# Root merge 2 of 5: learned AMM fee policies

This merge covers all 25 supplied results. **No result establishes a top-10 leaderboard placement.** The strongest directly measured candidate-minus-HA mean in this group is IB continuation **c3: +0.7831934928537729, SE 0.032544514488015694, n=512**, with afterSwap peak **235028** and **0 OOG**. Its improvement over unchanged IB is **+0.20872903151392597, SE 0.022977540549694257**. These are reported paired real-engine measurements, not a leaderboard score; c3 also has three source-verification flags and supplies volume changes rather than the requested retail/arb edge decomposition. [b13-c3]

The task's historical baseline is **HA = T (RS2 plus a per-run toxicity anchor) plus a learned 13-8-2 fee correction**, **+1.257 se 0.062 over T on 512 fresh real-engine seeds**, **0 OutOfGas**, **afterSwap peak 231k**; **about +1.53 over RS2 (board 520.05, 27th)**. The goal says the top 10 needs **about +3.1 over HA**, and pilot **282fbc03** predicted real gains within **about 0.1**, with every training curve still rising at its time limit. These are requester-supplied history, not new measurements. HA's supplied code SHA-256 is `99e3a2eceae0e5543d27473859e10b005d0ce0aa7067f2e1a24d1e0e4a7390fe`, from the [HA artifact](https://api.imd.fun/artifacts/2801efc8190e86318dd922d4a7179f63684d77a226047fc2f3c78232963eb690). No current leaderboard was consulted.

## Evidence, scope, and confidence

All 25 local JSON inputs were present, parseable, nonempty, and matched their SHA-256 values supplied in the assignment. Embedded zlib/base64 payloads were decoded in memory and checked against companion hashes where present; embedded plain Solidity in fu-b/fu-c also matched its reported hash. These checks establish input integrity only: this merge did not rerun training, EVM parity, or the real engine. Source identifiers below refer to the supplied JSON, and field names identify the evidence within it. Selected narrative figures retain source precision; ledger tables explicitly round measurements.

The requester reports failed external source rechecks for 18 inputs (counts are recorded per item below). The supplied material does **not** identify the individual failed URLs/passages or distinguish unreadability from unsupported quotations. Their external support remains unresolved. Local experiment claims are retained as **worker-reported measurements**, with lower confidence than independently reproduced results; a hash match or a `valid:true` flag does not repair an unsupported citation. Unflagged inputs are not independently certified either. This particularly affects Gate B's historical reference gains and claims about pinned upstream APIs, not just incidental documentation. No missing external quotation has been reconstructed or silently treated as verified.

The required protocol is each policy alone against 30 bps VanillaStrategy, `amm_sim_rs.run_batch(X, VanillaStrategy, configs, 8)`, with MatchRunner configuration sampling, fresh confirmation seeds from 7000000, no OOG, and afterSwap peak <=245000. Gate A requires <0.01 bps on 50 sequences of 2000 callbacks; Gate B requires replica gain over T within 0.15 of the real reference; Gate C requires frozen port/Solidity agreement <0.1 bps. These are task requirements. The reported tests below are empirical evidence, not a proof for every future callback. Confirmation-only and speed-fallback items are distinguished from trained candidates.

The replica and Rust engine intentionally use different random paths. Identical seed numbers align configuration distributions and paired comparisons within an engine, not full cross-engine paths. Shared monitor 30000000..30002047 and holdout 40000000..40008191 recur across many searches; their repeated use across workers means these are not independent holdout replications. No cross-candidate pooled ranking is inferred from those shared blocks. [b07-wide-b notes; b07-wide-c notes; b09-kit-b notes; b10-inputs-cross notes]

## Findings that survive the merge

- **c3 leads the direct HA comparisons here.** Its replica gains are +0.19188324927575884 over IB and +0.844035295949161 over HA; real gains +0.20872903151392597 and +0.7831934928537729. c2 adds only +0.011245308912959162 over IB, SE 0.015367309340913873. Different frozen policies and workloads explain why these are not repeat estimates of one policy. [b13-c2; b13-c3]
- **Unchanged IB and FA advantages replicate across distinct real blocks.** Their individual results and pooled calculations follow. No same-seed IB-versus-FA comparison was supplied. [b12-ib-a; b13-c1; b13-c4; b12-fa-a; b12-fa-b]
- **Flow memory produces the largest r2 real increment over IA here**, +0.32168536156486005, mostly reduced arb loss. Cross-products, constants-only, fees, path, widening, gating and joint search also report positive means, but gated-c is uncertain with a -27.615352541581103 tail. These changes overlap; their gains cannot be added or relabelled as real HA comparisons. [b10-inputs-flow; b10-inputs-cross; b08-const-a; b10-inputs-fees; b10-inputs-path; b07-wide-a; b07-wide-b; b07-wide-c; b06-gated-b; b06-gated-c; b08-joint-a]
- **Diet packing frees gas without a measured fee-policy change.** IAD matches all 100 regression seeds and kit-b's 512 fresh seeds. Regression afterSwap peaks fall 237739 ->131825; fresh peaks 237738 ->132084. Wider candidates remain under 137000 gas, but wider is not automatically better across differing seeds/budgets. [b07-wide-a; b07-wide-b; b07-wide-c; b09-kit-b]
- **FA union is mixed:** fu-b gains +0.11209640416092681 over FA but omits edge decomposition; fu-c has a negative inconclusive mean and incomplete Gate C; fu-a never trained. [b11-fu-a; b11-fu-b; b11-fu-c]

Derived aggregation, not new runs: pooling the two FA-minus-HA confirmations by their unrounded `n,sum,sumsq` gives **n=2048, mean=0.5720108377251768, SE=0.015742734417236053**. Pooling unchanged IB-minus-HA from ib-a, c1, c4 gives **n=4128, mean=0.6444734258661725, SE=0.011308620754186574**. For c1/c4, sums of squares were reconstructed as `n*(n-1)*SE^2+n*mean^2`; pooled SE is `sqrt((Q-S^2/N)/(N-1)/N)`. These blocks are disjoint as reported. Do not include c2/c3's changed candidates in the unchanged-IB pool or count identity comparisons as extra independent IB runs. Pooling does not remove any source-verification caveat. [b12-fa-a x_vs_ha; b12-fa-b x_vs_ha; b12-ib-a x_vs_ha; b13-c1 real.vs_HA; b13-c4 real.vs_HA]

## Disagreements, omissions, and qualifications

**Frozen IA versus optimizer restart is a genuine policy mismatch.** joint-b measured maximum fee discrepancy 94.2068099712106 bps and mean 1.201640429565419 bps when comparing IA code to restart weights; 64.38380809595202% of fee comparisons exceed 0.01 bps. The first callback differs by 0.40874441118700366 bps; the largest weight difference is 0.44391. Frozen IA params pass Gate A at 0.00006501554232485329 bps. joint-b therefore stopped rather than choose between incompatible initialization requirements. wide-b/c and const-a chose frozen params. joint-a and inputs-path/flow/fees/cross started from the optimizer restart while checking equality against frozen params separately. Zero new feature weights preserve whichever old weights are used; they do not make the restart equal to frozen IA. On the common monitor, restart is -0.013212781858070094 versus IA, SE 0.014527516303598702. These results remain useful experiments, but their initial-policy gate does not certify restart equality. [b08-joint-b gates/notes; b07-wide-b notes; b07-wide-c notes; b08-const-a notes; b08-joint-a notes; b10-inputs-path initial_restart_vs_IA; b10-inputs-flow search; b10-inputs-fees notes; b10-inputs-cross notes]

**gated-b mislabels a comparator.** Its holdout `vs_HA` and `vs_IA` are numerically identical (+0.08090535160872292, SE 0.004432768631605314). The real section also labels the IA arm as `HA` in gas output, while its notes say candidate, IA, and T were tested. Preserve those raw labels in the ledger but do not present this as an independently measured HA comparison. Other reports establish distinct HA and IA holdout baselines (+1.1168320624999497 and +1.4817112289868306 over T). A corrected gated-b HA comparison was not supplied. [b06-gated-b holdout/real/notes; b10-inputs-flow holdout]

**Volume is not edge.** fu-b leaves real retail/arb edge deltas null and supplies volumes. c1/c2/c3/c4 call fields `retail_delta`/`arb_delta` but explicitly define them as Y-volume changes. c1 volumes cover only 2560 of its 2592 pairs because the first chunk retained edge only. Do not add these volume deltas to edge or compare them with the reconstructed edge decomposition used by the kit-based r2 and r4-confirm reports. The wheel's absence of a direct edge split does not prove reconstruction impossible: those kit reports reconstruct arb edge, set retail=edge-arb, and cross-check arb volume. The missing requested edge splits are reporting limitations even where `valid:true` is claimed. [b11-fu-b real; b13-c1 real/notes; b13-c2 real; b13-c3 real; b13-c4 real; b12-ib-a notes]

**Validity is narrower than full brief compliance.** Reported successful real candidates have zero OOG and peaks within 245000, but several omit edge splits or parity tests. fu-c is invalid because its frozen Gate C was not completed. c4 is an unchanged-code speed fallback with A/B/C not run, not a new gated training success. ib-b and fu-a are setup failures; inputs-clock is unmeasured, not a losing policy. The three confirmation-only workers did not rerun replica gates. [b11-fu-c; b13-c4; b12-ib-b; b11-fu-a; b10-inputs-clock; b12-ib-a; b12-fa-a; b12-fa-b]

**Timing and seed qualifications remain.** wide-a discloses a command that ran 9m50s before being killed, exceeding the task's 8-minute rule; completed chunks were retained and one interrupted candidate chunk rerun. gated-c reran a chunk killed at the 8-minute limit, reporting no result read from the killed attempt. These are reported confirmation retries, so a strict no-reuse interpretation is not fully satisfied even though no confirmation-dependent selection was reported. inputs-flow resumed from ten-generation checkpoints, sometimes repeating training generations; its NumPy RNG state was saved only from generation 80, so full from-scratch reproducibility is not established. inputs-path briefly overlapped adjacent segments, later added a lock, and checked retained checkpoint order. inputs-fees samples random training seeds from a finite interval; that rule alone does not guarantee uniqueness. No fresh-confirmation violation is inferred from training seed reuse. [b07-wide-a notes; b06-gated-c notes; b10-inputs-flow notes; b10-inputs-path notes; b10-inputs-fees search]

**Do not extrapolate every curve as rising.** gated-c plateaued after generation 140 through 224. wide-c's best checkpoint was generation 500 although training reached 582. inputs-flow's best was generation 660 although its note calls the curve still rising at 720; reported final monitored scores remain below the best. c2's best was generation 43 and final monitoring remained below it. c3 rose substantially overall but fluctuated late. These observations weaken blanket promises that longer training must improve every run. fu-b's statement that the incremental gain is “within noise on the real engine” conflicts with its own +0.11209640416092681, SE 0.017097784125455272 (about 6.6 SE); keep the measured result, but nonzero learned feature weights alone do not establish causal feature importance. fu-c's search_minutes=72.98397571245829 conflicts with a curve ending at 37.92990431388219 and notes saying 38+ minutes; timing cannot be reconciled from the supplied fields. [b06-gated-c search; b07-wide-c freeze; b10-inputs-flow search/notes; b13-c2 search/curve; b13-c3 curve; b11-fu-b notes/real; b11-fu-c search_minutes/curve/notes]

**HA's historical +1.257 is not a fixed offset.** Later HA-minus-T confirmations are +1.1517558455576902, +1.137622278143045, and +1.0966422066032975 on different real blocks; the shared replica holdout is +1.1168320624999497. Keep each baseline with its seeds and SE. This variation does not by itself identify a changed policy or a calibration failure, and adding the historical +1.257 to a new candidate comparison would not be a paired measurement. [b12-ib-a ha_vs_t; b12-fa-a ha_vs_t; b12-fa-b ha_vs_t; b10-inputs-flow holdout]

## Per-result interpretation

The ledger supplies paired means, SEs, worst-five tails, units, seeds and gas, rounded as labelled. Every item is retained, including failures. Source flags apply to the whole associated report, not to an invented subset of URLs.

### b06-gated-b

Two firstInStep-gated 19-8-2 heads; 380-value frozen layout includes independent six-input affine columns per head, while training tied those columns and duplicates them at freeze. Diagonal CMA, population 16, seed 20260312, 128 markets/generation: 105 generations in 41.24761648357768 minutes at 452.5069537168777 markets/s. Common monitor gain over the IA comparator rises from 0.015818146685694673 at generation 10 to 0.07755447976215876 at 100, with an early negative score at 20. Nine kit tests passed. Gate B was repeated in a clean Numba cache after a stale port-cache result. See comparator-label warning above. [b06-gated-b]


### b06-gated-c

Two gated 19-8-2 heads, 380 parameters. Both-nets=IA starts at monitor +1.5395660516422922 over T, while GATED nets plus IA new columns start at +0.7944757663286321; choose both IA. Full-covariance CMA, population 16, seed 20260313, sigma0=0.06, 192 markets/generation: 224 generations/60.79 minutes, gated speed 185.73885074483644 markets/s (T 309.14338176653894). Frozen generation 140 wins at +1.6446695255028119 over T (+0.10510347386051966 over IA), then plateaus. The 16-market pretraining probe peaked at 232454; frozen real peak 235526. [b06-gated-c]


### b07-wide-a

Diet shell with a 19-12-2 head: IA's eight units retained, four new W1 rows sampled normal(0,0.1), new b1/W2 zero; the initial replica equals IA on four checked markets. Partial sums are combined before final division. Population 16, seed 20260321, 256 markets/generation, 144 generations/40.31787060499191 minutes; IA speed 285.6728865566038 markets/s (kit 357.16699075062206). Freeze generation 140 ES mean, monitor +0.1590816760466797 over IA; still rising by source account. Diet regression is exact on seeds 0..99. Retain the 9m50s command-limit violation. [b07-wide-a]


### b07-wide-b

Diet 19-16-2, 354 head weights; eight new W1 rows normal(0,0.1), new b1/W2 zero, frozen IA params as starting eight units. Full CMA, population 16, seed 20260322, 128 markets/generation: 130 generations/65.06669605181474 minutes. Speed kit 249.48173166666118, IA 162.69934032191028, wide 71.71243941756586 markets/s. Best monitor +1.627939829461402 over T versus initial +1.5395660516422922. Exact 100-seed IAD regression; no claim of resumed upstream optimizer covariance. [b07-wide-b]


### b07-wide-c

Diet 19-16-2, 354 parameters, seed 20260323, population 16, 128 markets/generation; 582 generations/65.10815574514757 minutes. Expanded IA speed 351.75746652437533 markets/s; HA kit 511.1594540616374. Zero new output weights yield exact IA EVM fees on 100000 swaps; nonzero 16-unit code compiles to 17197 bytes. Freeze best generation 500, monitor +0.1345392700268226 over IA. Its sampled 20-row curve represents 58 monitoring rows, not 20 total checks. T parity 1.2216616607219066e-09 bps and kit HA parity 7.60422131512728e-05 bps also reported. No function evaluates more than eight hidden units; final signed division follows combined sums. [b07-wide-c]


### b08-joint-a

Joint head/constants/anchor search, 178 head weights plus 30 continuous T constants and two anchor coefficients; positive constants in log scale, cap/count/zero/integer/minimum-gate settings fixed. Population 16, seed 20260331, 128 markets/generation; IA speed 745.5614067104533 (kit HA 874.1036620560881) markets/s. Total 48.4815793065005 minutes includes 5.307997741278571 minutes/105 generations constants-only warm start (best monitor +0.036089439790771394 over IA) and 43.17358156522193 minutes/832 joint generations. Restarts at 318 and 629, 80 IA injections at 421..500. Best monitor +0.040119171898595596, frozen generation 640. Real retail/arb reconstruction max error 0.008310770215757657, larger than most other kit reports. Frozen constants retained below. [b08-joint-a]


### b08-joint-b

Blocked at initial_policy_equivalence. No search, frozen candidate, holdout, real engine, or candidate gas. Speed 491.95793812193494 markets/s and frozen-IA gates A/B succeeded, but the requested restart fails initial fee equality. Planned population 16, seed 20260332, 178 head weights plus log-scale constants. Initial benchmark had a stale Numba cache/ZeroDivisionError; local cache clearing fixed it. The result asks whether restart or frozen params should govern; this merge records that historical unresolved choice without stopping its own task. [b08-joint-b]


### b08-const-a

Hold IA's 178 head weights and scaling fixed, train 33 log-multiplier constants including MIN_GATE and two anchor coefficients; cap/count/integer/zero parameters stay fixed, decays/alphas/tail slopes <=1. Population 16, seed 20260333, sigma0=0.05, 256 markets/generation: 348 generations/73.98719760227638 minutes. Corrected isolated speed 378.0059910287135 markets/s; initial cache result invalid and concurrent 177.44 is not isolated speed. Best generation 340, monitor +0.21070258891880841 over IA. Constants retained below. [b08-const-a]


### b09-kit-b

No search. Generalized kit packages simulator, configurable port, harness, differential tests, tests, IAD source, normalized weights, and IA restoration patch. Supports selected/reordered/repeated feature IDs 0..18; n,h=1..64 subject to compile/code-size/gas checks; single or firstInStep gating; 19-16-2 has 354 weights and two such heads 708. Affine features clip [-4,4], ReLU clips [0,4], outputs in bps and fees clamp [0,.1]; coefficients quantized 1e-6, Solidity division truncates toward zero. S=zeros(40); strat_init/strat_after use real units. At most eight units per emitted function; partial sums combined before final division. Freeze binds hashes; harness claims measured-gate guards, once-only holdout, locked seed ledger and gas invalidation. Original nine and new fifteen tests pass; reordered gated 5-9-2 tested on 400 callbacks, nonzero 19-16-2 compile-tested, 20-input port tested. These generic probes are not fully trained/validated variants. IA restored byte-for-byte from IAD plus patch; HA/T sources omitted. Generalized throughput 327.9006234923561 versus supplied IA 635.4346768129121 and original kit 765.4776653194682 markets/s: flexibility has overhead. Eight live storage words; packed slots 0/3/6/13, slots 2/5/14/15 unchanged. Fresh IAD equality supports gas optimization, not performance gain. Runtime binaries are not embedded; this is not a self-contained offline installation. [b09-kit-b]


### b10-inputs-clock

No experiment performed. Worker reports roughly 31 minutes remaining versus a required minimum 35-minute search plus setup/gates/holdout/confirmation; it deliberately returned measurements null. Input hashes were copied from the brief, not independently downloaded or verified by that worker. Design only: timestamp/10000, anchor sample count/674, min(callback gap,32), previous-step trade count, previous callback arb-like boolean. Proposed 178 old plus five new input columns, seed 20260341/population 16; no gas-based feature drops determined. Reserved real seeds beginning 7281920 were not used. The timing account is a source claim, not this merge's runtime condition. [b10-inputs-clock]


### b10-inputs-path

24-8-2 head with five path features, no dropped inputs. Population 16, seed 20260342, restart sigma0=0.524401550431785, 128 markets/generation; 449 generations/68.12335278958781 minutes. IA/expanded speeds 205.23446363296566/181.14784505779232 markets/s. Cohort batching preserves exact outputs, speeds up 1.203252574297755 times; 8/16/32-worker tests give 141.1138524066414/185.73814612941038/225.80160891739845 markets/s. Selected generation 390 mean, monitor +0.1824722471871834. Zero-column parity 0 bps; nonzero probe 0.00006636063988249563 bps. Two 16-market gas probes peak 221516 and 240034; frozen real peak 243585 is under 245000 but above the 243000 probe target, leaving little headroom. Packed 80-bit prices cover challenge range only, not arbitrary token pairs. [b10-inputs-path]


### b10-inputs-flow

25-8-2 head, six flow inputs; all kept. Four-process speed 418.1770921766897 markets/s (T 636.8824317165482), generation 5.075251969695091 seconds. Population 16, seed 20260343, 128 markets/generation; 720 generations/60.903023636341096 minutes. Freeze generation 660 mean, monitor +0.3503654246705974 over IA. Zero new weights match IA params exactly on 2000 callbacks. Nonzero probe peaks 215886 versus IA 211980 on 16 markets; full result peaks 239388. Packs new memory in spare slot[10] bits; no new storage word. Replica over-IA prediction +0.3408202870103751 differs from real +0.32168536156486005 by about 0.02. T and HA not rerun on real seeds. Checkpoint/RNG and curve caveats are retained above. [b10-inputs-flow]


### b10-inputs-fees

24-8-2 head, five fee features, all kept. Six processes: IA 256.80430363250304, T 444.7129004813607 markets/s; new-port 128-market timing 122.32994455885714 includes compilation. Population 16, seed 20260344, 128 markets/generation; 297 generations/59.9 minutes, selected generation 290 mean at +0.1496845166610067 over IA. New zero weights with frozen IA params match all output columns on 128 markets; nonzero emitter gate 0.00009844873301995127 bps. Probe peak 214585 versus IA 211968; frozen peak 238236. IA optimizer-restart start is about -0.013 versus frozen IA on the monitor. No real T/HA arms; feature memory fits existing bid/ask slots. [b10-inputs-fees]


### b10-inputs-cross

24-8-2 head, five products of IA's already scaled/clipped features; no new persistent state or dropped inputs. Expanded speed 186.63383107287544 markets/s (kit 227.15590521967428). Population 16, seed 20260345, sigma0=1, 128 markets/generation; 339 generations/65.89247309930312 minutes, 694272 training market evaluations, 33 monitor checks. Selected generation 320 mean: +1.7257565778023238 over T (+0.18619052616003162 over IA). Probe peak 229829 on 16 markets; nine adapted tests pass. Frozen Gate C error 0.0015037960903949155 bps, larger than other r2 candidates but below 0.1. `TH_HA` in the exported port aliases the frozen candidate; `run_batch_seconds.HA` is not evidence of a true HA comparison. Replica/real IA gains differ by 0.01765393495264406 (source comparison field). [b10-inputs-cross]


### b11-fu-a

Setup blocked: Linux x86_64 simulator wheel rejected as incompatible on Darwin arm64. No training, speed, gates, freeze, holdout, real, or gas results; valid=false. Source reports hashes verified for FA code/port, RS2 port and HA code. [b11-fu-a]


### b11-fu-b

UN is FA plus raw_size, step_trade_count, steps_since_arb_like_print, router_implied_vanilla_spot_dev: 23-8-2, 210 weights. Diagonal CMA, seed 312, population 16, sigma0=0.05 capped by 0.05*0.5^(minutes/30), 512 markets/generation; 146 generations/42.87 minutes. Holdout throughput 268.2 markets/s, training about 500 including T. Monitor over T starts 1.7285938423048703 and ends 1.8061432793106984. All four feature columns receive modest absolute-weight sums 1.091/0.553/1.092/0.79 (old inputs roughly 3..17); not an ablation. Nonzero epsilon probe 213710 gas, no dropped input; exact zero-weight EVM equality over 18000 callbacks. Gate C 56x2000, not merely 50x2000. valid=true in source, but required edge decomposition missing. Real gain differs from holdout over FA by about 0.0365; HA arm was retained. [b11-fu-b]


### b11-fu-c

Same union family, seed 313/population 16/sigma0=0.08, 512 replica markets/generation. Reported speed 339.2267857721121; 36 generations. Conflicting time fields retained above. Gates A/B are inherited FA references with hash checks, not freshly repeated empirical gates. Frozen Gate C unfinished. Once-only holdout weakly positive over FA, real mean negative and statistically inconclusive. Both real arms completed in 16-seed chunks; HA intentionally dropped for time. Empty worst5 is missing data, not absence of losses. valid=false despite 0 OOG and peak below cap. [b11-fu-c]


### b12-ib-a

Unchanged IB, no training or policy edits. All 16 ascending 64-seed chunks complete (1024 seeds); every config checked field-for-field against MatchRunner. Nine kit tests pass. Gas reset/read per arm; failure counters include all callbacks. Total edges from Rust, retail/arb reconstructed; maximum arb-volume reconstruction error 0.00010703265797928907. Rounded four-decimal per-seed x-minus-HA list has 1024 entries, while summary statistics use unrounded data. No replica benchmark/gates rerun; valid applies reported real gas criterion. [b12-ib-a]


### b12-ib-b

Setup blocked with exit 126: downloaded Linux uv binary cannot execute on Darwin arm64. Wheel SHA check passed; no other named inputs downloaded/hash-checked. Zero completed seeds/chunks beginning at 7510000. No kit tests, timing, real comparisons, gas, or validity measurement. No policy changes. [b12-ib-b]


### b12-fa-a

Unchanged FA, no training/code/weight edits. All 16 ascending 64-seed chunks complete (1024 seeds), each arm independently against Vanilla, nine kit tests pass. Rust total edge plus kit float retail/arb reconstruction; maximum arb-volume error 0.00009889002103591338. Replica training gates not rerun. [b12-fa-a]


### b12-fa-b

Independent unchanged-FA confirmation: all 16 64-seed chunks/1024 seeds, nine kit tests, all seven kit files and five named source hashes reported checked. No replica evaluation. Reconstruction error 0.00011042502228519879. First-chunk run_batch durations candidate/HA/T 16.774744525086135/16.356789972633123/12.586185064632446 seconds. Tail seed 7530024 hits FA-minus-T -21.236066425686545 and HA-minus-T -21.81424126051627; do not erase this by reporting only FA-minus-HA. [b12-fa-b]


### b13-c1

Speed fallback, unchanged IB. Isolated 146.36746410416507 markets/s on 512 10000-step markets/eight workers; initial overlapping compile timing 160.3435 also below 200. No search or holdout. Gate A reused legitimately as frozen C because code and weights unchanged; Gate B uses historical IB target +1.749. Real 2592 pairs in 32-seed chunks, engine_seconds=5649.764570685802 aggregate, not one command. vs_IB is identity from the same arm, not a second execution. Volumes cover only 2560 pairs; requested edge split missing. [b13-c1]


### b13-c2

Train all 178 IB head weights only; byte-for-byte generator round-trip reconstructs IB, freeze changes numeric weight literals only. Speed 238.0666982143386 markets/s, population 16, seed 402, 1024 fresh markets/generation, 62 generations/66.01403654532817 minutes. Normalize block perturbations by 0.05*RMS; normalized sigma capped by 0.5^(minutes/30). Workers change 8 ->32 at generation 15 ->8 at 20. Monitor best at generation 43/46.04800781806309 minutes: +1.8294615098106202 over T, only +0.019890087435309622 over initial IB. Holdout increment +0.00989863350072289 and real +0.011245308912959162 do not show a clear gain. HA port represented by original 13-feature weights padded with six zero columns. Real decomposition is volume, not edge. [b13-c2]


### b13-c3

Train only 178 IB head literals, seed 403, population 32, 512 markets/generation, 868 generations/45.23 minutes. Measured 6222.4 markets/s (512 batch), 6575.0 (2048 batch), search 5404.2 with 30 processes. Sigma0=.05, block RMS scaling, cap .05*0.5^(minutes/30). Six-generation smoke state discarded, its seeds excluded from real training. Seven <=7-minute segments. Monitor +1.8096 at start to best +1.9923362726676337 at generation 867/45.177 minutes; final +1.9895. Exact IB generator round-trip; candidate changes only head literals. Paired gain over IB transfers from replica +0.19188324927575884 to real +0.20872903151392597. [b13-c3]


### b13-c4

Speed fallback: 104.2 then 68.5 markets/s under high machine load, both below 200. No search, gates, or holdout. Return unchanged IB with identical code hash to c1 and ib-a, compare to HA on 512 fresh seeds in 32-seed chunks. vs_IB is zero by identity; empty worst5 is not an independent test. Real Y-volume differences are not edge decomposition. Source notes an earlier attempt's out-of-scope scratch checkout; this run reports work under test/scratch/amm. [b13-c4]


## Measurement ledger

All entries below are transcribed from the named source, not new measurements. `mean` and `mean_diff` are paired edge gains unless explicitly labelled volume; SE belongs to that comparison. Worst-five entries pair seed with edge difference. Empty arrays and nulls are preserved as missing/identity data, never filled. The source's valid flag is retained separately from this merge's compliance caveats. Tables round noninteger measurements to six significant digits; raw engine means and redundant sums/counters are omitted. Null or absent is shown as missing. Source JSON retains full precision; the narrative above preserves selected full-precision figures.

Common r2 IA gate values: A maximum = 6.501554232485329e-05 bps over 50x2000 swaps (plus initialization); B replica IA-minus-T = 1.4654026984931672, SE 0.04614738409203076, n=512, seeds 7116384..7116895. The quoted real reference is either rounded 1.513 (absolute gap 0.04759730150683272) or 1.5132488526085854 (gap 0.04784615411541826). A/B apply to frozen IA, not a different restart. Every r2 trained item below reports A/B passed and C passed except joint-b (no C) and inputs-clock (none). Repeated identical calibration values are not additional independent real tests. [b06-gated-b; b06-gated-c; b07-wide-a; b07-wide-b; b07-wide-c; b08-joint-a; b08-joint-b; b08-const-a; b09-kit-b; b10-inputs-path; b10-inputs-flow; b10-inputs-fees; b10-inputs-cross]

### Real paired edge comparisons

Source is the input named in each row, its `real` comparison field (or top-level comparison for b12). Retail/arb columns are edge deltas only; volume-only reports remain missing here and are preserved separately below.

| Source | Comparison | n | Mean diff | SE | Retail / arb edge delta |
|---|---|---:|---:|---:|---|
| b12-ib-a | x_vs_ha | 1024 | 0.628081 | 0.0231213 | 0.980428 / -0.352347 |
| b12-ib-a | x_vs_t | 1024 | 1.77984 | 0.031317 | 5.05326 / -3.27342 |
| b12-ib-a | ha_vs_t | 1024 | 1.15176 | 0.0278638 | 4.07283 / -2.92107 |
| b08-joint-a | vs_IA | 512 | 0.0758577 | 0.0186133 | -0.541148 / 0.617005 |
| b10-inputs-flow | vs_IA | 512 | 0.321685 | 0.0311208 | -0.199617 / 0.521302 |
| b11-fu-c | vs_FA | 512 | -0.0129353 | 0.0181153 | missing / missing |
| b06-gated-c | vs_IA | 512 | 0.0657026 | 0.0568563 | -0.21398 / 0.279683 |
| b07-wide-b | vs_IA | 512 | 0.0884601 | 0.0192208 | -0.113175 / 0.201635 |
| b08-const-a | vs_IA | 512 | 0.231421 | 0.0343056 | -0.777658 / 1.00908 |
| b07-wide-c | vs_IA | 512 | 0.128799 | 0.0169564 | -0.203972 / 0.33277 |
| b07-wide-a | vs_IA | 512 | 0.156775 | 0.0215165 | -0.409056 / 0.565831 |
| b10-inputs-fees | vs_IA | 512 | 0.155631 | 0.0234576 | -0.685931 / 0.841561 |
| b06-gated-b | vs_IA | 512 | 0.100926 | 0.0156988 | -0.0303963 / 0.131322 |
| b06-gated-b | vs_T | 512 | 1.60831 | 0.0449746 | 4.64723 / -3.03892 |
| b06-gated-b | IA_vs_T | 512 | 1.50739 | 0.0435398 | 4.67763 / -3.17024 |
| b13-c3 | vs_IB | 512 | 0.208729 | 0.0229775 | missing / missing |
| b13-c3 | vs_HA | 512 | 0.783193 | 0.0325445 | missing / missing |
| b10-inputs-path | vs_IA | 512 | 0.119195 | 0.0252795 | -0.474542 / 0.593737 |
| b11-fu-b | vs_FA | 512 | 0.112096 | 0.0170978 | missing / missing |
| b11-fu-b | vs_HA | 512 | 0.64789 | 0.0329437 | missing / missing |
| b13-c1 | vs_IB | 2592 | 0 | 0 | missing / missing |
| b13-c1 | vs_HA | 2592 | 0.654579 | 0.0143896 | missing / missing |
| b12-fa-a | x_vs_ha | 1024 | 0.569479 | 0.023971 | 0.489707 / 0.0797724 |
| b12-fa-a | x_vs_t | 1024 | 1.7071 | 0.0336454 | 4.56909 / -2.86198 |
| b12-fa-a | ha_vs_t | 1024 | 1.13762 | 0.0286758 | 4.07938 / -2.94176 |
| b12-fa-b | x_vs_ha | 1024 | 0.574542 | 0.0204255 | 0.498793 / 0.0757497 |
| b12-fa-b | x_vs_t | 1024 | 1.67118 | 0.0371068 | 4.52502 / -2.85384 |
| b12-fa-b | ha_vs_t | 1024 | 1.09664 | 0.0339633 | 4.02623 / -2.92959 |
| b10-inputs-cross | vs_IA | 512 | 0.238041 | 0.0298085 | -0.423378 / 0.66142 |
| b13-c2 | vs_IB | 512 | 0.0112453 | 0.0153673 | missing / missing |
| b13-c2 | vs_HA | 512 | 0.678421 | 0.0295721 | missing / missing |
| b09-kit-b | IAD_vs_IA | 512 | 0 | 0 | 0 / 0 |
| b13-c4 | vs_IB | 512 | 0 | 0 | missing / missing |
| b13-c4 | vs_HA | 512 | 0.6261 | 0.0294786 | missing / missing |

### Worst-five real paired edge differences

Each row comes from the same source/comparison as above; values are `seed:edge difference`.

| Source | Comparison | Worst five |
|---|---|---|
| b12-ib-a | x_vs_ha | 7500162:-4.41776; 7500098:-3.79263; 7500941:-2.78672; 7500142:-2.13005; 7500904:-2.11753 |
| b12-ib-a | x_vs_t | 7500162:-2.87269; 7500866:-2.60533; 7500904:-2.45096; 7500941:-2.21251; 7500679:-1.48527 |
| b12-ib-a | ha_vs_t | 7500866:-4.62715; 7500230:-4.22943; 7500679:-3.05149; 7500740:-2.22685; 7500244:-1.77919 |
| b08-joint-a | vs_IA | 7265567:-2.09276; 7265682:-1.56546; 7265929:-1.39666; 7265624:-0.95995; 7265554:-0.847172 |
| b10-inputs-flow | vs_IA | 7290442:-3.62398; 7290232:-2.31828; 7290221:-1.40636; 7290363:-1.26672; 7290277:-1.18959 |
| b11-fu-c | vs_FA | empty (source) |
| b06-gated-c | vs_IA | 7249206:-27.6154; 7249332:-1.87968; 7249656:-1.57622; 7249549:-1.56943; 7249435:-1.30328 |
| b07-wide-b | vs_IA | 7257399:-2.85336; 7257695:-1.51678; 7257504:-1.40082; 7257719:-1.29839; 7257455:-1.25408 |
| b08-const-a | vs_IA | 7274033:-2.64989; 7274193:-2.4645; 7274204:-1.68378; 7274194:-1.64529; 7273841:-1.53589 |
| b07-wide-c | vs_IA | 7261810:-1.24101; 7261453:-1.01907; 7261893:-0.785615; 7261511:-0.779869; 7261633:-0.737686 |
| b07-wide-a | vs_IA | 7253553:-1.76809; 7253632:-1.74929; 7253621:-1.67066; 7253743:-1.02698; 7253538:-0.96168 |
| b10-inputs-fees | vs_IA | 7294626:-3.38755; 7294320:-1.45522; 7294685:-1.36243; 7294284:-1.33406; 7294477:-1.31673 |
| b06-gated-b | vs_IA | 7245480:-1.02803; 7245424:-0.919361; 7245360:-0.912605; 7245266:-0.887641; 7245348:-0.773126 |
| b06-gated-b | vs_T | 7245204:-3.9757; 7245191:-2.0713; 7245519:-1.84703; 7245254:-1.08845; 7245102:-1.07663 |
| b06-gated-b | IA_vs_T | 7245204:-3.69957; 7245191:-2.39873; 7245519:-1.71631; 7245102:-1.05333; 7245434:-0.94133 |
| b13-c3 | vs_IB | 7548334:-1.79708; 7548052:-1.66561; 7548050:-1.40919; 7548044:-1.25919; 7548336:-1.10649 |
| b13-c3 | vs_HA | 7548050:-3.01156; 7548459:-2.13584; 7548035:-1.3455; 7548230:-1.17536; 7548177:-1.1201 |
| b10-inputs-path | vs_IA | 7286227:-5.36852; 7286281:-1.64592; 7286123:-1.45136; 7286499:-1.42441; 7286424:-1.38758 |
| b11-fu-b | vs_FA | 7380036:-1.21177; 7380266:-1.12875; 7380000:-0.918332; 7380410:-0.831879; 7380282:-0.82468 |
| b11-fu-b | vs_HA | 7380109:-9.12968; 7380139:-1.37623; 7380308:-1.23245; 7380498:-1.10155; 7380396:-0.951218 |
| b13-c1 | vs_IB | 7540000:0; 7540001:0; 7540002:0; 7540003:0; 7540004:0 |
| b13-c1 | vs_HA | 7541569:-7.29449; 7541365:-2.63212; 7541476:-2.1896; 7542180:-2.0188; 7542413:-1.96923 |
| b12-fa-a | x_vs_ha | 7520325:-1.92113; 7520275:-1.70459; 7520826:-1.64396; 7520205:-1.53181; 7520615:-1.43864 |
| b12-fa-a | x_vs_t | 7520642:-4.07951; 7520176:-2.1385; 7520323:-1.78836; 7521003:-1.74429; 7520065:-1.56695 |
| b12-fa-a | ha_vs_t | 7520250:-3.66961; 7520642:-3.30868; 7520808:-3.25708; 7520941:-2.42662; 7520797:-2.32379 |
| b12-fa-b | x_vs_ha | 7530557:-3.7649; 7530512:-2.07862; 7530850:-1.72981; 7530066:-1.66881; 7530833:-1.6289 |
| b12-fa-b | x_vs_t | 7530024:-21.2361; 7530547:-1.62954; 7530512:-1.43698; 7530047:-1.28546; 7530331:-1.02076 |
| b12-fa-b | ha_vs_t | 7530024:-21.8142; 7530547:-2.17709; 7530047:-1.95219; 7530719:-1.76429; 7530424:-1.62144 |
| b10-inputs-cross | vs_IA | 7298680:-1.98736; 7298418:-1.85733; 7298562:-1.60185; 7298810:-1.37907; 7298716:-1.23896 |
| b13-c2 | vs_IB | 7544477:-1.25299; 7544348:-1.20706; 7544025:-1.20288; 7544134:-1.16354; 7544095:-1.06953 |
| b13-c2 | vs_HA | 7544031:-1.49545; 7544294:-1.37804; 7544477:-1.24595; 7544507:-0.922143; 7544452:-0.753183 |
| b09-kit-b | IAD_vs_IA | 7277824:0; 7277825:0; 7277826:0; 7277827:0; 7277828:0 |
| b13-c4 | vs_IB | empty (source) |
| b13-c4 | vs_HA | 7552398:-3.22482; 7552425:-2.66413; 7552255:-2.10696; 7552113:-2.1069; 7552090:-1.29078 |

### Replica holdout comparisons

Sources: each input `holdout`; common block 40000000..40008191 unless explicitly absent. For fu-c, the sole supplied SE is attached here only to gain_over_T; an SE for gain_over_FA is missing. gated-b comparator warning above applies.

| Source | Comparison | n | Mean gain | SE |
|---|---|---:|---:|---:|
| b08-joint-a | vs_T | 8192 | 1.54435 | 0.0111095 |
| b08-joint-a | vs_HA | 8192 | 0.427522 | 0.00647816 |
| b08-joint-a | vs_IA | 8192 | 0.0626425 | 0.00509806 |
| b10-inputs-flow | vs_T | 8192 | 1.82253 | 0.0119941 |
| b10-inputs-flow | vs_HA | 8192 | 0.705699 | 0.00762853 |
| b10-inputs-flow | vs_IA | 8192 | 0.34082 | 0.00670164 |
| b10-inputs-flow | IA_vs_T | 8192 | 1.48171 | 0.011391 |
| b10-inputs-flow | HA_vs_T | 8192 | 1.11683 | 0.0102086 |
| b11-fu-c | gain_over_T | 8192 | 1.69215 | 0.0114355 |
| b11-fu-c | gain_over_FA | 8192 | 0.0110485 | missing |
| b06-gated-c | vs_T | 8192 | 1.60418 | 0.0112009 |
| b06-gated-c | vs_HA | 8192 | 0.487352 | 0.00661274 |
| b06-gated-c | vs_IA | 8192 | 0.122473 | 0.00436193 |
| b06-gated-c | IA_vs_T | 8192 | 1.48171 | 0.011391 |
| b07-wide-b | vs_T | 8192 | 1.58648 | 0.0114164 |
| b07-wide-b | vs_HA | 8192 | 0.46965 | 0.00660715 |
| b07-wide-b | vs_IA | 8192 | 0.104771 | 0.00432415 |
| b08-const-a | vs_T | 8192 | 1.71967 | 0.0128727 |
| b08-const-a | vs_HA | 8192 | 0.602834 | 0.00991066 |
| b08-const-a | vs_IA | 8192 | 0.237955 | 0.00889799 |
| b07-wide-c | vs_T | 8192 | 1.63879 | 0.011372 |
| b07-wide-c | vs_HA | 8192 | 0.521961 | 0.00660578 |
| b07-wide-c | vs_IA | 8192 | 0.157082 | 0.00464315 |
| b07-wide-a | vs_T | 8192 | 1.63556 | 0.0111191 |
| b07-wide-a | vs_HA | 8192 | 0.518733 | 0.00665596 |
| b07-wide-a | vs_IA | 8192 | 0.153853 | 0.00567616 |
| b07-wide-a | IA_vs_T | 8192 | 1.48171 | 0.011391 |
| b10-inputs-fees | vs_T | 8192 | 1.65955 | 0.0112313 |
| b10-inputs-fees | vs_HA | 8192 | 0.54272 | 0.00703282 |
| b10-inputs-fees | vs_IA | 8192 | 0.177841 | 0.00592767 |
| b10-inputs-fees | IA_vs_T | 8192 | 1.48171 | 0.011391 |
| b06-gated-b | vs_T | 8192 | 1.56262 | 0.0112662 |
| b06-gated-b | vs_HA | 8192 | 0.0809054 | 0.00443277 |
| b06-gated-b | vs_IA | 8192 | 0.0809054 | 0.00443277 |
| b13-c3 | over_T | 8192 | 1.96087 | 0.0113433 |
| b13-c3 | over_HA | 8192 | 0.844035 | 0.00825817 |
| b13-c3 | over_IB | 8192 | 0.191883 | 0.00554804 |
| b10-inputs-path | vs_T | 8192 | 1.65901 | 0.0112788 |
| b10-inputs-path | vs_HA | 8192 | 0.542179 | 0.00731976 |
| b10-inputs-path | vs_IA | 8192 | 0.1773 | 0.00601359 |
| b11-fu-b | gain_over_T | 8192 | 1.75672 | 0.0116184 |
| b11-fu-b | gain_over_FA | 8192 | 0.0756195 | 0.00430489 |
| b10-inputs-cross | vs_T | 8192 | 1.7021 | 0.0113835 |
| b10-inputs-cross | vs_IA | 8192 | 0.220387 | 0.00666768 |
| b10-inputs-cross | vs_HA | 8192 | 0.585267 | 0.00747543 |
| b13-c2 | over_T | 8192 | 1.77888 | 0.012006 |
| b13-c2 | over_HA | 8192 | 0.662051 | 0.00863929 |
| b13-c2 | over_IB | 8192 | 0.00989863 | 0.00496201 |

### Gas and final parity

Sources: named input `gas`/`gas_summary`, or gated-b `real.gas_per_arm`. Peaks are per callback, not average gas. All measured successful candidate reports claim zero OOG; missing per-row counters below are not inferred. c1 duplicate aggregate/by-strategy values represent the same measurement.

| Source | Arm | afterSwap peak | afterInitialize peak | OOG |
|---|---|---:|---:|---:|
| b12-ib-a | IB | 237663 | 243227 | 0 |
| b12-ib-a | HA | 234255 | 243147 | 0 |
| b12-ib-a | T | 226215 | 243115 | 0 |
| b08-joint-a | IA | 237644 | 243237 | 0 |
| b08-joint-a | Candidate | 237617 | 243237 | 0 |
| b10-inputs-flow | candidate | 239388 | 243247 | 0 |
| b10-inputs-flow | IA | 235527 | 243237 | 0 |
| b11-fu-c | candidate | 234688 | 243270 | 0 |
| b11-fu-c | FA | 231789 | 243227 | 0 |
| b06-gated-c | candidate | 235526 | 243327 | 0 |
| b06-gated-c | IA | 235654 | 243237 | 0 |
| b07-wide-b | IA | 235527 | 243237 | 0 |
| b07-wide-b | Candidate | 136214 | 176970 | 0 |
| b08-const-a | IA | 238626 | 243237 | 0 |
| b08-const-a | Candidate | 238596 | 243237 | 0 |
| b07-wide-c | Candidate | 135889 | 176964 | 0 |
| b07-wide-c | IA | 237868 | 243237 | 0 |
| b07-wide-a | candidate | 134132 | 176937 | 0 |
| b07-wide-a | IA | 235502 | 243237 | 0 |
| b10-inputs-fees | candidate | 238236 | 243237 | 0 |
| b10-inputs-fees | IA | 235666 | 243237 | 0 |
| b06-gated-b | T | 225591 | 243115 | 0 |
| b06-gated-b | HA | 237892 | 243237 | 0 |
| b06-gated-b | Candidate | 237885 | 243327 | 0 |
| b13-c3 | candidate | 235028 | 243227 | 0 |
| b13-c3 | IB | 234925 | not stated here | 0 |
| b13-c3 | HA | 231592 | not stated here | 0 |
| b10-inputs-path | IA | 236234 | 243237 | 0 |
| b10-inputs-path | Candidate | 243585 | 243272 | 0 |
| b11-fu-b | candidate | 239583 | 243260 | 0 |
| b11-fu-b | FA | 236702 | 243227 | 0 |
| b11-fu-b | HA | 233603 | 243147 | 0 |
| b13-c1 | candidate | 237546 | 243227 | 0 |
| b13-c1 | IB | 237546 | 243227 | 0 |
| b13-c1 | HA | 234135 | 243147 | 0 |
| b12-fa-a | FA | 236888 | 243227 | 0 |
| b12-fa-a | HA | 233761 | 243147 | 0 |
| b12-fa-a | T | 225591 | 243115 | 0 |
| b12-fa-b | T | 225591 | 243115 | 0 |
| b12-fa-b | HA | 233671 | 243147 | 0 |
| b12-fa-b | FA | 236778 | 243227 | 0 |
| b10-inputs-cross | IA | 237624 | 243237 | 0 |
| b10-inputs-cross | Candidate | 239241 | 243254 | 0 |
| b13-c2 | candidate | 237012 | 243227 | 0 |
| b13-c2 | IB | 237021 | 243227 | 0 |
| b13-c2 | HA | 233620 | 243147 | 0 |
| b09-kit-b | IA | 237738 | 243237 | 0 |
| b09-kit-b | IAD | 132084 | 176917 | 0 |
| b13-c4 | candidate | 237690 | 243227 | 0 |
| b13-c4 | IB | 237690 | 243227 | 0 |
| b13-c4 | HA | 234377 | 243147 | 0 |

| Source | Frozen Gate C max fee error (bps) | Source passed |
|---|---:|---|
| b08-joint-a | 9.69445e-05 | True |
| b10-inputs-flow | 0.000249527 | True |
| b06-gated-c | 3.87873e-05 | True |
| b07-wide-b | 0.000147206 | True |
| b08-const-a | 8.81763e-05 | True |
| b07-wide-c | 0.000188391 | True |
| b07-wide-a | 4.02442e-05 | True |
| b10-inputs-fees | 5.53286e-05 | True |
| b06-gated-b | 0.000126867 | True |
| b13-c3 | 9.5988e-05 | True |
| b10-inputs-path | 6.78745e-05 | True |
| b11-fu-b | 4.17e-05 | True |
| b13-c1 | 5.22488e-05 | True |
| b10-inputs-cross | 0.0015038 | True |
| b13-c2 | 9.13067e-05 | True |

Additional gates: fu-b A=4.06e-05 bps, B=1.7050245300047933 (SE 0.04995197116469658), real reference 1.65 on 7004000..7004511. c1 A=5.224879067594612e-05; c2/c3 A=6.64202309125228e-05. All three IB B checks give 1.708612109706731 (SE 0.04425219850674254) versus historical 1.749 on 7132768..7133279. fu-c inherits A/B only, C unfinished; c4 gates absent. [b11-fu-b gates; b11-fu-c gates; b13-c1 gates; b13-c2 gates; b13-c3 gates; b13-c4 gates]

### Volume-only reports (not edge)

- b11-fu-b `real`: `{"retail_delta":null,"arb_delta":null,"retail_volume_y_delta_vs_FA":{"mean":-133.4570656027376,"se":2.768229598202696},"arb_volume_y_delta_vs_FA":{"mean":-95.02325276413208,"se":2.7534819009236995},"retail_volume_y_delta_vs_HA":{"mean":246.50458379340674,"se":10.591410793436243},"arb_volume_y_delta_vs_HA":{"mean":278.40154306604217,"se":6.622379461106862}}`.
- b11-fu-c `real`: `{"retail_delta":null,"arb_delta":null}`.
- b13-c1 `real`: `{"retail_delta":{"n":2560,"mean_diff":629.9461735194828,"units":"Y volume per market","comparison":"IB minus HA"},"arb_delta":{"n":2560,"mean_diff":485.2748601327374,"units":"Y volume per market","comparison":"IB minus HA"}}`.
- b13-c2 `real`: `{"retail_delta":{"vs_IB":63.34175170672256,"vs_HA":693.721933921566},"arb_delta":{"vs_IB":104.15794254884361,"vs_HA":593.3167936098332},"delta_units":"mean Y volume per market; Rust LightweightSimResult has retail/arb volumes but no retail/arb edge decomposition"}`.
- b13-c3 `real`: `{"retail_delta":{"vs_IB":-39.112154455183074,"vs_HA":590.0698000210484,"unit":"mean retail_volume_y per market"},"arb_delta":{"vs_IB":-17.768151487736567,"vs_HA":471.6294153232245,"unit":"mean arb_volume_y per market"},"pnl_delta_vs_IB":0.20431648526566448,"pnl_delta_vs_HA":0.8290623133571557}`.
- b13-c4 `real`: `{"retail_delta":631.4524001714483,"arb_delta":477.5820658603568,"delta_unit":"IB minus HA retail_volume_y and arb_volume_y (engine exposes volumes, not edge split)"}`.

### Seed partitions

These are reported allocations, not an independently audited global reservation ledger. Unmeasured allocations are not evidence of seed consumption. Gate/monitor/holdout reuse is distinct from fresh real confirmation. Source: each named input `seeds` or `real.seeds`.

b12-ib-a: start=7500000; end=7501023; n=1024; chunks_done=16.

b10-inputs-clock: gate_b=7116384..7116895; holdout=40000000..40008191; real_confirm=7281920..7282431 (7281920 + 511); train_from=520000000; monitor=30000000..30002047.

b08-joint-a: train=516000000..516106495; monitor=30000000..30002047; replica_holdout=40000000..40008191; real_confirmation=7265536..7266047; gate_A_C_sequences=310000..310049; gate_B=7116384..7116895.

b10-inputs-flow: gate_A_C=310000..310049; gate_B=7116384..7116895; speed=6010000..6010511; scaling=numpy.default_rng(20260343).integers(522000000,532000000,256); train=522000000..522092159; monitor=30000000..30002047; holdout=40000000..40008191; real_confirmation=7290112..7290623; real_gas_probe=7290624..7290639.

b11-fu-c: real 7385000..7385511.

b06-gated-c: gate_A_C=310000..310049; speed=6010000..6010511; gate_B=7116384..7116895; monitor=30000000..30002047; train=512000000..512043007; gas_probe=7316000..7316015; holdout=40000000..40008191; real=7249152..7249663.

b07-wide-b: gate_A=310000..310049; gate_B=7116384..7116895; regression=0..99; train=514000000..514016639; monitor=30000000..30002047; holdout=40000000..40008191; confirmation=7257344..7257855; benchmark=6010000..6010511.

b08-const-a: train=518000000..518089087; monitor=30000000..30002047; gate_A_C=310000..310049; gate_B=7116384..7116895; speed=6010000..6010511; holdout=40000000..40008191; real=7273728..7274239.

b07-wide-c: gate_A_C=310000..310049; gate_B=7116384..7116895; regression=0..99; training=515000000..515074495; monitor=30000000..30002047; holdout=40000000..40008191; real=7261440..7261951.

b07-wide-a: gate_A_C=310000..310049; gate_B=7116384..7116895; init_check=7116376..7116379; train=513000000..513036863; monitor=30000000..30002047; holdout=40000000..40008191; regression=0..99; real=7253248..7253759.

b10-inputs-fees: train=523000000..532999999; scaling=numpy.default_rng(20260344).integers(523000000,533000000,256); monitor=30000000..30002047; replica_holdout=40000000..40008191; real_fresh=7294208..7294719; gas_probe=7290000..7290015; gate_A_C=310000..310049; gate_B=7116384..7116895.

b08-joint-b: differential=310000..310049; benchmark=6010000..6010511; gate_B=7116384..7116895; train_reserved_unused_start=517000000; monitor_reserved_unused=30000000..30002047; replica_holdout_reserved_unused=40000000..40008191; real_confirmation_reserved_unused=7269632..7270143.

b06-gated-b: gate_A_C=310000..310049; gate_B=7116384..7116895; train_start=511000000; monitor=30000000..30002047; holdout=40000000..40008191; real=7245056..7245567.

b13-c3: real 7548000..7548511.

b10-inputs-path: gate_A=310000..310049; gate_B=7116384..7116895; gate_C=310000..310049; training=521000000..521057471; monitor=30000000..30002047; replica_holdout=40000000..40008191; real=7286016..7286527; gas_probes=[7270000, 7270015]..[7270016, 7270031].

b11-fu-b: real 7380000..7380511.

b13-c1: real 7540000..7542591.

b12-ib-b: start=7510000; end=None; n=0; chunks_done=0.

b12-fa-a: start=7520000; end=7521023; n=1024; chunks_done=16.

b12-fa-b: start=7530000; end=7531023; n=1024; chunks_done=16.

b10-inputs-cross: port_differential=310000..310049; gate_B=7116384..7116895; scaling=numpy.default_rng(20260345).integers(524000000,534000000,256); train=524000000..524043391; monitor=30000000..30002047; replica_holdout=40000000..40008191; pretrain_gas=7296000..7296015; real_confirmation=7298304..7298815.

b13-c2: real 7544000..7544511.

b09-kit-b: callback_sequences=310000..310049; callbacks_per_sequence=2000; kit_replica_B=6020000..6020511; IA_replica_B=7116384..7116895; regression=0..99; confirmation=7277824..7278335; confirmation_chunks=[7277824, 7277888, 7277952, 7278016, 7278080, 7278144, 7278208, 7278272]; confirmation_reused=False; train_not_run=400000000..409999999; monitor_not_run=30000000..30002047; holdout_not_run=40000000..40008191.

b13-c4: real 7552000..7552511.

## Feature formulas and reusable technical findings

All formulas use TradeInfo and own stored history. These definitions are not feature-importance ablations. IA's features are sigmaHat, lambdaHat, sizeHat, toxEma, actEma, direction, spot/staleHat deviation, signed trade ratio, arb probability, firstInStep, uncertainty, clipped silent gap, anchor mean, raw size, step count, previous bid/ask fees, steps since arb-like print, router-implied Vanilla spot deviation. Affine features clip [-4,4], ReLU clips [0,4], output is bps added to T then clamped [0,.1]; coefficients use 1e6 fixed point. [b09-kit-b api]

- **Clock (design only):** timestamp/10000; anchor count/674; min(callback gap,32); previous-step trade count; previous callback arb-like 0/1. IA new-input scaling was proposed but not checked. No implementation or feature-drop result. [b10-inputs-clock new_input_formulas]
- **Path:** log(post-trade spot/initialSpot), computed in Solidity by binary range reduction and atanh series through z^15; (spot-pHat)/pHat and (spot-fastHat)/fastHat use updated estimates. On firstInStep, store `trunc(clip(spot/previousFirstSpot-1,-2,2)*1e9)/1e9`; divide stored move by current sigmaHat on every callback. Store sign(newMove*previousMove), zero if either move zero, held between first callbacks. Initial previousFirstSpot=initialSpot. slot[6] packs sigma and initialSpot in two 80-bit fields; slot[7] packs lambda/lastFirstSpot (80 bits each), signed 32-bit move and 2-bit sign. Existing reads mask low bits and writes preserve high bits; no new storage word. [b10-inputs-path new_input_formulas]
- **Flow:** `sf=(isBuy?+1:-1)*min(amountY/reserveY,10)*w`, where w=1-arbProbability on first callback, otherwise 1. Current-inclusive EMAs are .5*old+.5*sf and .95*old+.05*sf. Other features: min(raw_size/sizeHat,16), or zero if sizeHat=0; previous raw size; previous side (+1/-1/0); same-side run including current trade capped 8. slot[10] preserves low16 step count; signed WAD EMAs at bits16..79/80..143, prior size144..207, side208..209, run212..215. [b10-inputs-flow new_input_formulas]
- **Fees:** current T-path bid and ask after tail compression/clamps/first-step 30/36-bps caps but before head; their difference; prior bid/ask head outputs in bps truncated to 1e-6, initially zero. Prior head outputs occupy signed bits64..127 in slots[0]/[1]; low64 retain fee. No new word. [b10-inputs-fees new_input_formulas]
- **Cross:** with x the already IA-scaled/clipped vector, products x8*x13 (arb probability/raw size), x9*x6 (first/stale deviation), x0*x3 (sigma/tox), x5*x7 (direction/signed size), x9*x11 (first/silent). Added multiplier/offset pairs respectively 1.509575/0.189311, 0.417175/0.001677, 0.241944/-0.602769, 43.532918/-0.139428, 7.627522/-0.896066. First is +/-1, direction=dirState-1, silent=min(silent,8)/8; no persistent state added. [b10-inputs-cross new_input_formulas/notes]
- **FA union:** add raw size, step count, steps since arb-like print, router-implied deviation with affine pairs 647.054817/-0.618293, 1.395287/-2.037262, 0.412565/-0.823811, 274.335453/0.001772; all new weights start zero over FA. [b11-fu-b params]

Original-kit checks: HA A/C 7.60422131512728e-05 bps, T 1.2216616607219066e-09 bps; replica HA-T 1.141450774002562 versus historical 1.257 (gap 0.115549225997438 <0.15). kit-b's IAD/wide16/emitted C each 6.501554232485329e-05 bps with zero exact-reference EVM callback mismatches. Zero-output widened checks do not validate arbitrary learned heads. [b09-kit-b checks/gates]

## Frozen constant changes

Values below are real-unit source values, not bps unless named as such. Each pair is original T/anchor value -> frozen value. joint-a fixes MIN_GATE; const-a optimizes it. Both keep anchor count cap 674 and the named caps/count/integer/zero settings fixed. The integer anchor WAD coefficients in frozen params are joint-a linear 715509569811160704, quadratic 1296032273166540032; const-a linear 669627321299906944, quadratic 1411810287645655808. [b08-joint-a constants/params; b08-const-a constants/params]

| Constant | Original | joint-a frozen | const-a frozen |
|---|---:|---:|---:|
| ELAPSED_CAP | 8.0 | 8.0 | 8.0 |
| SIGNAL_THRESHOLD | 0.001538461538461538 | 0.0015225158143694218 | 0.001386600831159654 |
| DIR_DECAY | 0.75 | 0.7556178388251961 | 0.7056833605074135 |
| ACT_DECAY | 0.53 | 0.5456432290211817 | 0.6776678006160287 |
| SIZE_DECAY | 0.53 | 0.5166919175001468 | 0.5071233852581395 |
| TOX_DECAY | 0.91 | 0.9097891683830247 | 0.9284929484280295 |
| SIGMA_DECAY | 0.69 | 0.7117591006886598 | 0.7003266502040828 |
| LAMBDA_DECAY | 0.99 | 0.9911094574887175 | 0.9955186113044672 |
| SIZE_BLEND_DECAY | 0.828 | 0.8510501717234721 | 0.8572867152615148 |
| TOX_BLEND_DECAY | 0.051 | 0.0511443702799374 | 0.06166764560156666 |
| ACT_BLEND_DECAY | 0.985 | 0.9852215971905479 | 0.9887231557639652 |
| PHAT_ALPHA | 0.27 | 0.26304520811525933 | 0.25323361432837255 |
| PHAT_ALPHA_RETAIL | 0.06 | 0.0596733642868973 | 0.04587789303941362 |
| PHAT_ALPHA_SAME | 0.12 | 0.11959600143005024 | 0.11651458115432142 |
| REVERT_U | 0.0 | 0.0 | 0.0 |
| DIR_IMPACT_MULT | 4.0 | 4.0 | 4.0 |
| GATE_SIGMA_MULT | 7.2 | 7.280032636713512 | 5.8533522920776075 |
| MIN_GATE | 0.03 | 0.03 | 0.0230515200964478 |
| TRADE_TOX_BOOST | 0.35 | 0.3315949532344332 | 0.3299165698173456 |
| STALE_ATTRACT_FRAC | 1.05 | 1.0369999945313373 | 1.015292168061998 |
| RET_CAP | 0.02857142857142857 | 0.02857142857142857 | 0.02857142857142857 |
| TOX_CAP | 0.2 | 0.2 | 0.2 |
| TRADE_RATIO_CAP | 0.2 | 0.2 | 0.2 |
| LAMBDA_CAP | 1.82 | 1.82 | 1.82 |
| STEP_COUNT_CAP | 64.0 | 64.0 | 64.0 |
| BASE_FEE | 0.0004 | 0.0004016489583752763 | 0.00036716378142068 |
| SIGMA_COEF | 0.2 | 0.19713010511068885 | 0.16756472465067151 |
| LAMBDA_COEF | 0.0012 | 0.0012088389028354278 | 0.00114941673010786 |
| FLOW_SIZE_COEF | 0.4842 | 0.4684747456594109 | 0.6616432731552307 |
| TOX_COEF | 0.03 | 0.030119886264442754 | 0.03750174990690164 |
| TOX_QUAD_COEF | 1.5 | 1.4757686394722218 | 1.4433421342886186 |
| ACT_COEF | 9.1843 | 8.92694254722055 | 9.117984480349524 |
| DIR_COEF | 0.002 | 0.002127138457650291 | 0.002614832967025314 |
| DIR_TOX_COEF | 0.01 | 0.009818480174712017 | 0.011458913823147811 |
| STALE_DIR_COEF | 0.685 | 0.697242285070208 | 0.6812809868770082 |
| SIGMA_TOX_COEF | 0.05 | 0.04908605307169276 | 0.06288551586626327 |
| TAIL_KNEE | 0.06 | 0.05864400902657272 | 0.06068322980890787 |
| TAIL_SLOPE_PROTECT | 0.93 | 0.9257709860150077 | 0.6844570275108228 |
| TAIL_SLOPE_ATTRACT | 0.85 | 0.8490660732593678 | 0.6814955536710317 |
| ANCHOR_COUNT_CAP | 674 | 674 | 674.0 |
| ANCHOR_LINEAR_K | 0.7126960148076508 | 0.7155095698111607 | 0.669627321299907 |
| ANCHOR_QUADRATIC_K | 1.2952407538263107 | 1.2960322731665401 | 1.4118102876456557 |

## Artifact identity and provenance

All source references refer to the supplied local input bytes. Original locations are `artifacts/goal/batches/<batch>/items/<item>/result.json`; mirrors are `.imd/reads/artifacts/<input>`. Executable payloads, full weights/optimizer state, raw curves/per-seed arrays, duplicate counters, logs and reproduction scripts are not embedded. Code identities are retained; most port hashes are omitted. This is a write-up, not a runnable bundle; several workers did not bundle runtime dependencies. Missing evidence stays missing.

The table records each input JSON SHA-256.

| Input | Batch/item | JSON SHA-256 | Failed source rechecks reported |
|---|---|---|---:|
| b06-gated-b | `r2-gated/gated-b` | `8378f9f0040c666b5d1b0156c9b36a103add13f4da1263aa53fdf9029084eb86` | 0 |
| b06-gated-c | `r2-gated/gated-c` | `d70dcb499e0aa62c563a683f362583de45cb7b516bc82ef818e0de7a6eed491d` | 3 |
| b07-wide-a | `r2-diet-wide/wide-a` | `c00acd843105737c3fd38b436e8d2fba3ea026612da2920b86947298ae964ac5` | 1 |
| b07-wide-b | `r2-diet-wide/wide-b` | `9e5a5bce69a1d9d19cc4325d6ace7ef7bf173ad820435436c6279f0dcde4ba3a` | 2 |
| b07-wide-c | `r2-diet-wide/wide-c` | `238562192553219ecbbef3157c28a10e09ae454d18f9657422c530f52769136d` | 2 |
| b08-joint-a | `r2-joint/joint-a` | `84090c51554ba15d3f44c99d7507b2617a414d8e7cd06bcb69bfa0405e334078` | 2 |
| b08-joint-b | `r2-joint/joint-b` | `49a363f0ad725acf20be4c4aa9ea2e1e00a92d2739178cf9934833693dc4aa2f` | 3 |
| b08-const-a | `r2-joint/const-a` | `a9279075abf7398da55fe663698052b77902f1a0e7f81105c1589d474d39b246` | 2 |
| b09-kit-b | `r2-kit/kit-b` | `48cd6a10f39fc22a79c5f12398474bd65282803d59314819acd586f187015808` | 2 |
| b10-inputs-clock | `r2-inputs2/inputs-clock` | `5689cbf90c8543079b22e1af71431ce0257b0dc4fee6bc4311c6ed12c618313e` | 0 |
| b10-inputs-path | `r2-inputs2/inputs-path` | `70a6391476779bbae83fcca6151837ee89f5ac8e190e74177614821e84844db2` | 2 |
| b10-inputs-flow | `r2-inputs2/inputs-flow` | `f3b4ec1ee4e96397fadf99e2a78bb97ed3d4f31e256864e51d861753a79b4016` | 2 |
| b10-inputs-fees | `r2-inputs2/inputs-fees` | `d3601e2846d72d088203a5c8a866dafd518ce7582ee683d954d362c76005e39c` | 2 |
| b10-inputs-cross | `r2-inputs2/inputs-cross` | `90b13643d59f78fbde3387add271164774f716019c87a26b502b9e802541c391` | 2 |
| b11-fu-a | `r3-fa-union/fu-a` | `73c0aa6ea64dcc3e856f814de8b56e61606ee4b83eef1643df10a115cfb2c146` | 0 |
| b11-fu-b | `r3-fa-union/fu-b` | `e2590eb635f0ac283492b2611d9202c19f64c03a3a1e8a3ffbad442a1f6989e2` | 0 |
| b11-fu-c | `r3-fa-union/fu-c` | `9b2e0e1d78cde2e091f4f0a331da6b85c3fcf0c851a76ee0aadebb3eb1637baa` | 0 |
| b12-ib-a | `r4-confirm/ib-a` | `dbf6c209fbb4a30a2a02c353f696ba58c9ffb324f870abdaf0fb769d445f7c19` | 3 |
| b12-ib-b | `r4-confirm/ib-b` | `e69c0685be98897aec7ff0f58c5b515b45f3fdc0dac33eae920f4f71eea829c0` | 0 |
| b12-fa-a | `r4-confirm/fa-a` | `f08aee506aadb8e79f7ede0905b4e46c0966be7f9edd24f159a2ea5bd29aff67` | 2 |
| b12-fa-b | `r4-confirm/fa-b` | `1c211deffc4f09d4b4aba35617af9fde5f5d55c41a7b546ad8f567ca30296ac9` | 2 |
| b13-c1 | `r4-ib-continue/c1` | `a101f1f43ebb91e12d4eea6d6bda7de3b2dac50545b22bc307b791415d7b55ab` | 4 |
| b13-c2 | `r4-ib-continue/c2` | `778be5203ddb5b1ea274888f1aebb35458db5745b65b62646576aa6cea198759` | 5 |
| b13-c3 | `r4-ib-continue/c3` | `ff1979d22ebe2c22fc6388ed443a4ba5d06ef0a9b111a16260a0fb96e0902c8d` | 3 |
| b13-c4 | `r4-ib-continue/c4` | `67f7e09533a675ce694947f5f3a5b6e861156c37555bbf0816fd26e36a60230a` | 0 |

| Source | Frozen code SHA-256 |
|---|---|
| b06-gated-b | `e99d9209c4dc24c1032b766c65ce9f2c2927a24f9dbbfc999ed68ff27667a629` |
| b06-gated-c | `cd8b04bebdaac9b9c798651d7df3ecc9a46f777581842ec2064630edc0a3de56` |
| b07-wide-a | `187f7c5173de8df12a7b22ddd0579255745e34bb448f8a2ec374580b81414e0c` |
| b07-wide-b | `ad32f4da7cd1d24035bc2d7986a07fe34c5bf5b059652115390848259c4b4e0d` |
| b07-wide-c | `48a493006cbc6e49f3559d0c9dfcafff35ad3a031e80d6bc368e192bcf83320c` |
| b08-joint-a | `cbd94c38923e8c624191169305b7d7bc6dbabd64341835065e9ab27a1b7399fb` |
| b08-const-a | `e8a9459f6a3051467d9ed73360bab25c092127c4ab1db763e8618c28e83db800` |
| b09-kit-b | `7f6998c8df1b738dbd9715053edae7c9c8fdad71bf7228d67b6f1b7573c7126b` |
| b10-inputs-path | `51cb7bf07c0a877e79a436a5230a6ccacd4a01839c96d4b15c9f9972a17268c5` |
| b10-inputs-flow | `b15aa3c2414dddcbdd43031324e115bd25ee62ea6254b0f30c41bdf087c88ccc` |
| b10-inputs-fees | `c08ffab82be53c8e51845dd542366d244b3f988b87d9d58cf91ba50a924a1754` |
| b10-inputs-cross | `59c2b5276ed26e7f80eb63c41b85bf92187bf58bc23488bf678b26c4b7ab1ee5` |
| b11-fu-b | `64664c77c560218d085c340ec2d9da2d9f54a4d90ed02fb0c2be464f14507a12` |
| b11-fu-c | `5271446c93f1377aba607854fba7107b55b5e2dce166de8c37ef8b2a876a81dd` |
| b13-c1 | `301941829a44bb60977bb28181a5e4bfa6544f3696763f987054789b2611d686` |
| b13-c2 | `c0cc6defb40170c1e10825e21bf3c0917b1c5782c0355eded78b30993308b6cd` |
| b13-c3 | `df1711cccb640ceffbbb50546c099612196741f469b4cfa05b924e6844b8c91b` |
| b13-c4 | `301941829a44bb60977bb28181a5e4bfa6544f3696763f987054789b2611d686` |

Additional pinned identities reported by the workers: T code `ebd9de9bf6af6e55d5e308fd15dbb88edec11fd684891d02a30aec185acd19b3`; replica source `1e4c44b4bffbdeed65872ee15d46dbde209bf5aa11014329d13d100507b4040c`; RS2 port `9eac3553357528b06eeae3ed8d18283d83401e1968145ceba16be356a147e271`; IA code `956566e95f46dc1ed5980fae5e550de7b9da10f9b3ad8eed62f265682cfcecf2`, port `b29cd2a078b4c3fc57945a79e927624dd346c44d8cc149dae43f4929aa6d736b`; IB code `301941829a44bb60977bb28181a5e4bfa6544f3696763f987054789b2611d686`, port `f63227c5e0f63633efd132d6fb402ede665e5c9cb24481911471e8dc20a3362d`; FA code `aa2c4a5637a4060088263d71ab8b109339b101bb25d64a45e9c7de83710d8728`, port `243a04fe20939682a8ed127f07895fdf7dce610c30d9af0a6ea0b74a536202b9`. [b07-wide-b hashes; b12-ib-a hashes; b13-c1 gates.input_sha256; b11-fu-a notes]

IAD code in wide-a is `a4f2b410ffbcafaa6518c15674d060d719e13796da053e5efa49df7e1267f853`; wide-b/c and kit-b use `7f6998c8df1b738dbd9715053edae7c9c8fdad71bf7228d67b6f1b7573c7126b`. Exact fee/edge regressions, not identical source hashes, support their common gas-saving conclusion. The prescribed checkout is `6b6921f8d463314aa7c96234e1be25a25245f704`, solc 0.8.24, simulator wheel SHA-256 `aedb68280aca34c110968f00148387c95885207913fb2cf19a1ab4c5cfd6887a`. [b07-wide-a regression; b07-wide-b regression; b07-wide-c regression; b09-kit-b gates; b12-ib-b hashes]


[b06-gated-b]: ../../../../../.imd/reads/artifacts/b06-gated-b

[b06-gated-c]: ../../../../../.imd/reads/artifacts/b06-gated-c

[b07-wide-a]: ../../../../../.imd/reads/artifacts/b07-wide-a

[b07-wide-b]: ../../../../../.imd/reads/artifacts/b07-wide-b

[b07-wide-c]: ../../../../../.imd/reads/artifacts/b07-wide-c

[b08-joint-a]: ../../../../../.imd/reads/artifacts/b08-joint-a

[b08-joint-b]: ../../../../../.imd/reads/artifacts/b08-joint-b

[b08-const-a]: ../../../../../.imd/reads/artifacts/b08-const-a

[b09-kit-b]: ../../../../../.imd/reads/artifacts/b09-kit-b

[b10-inputs-clock]: ../../../../../.imd/reads/artifacts/b10-inputs-clock

[b10-inputs-path]: ../../../../../.imd/reads/artifacts/b10-inputs-path

[b10-inputs-flow]: ../../../../../.imd/reads/artifacts/b10-inputs-flow

[b10-inputs-fees]: ../../../../../.imd/reads/artifacts/b10-inputs-fees

[b10-inputs-cross]: ../../../../../.imd/reads/artifacts/b10-inputs-cross

[b11-fu-a]: ../../../../../.imd/reads/artifacts/b11-fu-a

[b11-fu-b]: ../../../../../.imd/reads/artifacts/b11-fu-b

[b11-fu-c]: ../../../../../.imd/reads/artifacts/b11-fu-c

[b12-ib-a]: ../../../../../.imd/reads/artifacts/b12-ib-a

[b12-ib-b]: ../../../../../.imd/reads/artifacts/b12-ib-b

[b12-fa-a]: ../../../../../.imd/reads/artifacts/b12-fa-a

[b12-fa-b]: ../../../../../.imd/reads/artifacts/b12-fa-b

[b13-c1]: ../../../../../.imd/reads/artifacts/b13-c1

[b13-c2]: ../../../../../.imd/reads/artifacts/b13-c2

[b13-c3]: ../../../../../.imd/reads/artifacts/b13-c3

[b13-c4]: ../../../../../.imd/reads/artifacts/b13-c4