SirRoboGarage/docs/selector_negative_value.md

# The gun selector is currently NEGATIVE value

**Measured. The best single gun beats the full rack, and not by a little.**

| arm | runs | shots | hits | real % | dmg/run | per-run range | exact two-sided p vs `full` |
|---|---:|---:|---:|---:|---:|---|---:|
| `full` (shipped rack) | 7 | 3898 | 270 | **6.93** | 159 | 2.54–10.37 | — |
| **`onlyPattern`** | 7 | 4582 | 494 | **10.78** | **287** | 9.14–11.85 | **0.0012** |
| `onlyKNN` | 7 | 4033 | 207 | 5.13 | 119 | 4.15–6.47 | 0.1340 |
| `onlyLinear` | 7 | 3215 | 105 | 3.27 | 65 | 1.85–4.69 | 0.0082 |
| `onlyGF` | 7 | 3193 | 72 | 2.25 | 45 | 1.52–3.47 | 0.0012 |

**Firing `Pattern` alone: +3.85 pp pooled hit rate, +80% damage per run, p = 0.0012.**
It also fires MORE shots (4582 vs 3898), so it dominates on rate *and* volume.

This is **not** "any single gun wins" — `full` beats `onlyLinear`, `onlyGF` and `onlyKNN`.
It is specifically **"Pattern alone beats the rack"**.

## Follow-up: a small rack of GOOD guns STILL loses to `Pattern` alone

The obvious objection to the result above — "the rack is bloated, so of course it loses; prune it" — was
tested directly. Same protocol (one frozen binary, rack knobs only, real DrussGT, server-side real hit
rate), **6 arms × 7 runs × 7 rounds**. `lean8`/`lean6` are the offline audit's recommended racks, with the
selector ACTIVE; `onlyPattern` is the no-selection control.

| arm | guns (selector active?) | runs | shots | hits | real % | dmg/run | per-run range | exact two-sided p vs `onlyPattern` |
|---|---|---:|---:|---:|---:|---:|---|---:|
| **`onlyPattern`** (control) | Pattern, NO selection | 7 | 4325 | 448 | **10.36** | **264** | 8.96–11.47 | — |
| `lean8` | HeadOn, Linear, Circular, Accel, Pattern, GF, KNN, WallBounce | 7 | 3711 | 234 | 6.31 | 146 | 2.64–12.99 | 0.0169 |
| `lean6` | lean8 − HeadOn | 7 | 4113 | 363 | 8.83 | 212 | 7.35–10.87 | 0.0262 |
| `pairPC` | Pattern + Circular | 7 | 3862 | 318 | 8.23 | 185 | 3.36–11.64 | 0.0460 |
| `pairPK` | Pattern + KNN | 7 | 4571 | 448 | 9.80 | 264 | 7.90–12.76 | 0.3998 |
| `pairPL` | Pattern + Linear | 7 | 4109 | 338 | 8.23 | 200 | 6.14–9.90 | 0.0035 |

**No small rack beats `Pattern` alone.** Every selector-active arm is worse or (in one case) statistically
indistinguishable:

- **`lean8` is far worse** — 6.31% vs 10.36%, **p = 0.0169**. Pruning to the audit's recommended good guns
  did NOT rescue the selector.
- **`lean6` (HeadOn removed) is still worse** — 8.83% vs 10.36%, **p = 0.0262**. Removing the single
  most-over-selected gun helps (+2.5 pp over lean8) but still does not beat no-selection.
- **`pairPK` is the only near-tie** — 9.80% vs 10.36%, p = 0.3998, identical dmg/run. It ties only because
  the selector happens to pick `Pattern` **86.4%** of the time on a 2-gun rack; it does not *beat* Pattern.

### Verdict: the selector is negative value on a good rack too — DISABLE it

`lean6` and `lean8` (selector active) are both significantly WORSE than `Pattern` alone. This contradicts
the standing directive to keep the virtual-fitness selection mechanism, so it is stated plainly: **the
selection apparatus should be disabled (`TR_RACK_<every gun but PATTERN>=off`) pending a better fitness
signal.** Keeping it costs ~1.5–4 pp of real hit rate on a good rack.

### Why the pruned racks still lose: the same mis-ranking, visible in the mix

The selected-gun mix (liveness proof) shows the virtual signal still over-selects the wrong guns:

- **`lean8`**: `HeadOn` **46.2%** of ticks → only **2.0% real** (34/1670); `WallBounce` **21.0%** → 6.5%
  real (35/540); `Pattern` only **14.4%** → **12.6% real** (65/514). HeadOn and WallBounce crowd out the
  best gun.
- **`lean6`** (HeadOn gone): `Pattern` **29.2%** → **10.4% real** (127/1216) — the best real gun — while
  `KNN` 25.3% → 8.1%, `Linear` 15.2% → 8.0%, `Accel` 13.8% → 8.7%. The rack is still diluted.
- **`pairPL`**: the selector gives `Linear` **57.7%** of ticks → **6.0% real** (103/1712), and `Pattern`
  only 42.3% → **11.1% real**. Over-selecting Linear is why `pairPL` is the *worst* 2-gun arm (p = 0.0035).

The failure is not "the rack is too big". It is that the **virtual fitness signal ranks the wrong guns
first**, on a big rack and on a small one alike.

### Ship recommendation

**`onlyPattern`** — `Pattern` alone, selector bypassed. Measured: **10.36% real hit rate, 264 damage/run**
(vs `lean8` 6.31%/146 and `full` 6.93%/159). This replicates the original finding (10.78%/287).

Method note: one frozen binary built from clean `e0666a5` via `git archive` (other agents had
`common_libs/guns/*` dirty and have since committed further changes; this build predates them), sha256
`df8d4f2e…`. 6 arms × 7 runs × 7 rounds, 8 concurrent bridge battles. Liveness verified for every arm from
`gun_stats.jsonl`; the `[rack] active=…` line confirmed each arm's membership at process start.

## Method

One frozen binary built from clean `HEAD` via `git archive` (other agents had `common_libs/guns/*`
dirty), sha256 `f02481d8…`. Arms selected with the rack knobs only (`TR_RACK_<GUN>=off`), no source
edits. 5 arms × 7 runs × 7 rounds, 8 concurrent bridge battles, real DrussGT, judged ONLY on
server-side real hit rate from the events sidecar. Exact two-sided permutation test on per-run rates
(C(14,7)=3432 splits). Each arm's liveness verified from the selected-gun mix.

Harness preserved at `tools/ab/which_gun_*.sh` and `tools/ab/which_gun_analyze.py`.

## Why: the virtual fitness signal mis-ranks guns vs real outcomes

From the `full` arm's own selection mix and per-gun real rates:

- **`HeadOn` is massively over-selected** — **31.4% of ticks**, the most real shots (1070), but only
  **4.5% real**. It alone drags the rack down.
- **`Pattern`** has the best *virtual* rank and near-best *real* rate (11.9%, real rank 2), yet is
  selected only **22.6%** of the time.
- **`Linear`'s apparent strength was SELECTION BIAS.** Conditional on being selected it looked like
  15.2% (n=33); its *unconditional* rate (`onlyLinear`) is **3.27%**. Every earlier per-gun "real
  rate" in this repo is conditional on selection and is therefore confounded — this experiment is
  the clean measurement.

## What this does NOT yet settle

- **One adversary.** Everything here is vs DrussGT. `Pattern` should be re-checked against other
  bots before it becomes the default on the strength of this alone. (Supporting evidence: an offline
  audit found `Pattern` is the only gun competitive in *every* distance/speed bucket.)
- **Whether a SMALL good rack beats `Pattern` alone — SETTLED (follow-up above): NO.** A 6-arm
  follow-up (`lean8`, `lean6`, `pairPC`, `pairPK`, `pairPL`) found every selector-active rack worse or tied;
  `lean6` 8.83% and `lean8` 6.31% both lose to `Pattern` alone 10.36% (p = 0.026 / 0.017). The selector is
  negative value on a good rack too.
- **The user's directive was to KEEP the virtual-fitness selection mechanism — now directly tested.** The
  follow-up tested the selector on small, good racks instead of assuming either answer; it lost there too,
  so the directive conflicts with the evidence and disabling is recommended pending a better signal.

## Prior context: three failed selection-side attempts

| attempt | result |
|---|---|
| hysteresis (commit to incumbent) | 7.02% → 5.10%, p=0.002 |
| commitment (remove the random draw) | 7.17% → 4.44%, p=0.0012 |
| arrival-accuracy tie-break (rank by path, narrow by point) | 7.08%, p=0.88 — null |

So the per-tick random draw is load-bearing on three independent measurements, and no attempt to
"smarten" the tied band has helped. This experiment shows the problem is one level up: **which guns
are in the rack, and the fact that the virtual signal ranks them wrongly.**