table/grid-1",
    "b10-fresh-tell/fresh-1",
    "b10-fresh-tell/fresh-2"
  ],
  "cancel": [
    "b8-ambiguity-bonus"
  ],
  "wake": {
    "batches": [
      "a10-confirm-t1-a",
      "a10-confirm-t1-b",
      "a10-confirm-t1-c",
      "a10-confirm-t1-d",
      "a10-tell-compress",
      "b10-tell-sigma",
      "b10-tell-rival",
      "b10-finer-table",
      "b10-fresh-tell"
    ],
    "afterMinutes": 90
  },
  "finish": false
}