No portfolio replacement: the five frozen 334 selections admitted zero post entries. The post cohort is [3 Oct 10:45, 4 Oct 15:00 UTC), with each member's full frozen horizon required before the cut. This is a static exact-composer readout; the interactive tape below remains the sealed 3 Oct in-sample/pre-window tape.
| Frozen portfolio | Post PnL USD | Max DD USD | Positive UTC days | Entries | Entry turnover USD | 3 Oct / 4 Oct PnL USD |
|---|---|---|---|---|---|---|
| A1 | +0.00 | 0.00 | 0/2 | 0 | 0 | +0.00 / +0.00 |
| operator_core_2 | +2.93 | 0.00 | 1/2 | 1 | 181 | +0.00 / +2.93 |
| arc_steady_15 | -1.34 | -28.82 | 1/2 | 5 | 683 | +24.55 / -25.89 |
The 4 Oct rollout policy requires an as-of score below 4 and blocks missing; the older policy blocks explicit BUNDLED verdicts and admits missing. Scores were published on 4 Oct, so both are counterfactual historical backfills.
| Portfolio | Unfiltered PnL / entries | Rollout PnL / entries | Older PnL / entries |
|---|---|---|---|
| A1 | +0.00 / 0 | +0.00 / 0 | +0.00 / 0 |
| operator_core_2 | +2.93 / 1 | +2.93 / 1 | +2.93 / 1 |
| arc_steady_15 | -1.34 / 5 | -1.34 / 5 | -5.78 / 4 |
The FINAL-free research-replica pull was rebuilt in a fresh lineage 7,377 seconds later: 556,422 swaps, 14,885 liquidity rows, 119,988 pools and 108,617 tokens matched row for row. All 29 frozen selected/comparator overlays matched 28,728 sealed pre-cut chosen legs. The unvalidated 4 Oct FINAL tape was not used.
Across those overlays only 41 mature post candidate legs appeared. A1, A2 and A6 had none; A3 and A5 had one each, the same j05 event, whose $6.12 10-minute participation support was below the composer minimum ($50 for A3, $25 for A5). Zero admitted PnL does not measure model quality. No selection met the registered 20-entry replacement gate. Frozen hook-cost coverage was also sparse: 1,073 of 1,143 hooked pools at the 4 Oct head lacked an empirical estimate. See the experiment report for limits and source receipts.
Selection is in-sample: 27 Sep 00:00–3 Oct 10:45 UTC exclusive on the validated Arc source. Pre-27 Sep is a holdout for this iteration. A validated post-window addendum appears above this frozen 3 Oct interactive tape. Pre-window books were previously screened in legacy research, so this is not an untouched model-universe test. All metrics use complete entry cohorts and the same composer settings.
| Portfolio | Period | PnL USD | Max DD USD | Positive days | Entries | Entry turnover USD |
|---|---|---|---|---|---|---|
| A1 | In-sample 27 Sep–3 Oct 10:45 | +1,339.84 | -61.44 | 86% | 65 | 13,257 |
| A1 | Pre-27 Sep holdout | +6,248.80 | -1,904.25 | 64% | 467 | 120,469 |
| operator_core_2 | In-sample 27 Sep–3 Oct 10:45 | +48.20 | -861.48 | 57% | 123 | 33,554 |
| operator_core_2 | Pre-27 Sep holdout | +18,802.34 | -3,235.36 | 100% | 764 | 345,489 |
| arc_steady_15 | In-sample 27 Sep–3 Oct 10:45 | -1,149.06 | -1,736.41 | 14% | 287 | 64,313 |
| arc_steady_15 | Pre-27 Sep holdout | +40,087.06 | -4,964.37 | 91% | 1855 | 538,111 |
Daily PnL is in UTC date order, including zero-entry days; the final selection day is partial. Values are USD.
A1 · selection: +343, +300, +433, +135, +85, +45, +0
A1 · pre: +1,366, +2,558, +1,957, -23, -705, -30, +418, -222, +169, +449, +311
operator_core_2 · selection: -31, +323, +376, +43, +170, -128, -706
operator_core_2 · pre: +3,909, +864, +5,604, +2,662, +499, +225, +1,847, +75, +781, +630, +1,707
arc_steady_15 · selection: -238, -2, +206, -26, -239, -46, -805
arc_steady_15 · pre: +15,459, -719, +3,741, +5,684, +2,188, +2,280, +2,944, +843, +3,101, +2,037, +2,529
| Member | j40 | j30 |
|---|---|---|
| j40 | 1.00 | 0.74 |
| j30 | 0.74 | 1.00 |
The 4 Oct rollout filter is a counterfactual historical backfill, not a claim it operated during this period. Its rule requires a causal t500/t1000/t1500 score <4 and blocks missing scores. The earlier explicit BUNDLED-only rule admits missing. This interactive tape is unfiltered; filter results below are exact Python settlements.
| Admission | Period | PnL USD | Entries | As-of score coverage |
|---|---|---|---|---|
| No bundle scorer | In-sample | +1,339.84 | 65 | 144/145 |
| No bundle scorer | Pre-window holdout | +6,248.80 | 467 | 509/834 |
| 4 Oct rollout: require score <4 | In-sample | +1,342.20 | 64 | 144/145 |
| 4 Oct rollout: require score <4 | Pre-window holdout | +5,844.90 | 301 | 509/834 |
| Earlier rule: explicit BUNDLED blocks | In-sample | +1,358.93 | 61 | 144/145 |
| Earlier rule: explicit BUNDLED blocks | Pre-window holdout | +4,907.03 | 405 | 509/834 |
f_secondary membership is frozen from RH. The blue band starts OOS at 25 Sep 00:00; the second marker is 28 Sep 02:00; tape cutoff is 3 Oct 10:45 UTC. The old 330d main is RETRACTED. Empirical hooked-pool costs fail closed.Scale multiplies each frozen RH leg's Arc notional. Sleeve cap limits that member's live capital; blank = none. Solid lines = all tokens, dashed = hooked-only.
Live trades are the real on-chain swaps the eight production Arc wallets made, read from the research replica and grouped into cells of (pool, UTC minute, side). Replacement estimates what would happen if this portfolio took their place.
On (replacement): in every cell, the first L dollars of our same-side flow (L = what the live fleet already bought, or sold, in that pool and minute) pay no new own impact — the price had already moved by that much in the history the backtest reads. Only impact above L is charged. Off (additive): our impact is charged in full on top of the live fleet's (the 005 convention) — that double counts the pool move the live fleet caused.
Limits: credit applies only to the same pool, minute and side. Opposite-side flow never offsets. Hooked-v4 pools in the corrected Arc books rarely overlap the live fleet, so the credit may be small. The fleet baseline is shown in the table below and is excluded from the main PnL chart.
Joint = this member's admitted legs inside the portfolio after Arc participation, crowding, caps and live replacement. Solo IS is the corrected frozen book at scale 1.