# d-ideas — merge 02 (10 idea-sweep items)

Credit: Taylor (gtmcknight), github.com/gtmcknight/amm-sim@4812a6382311624e1506aca131ca35c378f930d9 research/solved. The replica, Gaussian filter, fee rule, qm tables and exact-arithmetic arb/retail tell that every item runs on and edits are Taylor's work.

Sources: the ten item results `artifacts/goal/batches/d-ideas/items/<item>/result.json` (SHA-256 values as given in the merge brief). Every number below is copied from those files. This merge adds no numbers of its own.

**Summary.** None of the 10 items is promoted, and none reaches the +0.03 bar. The largest `on_age` mean_diff is +0.004592 (tell-coverage), and that item's own verdict says it could just as well be a fee-level effect. Markout is clearly negative, at −0.037481 (se 0.003495). Every item is replica-only (fast LEAN replica, with LEAN and AGE baselines). None ran on the engine, ran a gate, was compared with QLX5 or qt1, or was submitted.

**Source caveat.** When their machines re-read their sources, these results cited sources that no longer held: sd-extrap (4 sources), second-order (1), sync-grade (2) and endgame (1). The result files list no separate sources field, so this merge cannot tell which claims depend on the failed sources. For those four items, give less weight to claims that rest on external pages (for example provenance and contract-structure statements). This does not change the conclusion: all four are non-promoted, near-zero results.

```json
{
 "credit": "Taylor (gtmcknight), github.com/gtmcknight/amm-sim@4812a6382311624e1506aca131ca35c378f930d9 research/solved",
 "rows": [
  {"item": "tell-coverage", "mech": "tell-coverage", "already_present": false,
   "chosen": {"name": "gamma_2^36_higher_fee", "parameters": {"mode": 1, "par": 36, "rnd": 1}},
   "report": {"vs_lean": {"n": "2048", "mean_diff": "0.006329", "se": "0.002453", "seeds_positive": "1076"},
              "on_age": {"n": "2048", "mean_diff": "0.004592", "se": "0.002451", "seeds_positive": "1094"}},
   "promote": false,
   "contract_cost": "estimate, not measured: storage_words 0; gas about 2,000 to 2,600 per afterSwap on every path (peak about 210,000 to 211,000 of 245,000); runtime_bytes about 250 to 400; characters about 600 to 900; fits true",
   "definition": "Snap each fee's engine gamma (1 - w/1e18 in binary64) to a coarser lattice of multiples of 2^par before the step off Vanilla's gamma, to change which trades the tell can classify.",
   "note": "Item verdict: tell coverage did not move; the gain could be a fee raise of up to 0.0763 bps; no control was run."},
  {"item": "lookahead", "mech": "lookahead: additive forecast variance used only for the fee-table lookup", "already_present": false,
   "chosen": {"h": 0.001},
   "report": {"vs_lean": {"n": "2048", "mean_diff": "0.003492132563304662", "se": "0.002146116474430133", "seeds_positive": "1029"},
              "on_age": {"n": "2048", "mean_diff": "0.0005059215832921293", "se": "0.0022894789176444662", "seeds_positive": "1044"}},
   "promote": false,
   "contract_cost": "not measured: storage_words 0; one constant addition inside the existing sqrt (about 16 characters at the expression site); no compiled gas or size claim",
   "definition": "The fee-table lookup uses sd = 10000*sqrt(v + h*sigma0^2) in place of sqrt(v), without changing the stored belief.",
   "note": "Item: h=0.001 is a nearly disabled forecast and is not evidence for useful lookahead."},
  {"item": "sd-extrap", "mech": "sd-extrap: what the qm lookup does outside its tables' range of offset and belief standard deviation (linear extrapolation in place of clipping, and a linear add-on per bps of sd beyond the last level)", "already_present": false,
   "chosen": {"name": "w_ext_uncapped", "parameters": {"sd_ext_bps": 0, "sd_add": 0, "w_ext_bps": 1000, "w2_ext_bps": 0}},
   "report": {"vs_lean": {"n": "2048", "mean_diff": "-0.000105", "se": "0.000258", "seeds_positive": "480"},
              "on_age": {"n": "2048", "mean_diff": "0.000152", "se": "0.000432", "seeds_positive": "493"}},
   "promote": false,
   "contract_cost": "estimate from reading the QLX5 source, not compiled: storage_words 0; chosen offset extrapolation about +50 to +150 gas per afterSwap (peak about 208,100..208,500 of 245,000)",
   "definition": "Extrapolate the qm table bilinear form linearly past its edges (sd, offset and rival-offset axes) in place of clipping.",
   "note": "Source caveat applies (4 sources did not hold on re-read). Item: sd passes 34 bps on about one callback in 138,000."},
  {"item": "endgame", "mech": "endgame: callback-timestamp-gated symmetric fee offsets at the beginning or end of a fixed 10,000-step run; tested independently at each end.", "already_present": false,
   "chosen": {"id": 3, "parameters": {"first_N": 0, "first_delta_bps": 0.0, "last_N": 10, "last_delta_bps": 2.0}},
   "report": {"vs_lean": {"n": "2048", "mean_diff": "0.00035885772081734046", "se": "0.0003227178901749649", "seeds_positive": "879"},
              "on_age": {"n": "2048", "mean_diff": "3.906800292433066e-05", "se": "0.00032759570186093344", "seeds_positive": "870"}},
   "promote": false,
   "contract_cost": "estimate: storage_words 0; additional peak gas 30..200; characters 200..500; runtime_bytes 40..200; not measured",
   "definition": "Add delta_bps to both fees in the first first_N or last last_N steps of the 10,000-step run.",
   "note": "Source caveat applies (1 source did not hold on re-read). The chosen schedule's selection mean was negative (-0.00044629770173748584)."},
  {"item": "drift", "mech": "drift: a per-step drift of the mean of ln p in the Gaussian filter's prediction step (LEAN keeps the mean and grows only the variance)", "already_present": false,
   "chosen": {"parameters": {"c": -16, "dmode": 0, "drift_ln_per_step": 7.1442e-06}},
   "report": {"vs_lean": {"n": "2048", "mean_diff": "0.003879", "se": "0.002265", "seeds_positive": "1014"},
              "on_age": {"n": "2048", "mean_diff": "-0.000171", "se": "0.002565", "seeds_positive": "1022"}},
   "promote": false,
   "contract_cost": "measured (solc 0.8.24, viaIR): storage_words 0; characters +34 (c=-16), +32 (c=1); runtime_bytes 22440 -> 22464 (c=-16), 22462 (c=1); gas an opcode estimate of the order of 100-200",
   "definition": "Shift the filter mean m by d = c*(-SIG2/2) per elapsed step in the prediction step, alongside the variance growth.",
   "note": "Item: the chosen c=-16 is a noise pick whose sign is opposite to the true drift; the exact drift (c=1) scored -0.0013 se 0.0027 on the choosing seeds."},
  {"item": "mixture-branch", "mech": "mixture-branch", "already_present": false,
   "chosen": {"id": "s01", "parameters": {"wU": 1.0, "wN": 1.0, "eps": 0.005, "own": 0.0, "gamma": 1.0}},
   "report": {"vs_lean": {"n": "2048", "mean_diff": "0.000526", "se": "0.000736", "seeds_positive": "1052"},
              "on_age": {"n": "2048", "mean_diff": "-0.000455", "se": "0.000733", "seeds_positive": "978"}},
   "promote": false,
   "contract_cost": "not measured: storage_words 0; one more qm table lookup, about 2 ln and 4 multiplications on a mixed callback (about 26 per run); the Python mechanism lines are 2548 characters",
   "definition": "On an undecided first trade (eps < arb posterior < 1-eps), return the posterior-weighted mix of the arb-branch and retail-branch fees in place of the hard branch.",
   "note": "Item: the over/unsupported half could not be exercised on the replica (1 in 14.4 million first trades)."},
  {"item": "second-order", "mech": "second-order: residual log-price-to-price center shift, implemented as opposite ask/bid fee adjustments proportional to posterior log-price variance", "already_present": false,
   "chosen": {"parameters": {"c": 1.0}},
   "report": {"vs_lean": {"n": "2048", "mean_diff": "0.004102156314634814", "se": "0.001829279029610601", "seeds_positive": "1021"},
              "on_age": {"n": "2048", "mean_diff": "-0.0005802840008916621", "se": "0.0018498706799454404", "seeds_positive": "986"}},
   "promote": false,
   "contract_cost": "estimate, not measured: storage_words 0; gas +50..1000 per callback; characters +100..400; runtime_bytes +40..250",
   "definition": "Lower the bid and raise the ask by c times the posterior variance of ln price, before the existing fee rounding.",
   "note": "Source caveat applies (1 source did not hold on re-read). already_present is given as {\"present\": false}."},
  {"item": "sync-grade", "mech": "sync-grade: weight of the synced fee table graded by the age (in steps) of the last exact reading of the rival, in place of the synced flag", "already_present": false,
   "chosen": {"l1": 1.0, "linf": 0.5, "gtau": 1.0},
   "report": {"vs_lean": {"n": "2048", "mean_diff": "0.003221", "se": "0.001723", "seeds_positive": "1059"},
              "on_age": {"n": "2048", "mean_diff": "-0.00349", "se": "0.003145", "seeds_positive": "1023"}},
   "promote": false,
   "contract_cost": "cost sketch on QLX5 compiled with solc 0.8.24 (unedited 22,440 runtime bytes); storage_words new 0 (bits 224..254 of slot 2); fee-rule gas measured with forge only, not an engine afterSwap peak",
   "definition": "Blend synced and unsynced table values with weight lam = linf + (l1-linf)*gtau/(u-1+gtau) by age u of the last exact rival reading.",
   "note": "Source caveat applies (2 sources did not hold on re-read)."},
  {"item": "side-imbalance", "mech": "side-imbalance: exponentially weighted running classified-retail sign imbalance, applied as an antisymmetric fee skew", "already_present": false,
   "chosen": {"index": 9, "parameters": {"memory": 256, "amplitude_bps": -0.5}},
   "report": {"vs_lean": {"n": "2048", "mean_diff": "-0.001620398002443485", "se": "0.0020789493224039174", "seeds_positive": "1013"},
              "on_age": {"n": "2048", "mean_diff": "-0.004243779899149924", "se": "0.0032861145878138203", "seeds_positive": "972"}},
   "promote": false,
   "contract_cost": "estimate, not measured: storage_words 1; gas +22500..24000 first nonzero update, +5500..8000 later retail callback, +2300..3500 arb callback; characters +500..1200; runtime_bytes +200..700",
   "definition": "An EMA I of retail trade sides (memory L) skews the fees: bid += A*I and ask -= A*I.",
   "note": ""},
  {"item": "markout", "mech": "markout: an online controller on realised markout; a per-side EMA of next-step markout (bps) raises or lowers that side's fee within the run", "already_present": false,
   "chosen": {"name": "cross_b005_k+0.03", "parameters": {"k": 0.03, "beta": 0.005, "mode": 2, "target_bps": 0.0, "cap_bps": 3.0, "fills": 1, "beta_slow": 0.0, "clip_bps": 50.0}},
   "report": {"vs_lean": {"n": "2048", "mean_diff": "-0.035103", "se": "0.003312", "seeds_positive": "812"},
              "on_age": {"n": "2048", "mean_diff": "-0.037481", "se": "0.003495", "seeds_positive": "812"}},
   "promote": false,
   "contract_cost": "estimate, not measured: storage_words 1 to 2 new words; gas about +7,000 to +14,000 on the peak callback; characters about +900 to +1,300; runtime_bytes about +350 to +600",
   "definition": "Value each side's fills at the filter mean after the next step, and feed a per-side EMA of that markout back into that side's fee.",
   "note": "Item: all 12 settings were dropped after the first 128 seeds (mean_diff + 3*se < 0)."}
 ],
 "failed": []
}
```

Rows are sorted by `report.on_age.mean_diff`, largest first.

## Failed / missing
None. All 10 items (lookahead, side-imbalance, markout, tell-coverage, sd-extrap, drift, second-order, mixture-branch, sync-grade, endgame) returned a result with a report.

## Where results agree or disagree
- Several items give the same replica baseline means: LEAN 536.603391 and AGE 536.6231 on seeds 10000..12047. The drift, markout and sync-grade items each report AGE minus LEAN on those seeds as +0.0197 (se 0.0041; written 0.019709 / 0.004066 in markout and sync-grade). This is consistent across items, but it is replica only, not engine.
- Lookahead, drift, second-order and sync-grade show small positive vs_lean values (about 1.6 to 2.2 se) that disappear or turn negative when stacked on AGE. None of these effects survives stacking.
- Every item's not_done lists the same gaps: no Solidity port measured on the engine, no gates P/Q/T/F/G/E, no QLX5 or qt1 pairing, and nothing submitted. The fast replica has no QLX5 rival split test, so none of these results was measured on QLX5 itself.

## Promotion
No item reaches >= +0.03, so nothing is promoted into a pipeline from this merge.