# Config Watch #1

**WEEKLY · CONFIG WATCH**  ·  [NEW SERIES -- Issue #1]

Issue date: August 13, 2026  ·  Language: English  ·  Methodology v1.1 + Config Watch amendments (owner-approved 12.08, hash e66c7e8c864a2233)

Search windows: **7d** Aug 5-12, 2026 and **30d** Jul 13-Aug 12, 2026 (frozen date ranges).

*Both windows are frozen date ranges, not the Mon-Mon reporting window: issue #1 ran ad-hoc. From issue #2 onward, windows align to Mondays.*

---

## Snapshot

- **Windows:** 7d = Aug 5-12, 2026; 30d = Jul 13-Aug 12, 2026 (both frozen; not yet the Mon-Mon reporting window -- see note above)
- **Scale:** 12,199 simulations total -- 7d branch 7,307 (600+2,160+4,547) + 30d branch 4,892 (600+2,172+2,120)
- **Search stages:** council scan -> skeleton grid -> coordinate-ascent waves (stage design: owner)
- **Engine economics (fixed across all sims):** $1,000 deposit, 10x leverage, position mode, risk-budget stake table, fees modeled, liquidation modeled by the engine
- **Publication/selection rules for TOP-3:** at least 2 coins, pairwise different universes across the three picks, deduplicated trade streams; single-coin bests are reported separately as a side note, not ranked in the TOP-3


---

## KEY FINDING

> **[OBSERVATION]** Search found configurations that materially outperformed the default setup in-sample; whether that edge survives starts next week.

---

## TL;DR

- OBSERVATION -- Search found configurations that materially outperformed the default setup in-sample on both windows; whether that edge survives out-of-sample starts next week.
- Defaults lost on both windows: 7d grid median -$99.4 (24.7% positive); 30d grid median -$221.0 (2.4% positive). All 7 solo-model defaults and the council-of-7 default were negative on both windows.
- Top result per window beat the best hand-tuned pre-search config: +$439.2 (7d) and +$868.4 (30d) -- search validation passed against independent, pre-search baselines.
- Cross-window robustness check: most tuned configs are fragile across windows; one family (fable-5+deepseek-v4-pro+opus-5 core, 4h/1h, ladder+BE, conf>=60, farthest source) showed a cross-window robustness hint -- this is not independent validation. Real out-of-sample testing starts next week.

---

## Stage-1b grid: where the defaults sit (7d)

![7d stage-1b grid net PnL histogram](hist_7d.png)

*Distribution of net PnL across all 2,160 stage-1b systematic-grid simulations on the 7d window. Grid median -$99.4, 24.7% positive. Markers show the best default and the three published TOP-3 configs (all from the waves stage, which searches beyond this grid -- see percentile note on each card). The matching 30d histogram appears later in this report, in the TOP-3 30d section.*

---

## Why it matters

MarketMania publishes default LLM-council trading configs. This series asks a narrower, more useful question every week: how much apparent performance can large-scale in-sample search extract -- and how much of it survives out-of-sample? Issue #1 establishes the in-sample (IS) baseline: search scale, what wins, and a first, honest look at whether winners on one window still win on another. It is not yet an out-of-sample (OOS) verdict -- that begins next week.

---

## How we searched

Three stages per window, run independently for 7d and 30d:

1. **Council scan (stage-1a)** -- 600 sims/branch across council compositions and coarse settings.
2. **Skeleton grid (stage-1b)** -- a systematic grid (2,160 configs on 7d; 2,172 on 30d) sweeping declared axes: archetype (ladder / pure / filtered), forecast horizon vs. timeframe, entry (max-sideways / max-diff-side), universe size, council size, and model membership. This grid is the denominator for every percentile claim in this report.
3. **Coordinate-ascent waves** -- local refinement beyond the grid (4,547 sims on 7d; 2,120 on 30d), producing the published TOP-3 and the solo-coin side note.

Stage design: owner. All fixed economics (deposit, leverage, stake table, fees, liquidation modeling) are held constant across every sim in every stage.

*Models under test (exact engine IDs): claude-fable-5, claude-opus-5, gpt-5.6-sol, deepseek-v4-pro, gemini-3.1-pro, qwen-3.8-max, grok-4.5.*

---

## TOP-3 -- 7d window (Aug 5-12, 2026)

| # | Council | Coins | FH/TF | Net | Net % | maxDD | Trades | WR |
|---|---|---|---|---|---|---|---|---|
| #1 | fable-5, deepseek-v4-pro, opus-5, gpt-5.6-sol | SOL+XRP | 4h/1h | **+$439.2** | +43.9% | 5.0% | 32 | 68% |
| #2 | fable-5, qwen-3.8-max, gemini-3.1-pro, opus-5 | XRP+BNB | 4h/4h | **+$429.5** | +43.0% | 8.0% | 19 | 76% |
| #3 | gpt-5.6-sol, opus-5 | BNB+XRP+SOL | 4h/1h | **+$398.6** | +39.9% | 0.4% | 21 | 85% |

| # | Ladder | BE | Shifts | Conf filter | Source | Entry | Positional | Stake |
|---|---|---|---|---|---|---|---|---|
| #1 | 50:40, 100:60 | on | TP -10 / SL +50 | >=60 (each) | farthest | 4/0 | reenter | $450 |
| #2 | 100:100 | off | TP -10 / SL +40 | >=50 (each) | median | 4/2 | keep | $450 |
| #3 | 80:50, 100:50 | off | TP +10 / SL +20 | >=60 (each) | nearest | 1/1 | keep | $300 |

*All three are waves-stage results -- best of 7,307 sims in the 7d branch; better than 99.0% / 99.0% / 98.4% of the 2,160-config stage-1b grid (#1/#2/#3; the waves stage searches beyond that grid). Grid median -$99.4.*

*Net % = net PnL as a percentage of the fixed $1,000 starting deposit; every simulation runs at **10x leverage** (liquidation modeled by the engine).*

*Reproduce in sandbox: **marketmania.ai/s/cw1-7d1** · **marketmania.ai/s/cw1-7d2** · **marketmania.ai/s/cw1-7d3** -- each link opens this exact frozen window; results visible without sign-in (embedded share signature).*

![7d TOP-3 equity](equity_7d.png)

*Daily settled-equity curves for the three councils above -- $1,000 start, 10x leverage, frozen window. Ranks match the table; each step is that day's settled trades; a curve running past the window end is an open-at-cutoff trade settling at its horizon.*

---

## TOP-3 -- 30d window (Jul 13-Aug 12, 2026)

| # | Council | Coins | FH/TF | Net | Net % | maxDD | Trades | WR |
|---|---|---|---|---|---|---|---|---|
| #1 | fable-5, deepseek-v4-pro, opus-5, qwen-3.8-max | ETH+SOL | 4h/1h | **+$868.4** | +86.8% | 9.4% | 42 | 80% |
| #2 | gemini-3.1-pro, gpt-5.6-sol, deepseek-v4-pro | SOL+XRP+ETH | 1d/1d | **+$590.6** | +59.1% | 11.0% | 24 | 83% |
| #3 | fable-5, deepseek-v4-pro, opus-5, qwen-3.8-max | BTC+ETH+SOL | 4h/1h | **+$580.5** | +58.1% | 12.8% | 57 | 78% |

| # | Ladder | BE | Shifts | Conf filter | Source | Entry | Positional | Stake |
|---|---|---|---|---|---|---|---|---|
| #1 | 50:80, 100:20 | on | TP +10 / SL +35 | >=60 (each) | farthest | 1/1 | update | $450 |
| #2 | 40:80, 100:20 | on | TP +30 / SL +20 | >=60 (each) | nearest | 2/0 | update | $300 |
| #3 | 50:80, 100:20 | on | TP +10 / SL +35 | >=60 (each) | farthest | 1/1 | update | $300 |

*All three are waves-stage results -- best of 4,892 sims in the 30d branch; better than 99.9% / 99.6% / 99.6% of the 2,172-config stage-1b grid (#1/#2/#3; the waves stage searches beyond that grid). Grid median -$221.0.*

*Net % = net PnL as a percentage of the fixed $1,000 starting deposit; every simulation runs at **10x leverage** (liquidation modeled by the engine).*

*#2: time_stop = 3x (only non-default extra on this config).*

*Reproduce in sandbox: **marketmania.ai/s/cw1-30d1** · **marketmania.ai/s/cw1-30d2** · **marketmania.ai/s/cw1-30d3** -- each link opens this exact frozen window; results visible without sign-in (embedded share signature).*

![30d TOP-3 equity](equity_30d.png)

*Daily settled-equity curves for the three councils above -- $1,000 start, 10x leverage, frozen window. Ranks match the table; each step is that day's settled trades; a curve running past the window end is an open-at-cutoff trade settling at its horizon.*

---

## Stage-1b grid: where the defaults sit (30d)

![30d stage-1b grid net PnL histogram](hist_30d.png)

*Distribution of net PnL across all 2,172 stage-1b systematic-grid simulations on the 30d window. Grid median -$221.0, only 2.4% positive -- a materially harder window than 7d.*

---

## Cross-window robustness check

Each TOP-3 config, replayed on the *other* window's history, using the exact same config values (no re-tuning). This checks whether a winner on one slice of history is still a winner on an overlapping, longer slice -- a same-history consistency check, not an out-of-sample test.

| Config | Family | Home window (net) | Other window (net) | Trades (other) | Verdict |
|---|---|---|---|---|---|
| 7d#1 | fable-5, deepseek-v4-pro, opus-5, gpt-5.6-sol | 7d: +$439.2 | 30d: +$178.8 | 108 | Held up |
| 7d#2 | fable-5, qwen-3.8-max, gemini-3.1-pro, opus-5 | 7d: +$429.5 | 30d: -$551.3 | 16 | Collapsed |
| 7d#3 | gpt-5.6-sol, opus-5 | 7d: +$398.6 | 30d: -$76.3 | 79 | Collapsed |
| 30d#1 | fable-5, deepseek-v4-pro, opus-5, qwen-3.8-max | 30d: +$868.4 | 7d: -$21.8 | 13 | Collapsed |
| 30d#2 | gemini-3.1-pro, gpt-5.6-sol, deepseek-v4-pro | 30d: +$590.6 | 7d: +$223.1 | 5 | Insufficient (n=5) |
| 30d#3 | fable-5, deepseek-v4-pro, opus-5, qwen-3.8-max | 30d: +$580.5 | 7d: -$11.5 | 15 | Collapsed |

**Honest read:** most tuned configs are fragile across windows -- 7d#2 and 7d#3 collapsed on 30d, and 30d#1 and 30d#3 went mildly negative on 7d. One family did not collapse in either direction: the **fable-5+deepseek-v4-pro+opus-5 core, 4h/1h, ladder+BE, conf>=60, farthest source** (7d#1 approx-equals the 30d#1/#3 family). 7d#1 held up when run on the 30d window (+$178.8, 108 trades); the 30d#1/#3 family only mildly underperformed on the 7d window (-$21.8, 13 trades). 30d#2 looked positive on 7d (+$223.1) but on only 5 trades -- flagged insufficient, not evidence either way. This is a same-history overlap check, **not** an out-of-sample (OOS) test -- both windows are in-sample and the 30d window contains the 7d window. Real OOS starts next week.

---

## Secondary analysis -- solo-model top (dedicated search)

**Sidebar to the main council-robustness narrative.** The council search never contained single-model configs (councils were always >=2 models), so the solo question got its own dedicated run: the same three-stage design (840-sim systematic grid + coordinate-ascent waves per branch), same frozen windows, same fixed economics, same publication gates (>=2 coins; >=15 trades on 7d, >=20 on 30d). One config per model; the ranks below are the best three *models*. From issue #2 the solo search runs alongside the council search every week.

### Solo top-3 -- 7d window (Aug 5-12, 2026)

| # | Model | Coins | FH/TF | Net | Net % | maxDD | Trades | WR |
|---|---|---|---|---|---|---|---|---|
| #1 | gemini-3.1-pro | XRP+BNB | 1h/1h | **+$417.8** | +41.8% | 14.9% | 39 | 54% |
| #2 | gpt-5.6-sol | XRP+BNB | 4h/1h | **+$366.3** | +36.6% | 0.0% | 30 | 87% |
| #3 | qwen-3.8-max | XRP+BNB | 4h/4h | **+$267.0** | +26.7% | 2.3% | 30 | 57% |

| # | Ladder | BE | Shifts | Conf filter | Source | Entry | Positional | Stake |
|---|---|---|---|---|---|---|---|---|
| #1 | 100:100 | off | TP +10 / SL +20 | >=0 (each) | median | 2/1 | keep | $450 |
| #2 | 30:40, 100:60 | on | TP +10 / SL +50 | >=0 (each) | median | 2/1 | update | $450 |
| #3 | 100:100 | off | TP +10 / SL -10 | >=0 (each) | median | 2/1 | update | $450 |

*Net % = net PnL as a percentage of the fixed $1,000 starting deposit; every simulation runs at **10x leverage** (liquidation modeled by the engine).*

*#3: time_stop = 3x (only non-default extra on this config).*

*Reproduce in sandbox: **marketmania.ai/s/cw1-7ds1** · **marketmania.ai/s/cw1-7ds2** · **marketmania.ai/s/cw1-7ds3** -- each link opens this exact frozen window; results visible without sign-in (embedded share signature).*

![7d solo top equity](equity_solo_7d.png)

### Solo top-3 -- 30d window (Jul 13-Aug 12, 2026)

| # | Model | Coins | FH/TF | Net | Net % | maxDD | Trades | WR |
|---|---|---|---|---|---|---|---|---|
| #1 | opus-5 | SOL+XRP | 4h/1h | **+$442.8** | +44.3% | 31.7% | 67 | 68% |
| #2 | grok-4.5 | SOL+XRP | 4h/1h | **+$262.9** | +26.3% | 28.0% | 47 | 70% |
| #3 | deepseek-v4-pro | XRP+BNB | 4h/1h | **+$119.0** | +11.9% | 71.4% | 54 | 62% |

| # | Ladder | BE | Shifts | Conf filter | Source | Entry | Positional | Stake |
|---|---|---|---|---|---|---|---|---|
| #1 | 60:50, 100:50 | on | TP -10 / SL +30 | >=55 (each) | median | 2/1 | update | $450 |
| #2 | 50:50, 100:50 | on | TP -10 / SL +0 | >=60 (each) | median | 2/1 | update | $450 |
| #3 | 80:50, 100:50 | on | TP +10 / SL +50 | >=55 (each) | median | 2/1 | keep | $450 |

*Net % = net PnL as a percentage of the fixed $1,000 starting deposit; every simulation runs at **10x leverage** (liquidation modeled by the engine).*

*Reproduce in sandbox: **marketmania.ai/s/cw1-30ds1** · **marketmania.ai/s/cw1-30ds2** · **marketmania.ai/s/cw1-30ds3** -- each link opens this exact frozen window; results visible without sign-in (embedded share signature).*

![30d solo top equity](equity_solo_30d.png)

**Council vs. solo:** on 7d the best solo (+$417.8) beats the council #3 but stays under the council #1 (+$439.2); on 30d every solo finalist sits below the council TOP-3 -- the council keeps the #1 spot on both windows, but a well-tuned solo model is competitive with the council mid-ranks. Same in-sample caveats apply.

---

## Baselines

Solo-model and council-of-7 defaults, plus three hand-tuned configs that were saved *before* this search began (independent baselines, not derived from the search).

### 7d (Aug 5-12, 2026)

| Baseline | Net PnL | Trades | WR | maxDD | PF |
|---|---|---|---|---|---|
| Best solo default (gemini-3.1-pro) | -$14.4 | 20 | 40% | 2.2% | 0.93 |
| Worst solo default (grok-4.5) | -$329.1 | 32 | 14% | 33.4% | 0.27 |
| Council-of-7 default | -$164.9 | 34 | 26% | 17.3% | 0.59 |
| Hand-tuned config A (pre-search) | +$137.3 | 7 | *80% (N=7)* | 0.0% | *51.07 (N=7)* |
| Hand-tuned config B (pre-search) | +$295.9 | 27 | 84% | 1.6% | 2.84 |
| Hand-tuned config C (pre-search) | -$21.8 | 13 | 67% | 10.3% | 0.92 |

*WR and PF italicized with (N=X) where fewer than 10 trades closed in the window -- insufficient sample; shown for completeness, not comparison.*

### 30d (Jul 13-Aug 12, 2026)

| Baseline | Net PnL | Trades | WR | maxDD | PF |
|---|---|---|---|---|---|
| Best solo default (fable-5) | -$432.5 | 99 | 30% | 43.6% | 0.58 |
| Worst solo default (grok-4.5) | -$770.1 | 65 | 14% | 77.0% | 0.26 |
| Council-of-7 default | -$693.5 | 63 | 16% | 69.3% | 0.26 |
| Hand-tuned config A (pre-search) | +$240.1 | 16 | 79% | 9.3% | 2.13 |
| Hand-tuned config B (pre-search) | +$112.5 | 92 | 71% | 36.1% | 1.10 |
| Hand-tuned config C (pre-search) | +$793.7 | 42 | 78% | 10.3% | 2.41 |

Search validation: the TOP-3 headline beat the best hand-tuned pre-search config on both windows -- +$439.2 vs +$295.9 (7d); +$868.4 vs +$793.7 (30d).

---

## Patterns: what the top decile looks like

Top decile vs. the full stage-1b pool (7d: 43 of 434 candidates; 30d: 62 of 623). For each axis, the value shown is the one most overrepresented in the top decile relative to the full pool (descriptive only -- not a causal claim).

| Axis | 7d: overrepresented value | 7d: top-decile vs grid | 30d: overrepresented value | 30d: top-decile vs grid |
|---|---|---|---|---|
| Archetype | ladder | 58.1% (all 38.9%) | ladder | 51.6% (all 38.0%) |
| Forecast horizon / timeframe | 1d/4h | 46.5% (all 7.8%) | 4h/1h | 98.4% (all 59.9%) |
| Entry (max-sideways / max-diff-side) | 4/2 | 44.2% (all 40.3%) | 0/0 | 38.7% (all 24.4%) |
| Universe size (coins considered) | 5 | 81.4% (all 59.0%) | 1 | 93.5% (all 20.9%) |
| Council size (models in ensemble) | 2 | 46.5% (all 22.1%) | 5 | 51.6% (all 30.3%) |
| Model membership | qwen-3.8-max | 79.1% (all 50.7%) | qwen-3.8-max | 62.9% (all 36.8%) |

**Winning knobs differ by window** (waves top-20, i.e. the 20 best waves-stage results per branch):

| Knob | 7d top-20 mode | 30d top-20 mode |
|---|---|---|
| Ladder shape (steps) | 100:100 (20/20) | 50:40,100:60 (18/20) |
| Break-even stop (BE) | 0 (19/20) | 1 (20/20) |
| Min confidence | 50 (18/20) | 0 (8/20) |
| SL shift | -10 (18/20) | 50 (16/20) |
| TP/SL source | nearest (15/20) | median (20/20) |

7d top-20 favors a single rung (100:100), min-conf 50, SL shift -10, nearest source. 30d top-20 favors a 2-rung ladder, break-even on, SL shift +40..+50, median source. This is an honest observation against a one-size-fits-all config -- it is descriptive, not a recommendation to run either knob set live.

---

## Solo-coin side note (excluded from TOP-3 by the >=2-coin rule)

- **7d:** fable-5+qwen-3.8-max+gemini-3.1-pro on XRP alone scored +$618.4 (16 trades, WR 67%, maxDD 13.5%) -- higher than every published 7d TOP-3 config, but excluded because the publication rule requires >=2 coins with pairwise different universes.
- **30d:** opus-5+fable-5+gemini-3.1-pro+qwen-3.8-max+grok-4.5 on SOL alone scored +$1,225.3 (22 trades, WR 90%, maxDD 27.9%) -- the single highest score in the entire 30d search, excluded for the same reason.

Both are flagged as **idiosyncratic-coin risk**: a single-coin config's edge (or lack of one) can be dominated by that one coin's specific path during the window, which is a weaker generalization claim than a multi-coin, multi-universe result.

---

## Liquidation note

Liquidation is modeled by the simulation engine at the fixed 10x leverage used throughout this search. Across the published configs, observed stop-loss placements stay well inside the distance that would approach the liquidation threshold. Max drawdown (maxDD) is reported on every config in this issue so realized risk is visible directly, rather than relying on this note alone.

---

## Walk-forward

This is an **in-sample (IS)** issue: every number above comes from the same historical windows the search optimized against. The out-of-sample (OOS) fate of all six TOP-3 configs (three per window) becomes a **standing section starting issue #2**. Tracking a decay series over 5-10 weeks -- does performance hold, fade, or reverse -- is the core research question of this series: how robust (or overfit) are LLM-ensemble configs found by large-scale in-sample search?

---

## Limitations

- **Multiple testing:** ~15,416 simulations were searched (council + dedicated solo); with a search this large, some winners are expected by chance alone. Antidotes used in this issue: the full stage-1b histogram is shown (not just the winners), search axes are declared, independent hand-tuned pre-search baselines are included, in-sample (IS) is labeled throughout, and a walk-forward plan starts next issue. Issue #1 does not apply a formal multiple-testing correction (see What we're testing next).
- **Single regime:** both windows sit inside one market regime; no regime-change evidence yet.
- **Small trade counts:** several cross-window reads run on as few as 5-16 trades -- directional, not conclusive.
- Research output, not financial advice.

---

## Counters & lineage

| Branch | Council scan | Skeleton grid | Waves | Total |
|---|---|---|---|---|
| 7d (Aug 5-12, 2026) | 600 | 2,160 | 4,547 | 7,307 |
| 30d (Jul 13-Aug 12, 2026) | 600 | 2,172 | 2,120 | 4,892 |
| **Total (council search)** | | | | **12,199** |
| 7d solo search (dedicated) | -- | 840 | 812 | 1,652 |
| 30d solo search (dedicated) | -- | 840 | 725 | 1,565 |

*Solo-search sims are counted separately from the council-search total above.*

Generated at: 2026-08-13T04:32:27Z  ·  Methodology v1.1 + Config Watch amendments (owner-approved 12.08), hash `e66c7e8c864a2233`.

**Snapshot principle:** this report is a frozen snapshot of the search taken at the generated-at timestamp above; numbers are not updated retroactively. Each week's issue re-runs the full 3-stage pipeline fresh over that week's windows.

---

## Research to date

> **THIS REPORT** -- council search: 12,199 sims · solo search: 3,217 sims · total search effort: **15,416 simulations**
>
> **MARKETMANIA RESEARCH TO DATE (as of August 13, 2026, 06:57 UTC):** 59,253 resolved forecasts since Jul 11 · 27,341 closed simulated trades · 9 models tracked · 5 assets · 5 horizons · hourly cadence · 5 published research reports
>
> *Resolved = horizon expired. Counting rule tightened in this issue; not comparable with the 14,151+ figure quoted in earlier issues. Source: platform_counters_2026-08-13*

**[NEW SERIES -- Issue #1]** Config Watch joins MarketMania's published research line.

---

## What we're testing next

- Next issue: the OOS fate of all six configs above -- does the surviving family keep its edge?
- Council vs. solo OOS: does the ensemble advantage persist after tuning?
- Monthly: formal multiple-testing correction once several OOS windows accumulate.

---

## Related research

- Stability Index Run 2 -- marketmania.ai/research/reports/si-run-2.pdf
- Weekly Calibration #1 -- marketmania.ai/research/reports/weekly-calibration-2026-08-03.pdf
- Weekly Model Watch #1 -- marketmania.ai/research/reports/model-watch-2026-08-03.pdf
- Consensus Watch #1 -- marketmania.ai/research/reports/consensus-watch-2026-08-03.pdf

---

## Cite this report

```bibtex
@techreport{mm_config_watch_2026w33,
  title        = {Config Watch #1},
  author       = {{MarketMania Research}},
  institution  = {MarketMania},
  year         = {2026},
  month        = aug,
  day          = {13},
  type         = {Weekly Research Report},
  series       = {Config Watch},
  number       = {1},
  note         = {Methodology v1.1 + Config Watch amendments, hash e66c7e8c864a2233},
  url          = {https://marketmania.ai/research/reports/config-watch-2026-08-05.pdf}
}
```