SirStone/SirRoboGarage
1
0
Fork 0
Code Issues 28 Pull Requests Actions Projects Releases Wiki Activity

j127: allocation batch results - forced shares do not beat pattern; allocation not a lever (closed axis)

Browse Source
This commit is contained in:
Davide Cappellini
2026-09-26 10:45:10 +02:00
parent a436e9f16a
commit b4e54fc3cd
1 changed files with 123 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files
docs/gun_campaign.md
+123 -1
View File
@@ -36,6 +36,7 @@ controls.
| other single guns | `knn` detectably worse; `bitbrain` a wash measured 3× (−3.6 / +10.3 / +1.8 dmg/run, all inside their MDEs); `TMHorizon` worse on 33 (`docs/gauntlet_bitbrain_vs_pattern.md`, Batches 1–2) | | other single guns | `knn` detectably worse; `bitbrain` a wash measured 3× (−3.6 / +10.3 / +1.8 dmg/run, all inside their MDEs); `TMHorizon` worse on 33 (`docs/gauntlet_bitbrain_vs_pattern.md`, Batches 1–2) |
| 2-gun selector racks | `rack_pk` nominally below `pattern` on wins; `rack_pt` == `tmhorizon`-alone → the selector just picks the corrector (`## Batch 1`) | | 2-gun selector racks | `rack_pk` nominally below `pattern` on wins; `rack_pt` == `tmhorizon`-alone → the selector just picks the corrector (`## Batch 1`) |
| selector mechanics | the virtual-fitness selector is negative value at every rack size tested (16, lean8, lean6, pairs); it never manufactured a win (`docs/selector_negative_value.md`) | | selector mechanics | the virtual-fitness selector is negative value at every rack size tested (16, lean8, lean6, pairs); it never manufactured a win (`docs/selector_negative_value.md`) |
| **gun allocation (share)** | **forced `TR_RACK_SHARE` schedules (50/50, 70/30; PB/PT/PK) never beat `pattern` by the pre-registered rule; the only detectable effect is negative (50/50 KNN, −11 dmg/run, p=0.042); the observed ~66/34 split was already right (`## Allocation`) |
| lead amplitude | a full gain sweep (1.0/1.5/2.0/3.0) makes Pattern strictly worse at every range band — the lever is lead *information*, not amplitude (`docs/bitbrain_campaign.md` §0.3.2) | | lead amplitude | a full gain sweep (1.0/1.5/2.0/3.0) makes Pattern strictly worse at every range band — the lever is lead *information*, not amplitude (`docs/bitbrain_campaign.md` §0.3.2) |
| Pattern's own match-length / history params | `len6`'s single-session +0.49 wins (p=0.039 on 15) reversed to a wash on 33 (−0.04, p=1); `len16`/`depth100` flat; two batches, no replication (`## Phase 2`) | | Pattern's own match-length / history params | `len6`'s single-session +0.49 wins (p=0.039 on 15) reversed to a wash on 33 (−0.04, p=1); `len16`/`depth100` flat; two batches, no replication (`## Phase 2`) |
| radial knobs (`TR_PATTERN_RAD_*`) | bearing-invariant **by construction** (job j99 proved `bmPath` is a structural no-op and scale/offset cannot change the aim bearing); in Batch 2 `rad_offset` is nominally *negative* | | radial knobs (`TR_PATTERN_RAD_*`) | bearing-invariant **by construction** (job j99 proved `bmPath` is a structural no-op and scale/offset cannot change the aim bearing); in Batch 2 `rad_offset` is nominally *negative* |
@@ -1030,4 +1031,125 @@ different split beats `pattern` alone.** This batch builds the mechanism
### RESULTS ### RESULTS
_(filled in after the batch)_ **Session** `/tmp/ab/j127_alloc2`, commit `2a3a62a`, frozen binary sha256 `1b663dbbd365…`,
15 opponents × 6 arms × 3 runs × 3 rounds = **270 battles, 0 failed, 0 never started,
0 liveness exclusions** (827 s wall). Analyzer: `tools/ab/tournament_analyze.py`.
**Mechanism bug found and fixed before the ledger run (MEASURED).** The first
session (`/tmp/ab/j127_alloc`) exposed that `bot.tick` RESETS to 0 each round
while `tracker.currentSince` persists, so the dwell hold `(tick - currentSince) <
DWELL` went negative and LOCKED the round-1-end gun for every later round
(per-round applied share was `[50, 100, 100]` instead of `[50, 50, 50]`). The
share path now requires `tick >= currentSince` (commit `2a3a62a`); the shipped
ranking path is untouched. **The same latent lock exists in the shipped
selector's hysteresis for any multi-gun rack** — the `sel_pb` arm's per-round
share is likewise bimodal (e.g. `[98, 100]`, `[91, 0, 51]`), which is why its
aggregate ~69/31 is itself a round-boundary artifact, not a stable policy. This
is a MEASURED property of the shipped mechanism, not fixed here (no shipping).
**Applied share and per-gun per-shot hit rate (MEASURED, per-run `gun_stats.jsonl`).**
All 45 runs per forced arm print `TR_RACK_SHARE active`; 0 errors, 0 unrecognized.
The forced shares land on target (50.1/49.9, 70.0/30.0). Per-shot hit rate is
`realHits/realShots` summed over the same battles.
| arm | applied share (`selected` ticks) | shots | per-gun per-shot hit |
|---|---|---:|---|
| `pattern` | Pattern 100.0% | 7016 | Pattern **15.94%** |
| `sel_pb` (shipped policy) | Pattern 69.4% / BitBrain 30.6% | 4470 / 2024 | Pattern **19.08%** / BitBrain 12.85% |
| `share_5050_pb` | Pattern 50.1% / BitBrain 49.9% | 3034 / 3042 | Pattern **17.77%** / BitBrain 14.14% |
| `share_7030_pb` | Pattern 70.0% / BitBrain 30.0% | 4405 / 1877 | Pattern **17.68%** / BitBrain 16.36% |
| `share_5050_pt` | Pattern 50.1% / TMHorizon 49.9% | 3404 / 3407 | Pattern **18.18%** / TMHorizon 15.59% |
| `share_5050_pk` | Pattern 50.1% / KNN 49.9% | 3164 / 2633 | Pattern **17.98%** / KNN 12.53% |
**Per-opponent results (Δdmg/run / Δwins/run, arm − `pattern`).**
| opponent | `sel_pb` | `share_5050_pb` | `share_7030_pb` | `share_5050_pt` | `share_5050_pk` |
|---|---:|---:|---:|---:|---:|
| DrussGT | +26.5 / +0.67 | -36.8 / -0.67 | -10.3 / -0.33 | -1.1 / -0.33 | -5.1 / +0.33 |
| Diamond | +7.6 / +0.33 | -4.1 / -0.33 | -15.0 / +0.00 | +12.2 / -0.33 | -6.3 / -0.33 |
| Dookious | +5.0 / +0.00 | -12.6 / -0.33 | -13.7 / -0.33 | -12.9 / -0.67 | -13.9 / +0.00 |
| GresSuffurd | +29.7 / +0.33 | -3.0 / -0.33 | +16.5 / -0.67 | +23.6 / +0.00 | -11.4 / -0.67 |
| CassiusClay | +25.7 / +0.33 | +33.0 / +0.33 | +21.4 / +0.67 | +29.5 / +0.67 | -7.5 / +0.00 |
| RetroGirl | +13.5 / +1.33 | -19.6 / +1.00 | +7.7 / +0.67 | +73.7 / +2.00 | +20.4 / +0.67 |
| TripHammer | +12.9 / +0.33 | +1.1 / +0.33 | +4.8 / +0.67 | -6.8 / +0.33 | +1.2 / +0.00 |
| Coriantumr | -1.9 / -0.33 | -25.7 / -0.67 | -16.9 / +0.33 | -13.3 / -1.33 | -44.9 / -1.67 |
| WallAvoider | -31.8 / +1.00 | +28.9 / +1.33 | +9.0 / +1.00 | -8.2 / +1.00 | -27.7 / -0.33 |
| HawkOnFire | +0.2 / -1.33 | +4.7 / -0.33 | +37.6 / +0.00 | -9.3 / -0.33 | -41.2 / -2.00 |
| SpinBot | -14.7 / +0.00 | +12.7 / +0.00 | +6.0 / +0.00 | +16.0 / +0.00 | -15.3 / +0.00 |
| DiamondStealer | +27.4 / +0.00 | +16.0 / +0.67 | -24.1 / -0.33 | -4.3 / +0.00 | +8.9 / +0.67 |
| BlitzBat | +3.5 / +0.33 | +26.3 / +0.00 | +10.2 / +0.33 | +11.1 / +0.33 | +10.6 / +0.00 |
| YersiniaPestis | -18.5 / -1.33 | -15.9 / -0.67 | -11.9 / -0.67 | +4.4 / -0.33 | -0.3 / -1.00 |
| Ascendant | -0.0 / +0.00 | -5.3 / -0.33 | +31.4 / +1.67 | -6.5 / +0.33 | -33.1 / -0.33 |
**Cross-opponent tests/CI/MDE vs `pattern` (the verdict layer).** Primary metrics
only; `p` is the exact sign-flip permutation on the per-opponent mean.
| arm | metric | mean Δ | 95% CI | sign test | p(sign-flip) | MDE | verdict |
|---|---|---:|---|---:|---:|---:|---|
| `sel_pb` | damage | +5.68 | [-4.29, +15.64] | 10/15 | 0.2393 | 13.01 | not distinguishable |
| `sel_pb` | wins | +0.11 | [-0.29, +0.51] | 8/11 | 0.6387 | 0.52 | not distinguishable |
| `share_5050_pb` | damage | -0.02 | [-11.41, +11.36] | 7/15 | 0.9968 | 14.87 | not distinguishable |
| `share_5050_pb` | wins | -0.00 | [-0.34, +0.34] | 5/13 | 1 | 0.45 | not distinguishable |
| `share_7030_pb` | damage | +3.50 | [-6.73, +13.74] | 9/15 | 0.4725 | 13.37 | not distinguishable |
| `share_7030_pb` | wins | +0.20 | [-0.16, +0.56] | 7/12 | 0.3193 | 0.47 | not distinguishable |
| `share_5050_pt` | damage | +7.21 | [-5.37, +19.80] | 7/15 | 0.2582 | 16.44 | not distinguishable |
| `share_5050_pt` | wins | +0.09 | [-0.34, +0.52] | 6/12 | 0.7539 | 0.56 | not distinguishable |
| `share_5050_pk` | damage | -11.03 | **[-21.54, -0.52]** | 4/15 | **0.04205** | 13.72 | **WORSE (damage)** |
| `share_5050_pk` | wins | -0.31 | [-0.73, +0.11] | 3/10 | 0.1777 | 0.55 | not distinguishable |
**Policy isolation — forced arms vs `sel_pb` (MEASURED, `--reference sel_pb`).**
This is the arm that holds gun choice fixed and asks whether the POLICY (ranking
vs forced schedule) changes anything. `share_7030_pb` deliberately reproduces
`sel_pb`'s applied share (70.0/30.0 vs 69.4/30.6), so it is the cleanest test.
| arm | metric | mean Δ | 95% CI | sign test | p(sign-flip) | MDE |
|---|---|---:|---|---:|---:|---:|
| `share_7030_pb` | damage | -2.17 | [-16.90, +12.55] | 6/15 | 0.7542 | 19.23 |
| `share_7030_pb` | wins | +0.09 | [-0.34, +0.52] | 6/12 | 0.7432 | 0.56 |
| `share_5050_pb` | damage | -5.70 | [-21.93, +10.54] | 6/15 | 0.4659 | 21.20 |
| `share_5050_pb` | wins | -0.11 | [-0.44, +0.22] | 4/12 | 0.5771 | 0.43 |
| `share_5050_pt` | damage | +1.54 | [-12.09, +15.16] | 7/15 | 0.8182 | 17.80 |
| `share_5050_pt` | wins | -0.02 | [-0.37, +0.33] | 5/10 | 1 | 0.46 |
| `share_5050_pk` | damage | -16.71 | **[-27.92, -5.50]** | 4/15 | **0.00769** | 14.64 |
| `share_5050_pk` | wins | -0.42 | **[-0.73, -0.11]** | 2/13 | **0.01733** | 0.40 |
Pooled dashboard (descriptive, not the verdict): `pattern` 101.9 dmg/1.47 wins;
`sel_pb` 107.5/1.58; `share_5050_pb` 101.8/1.47; `share_7030_pb` 105.4/1.67;
`share_5050_pt` 109.1/1.56; `share_5050_pk` 90.8/1.16.
### VERDICT
> **DIRECT ANSWER (MEASURED): NO — gun ALLOCATION is not a lever on the frozen
> 15-opponent panel at this resolution. No forced-share arm beats `pattern`
> alone by the pre-registered rule (CI excluding 0 AND sign-flip p<0.05). The
> selector's ~66/34 split was already right within the MDE; the rack question is
> genuinely CLOSED.**
* The best forced arm is `share_7030_pb` (applied 70/30, i.e. re-creating the
observed ~66/34 split deliberately): **+0.20 wins/run, +3.5 dmg/run vs
`pattern`, both inside their MDEs** (0.47 wins, 13.4 dmg), sign-flip p=0.32 /
0.47. A nominal, non-significant lean toward the incumbent's own split.
* The deliberate 50/50 arms gain nothing (`share_5050_pb` ≈ 0/0) or cost damage
(`share_5050_pt` +7.2 dmg but +0.09 wins, not distinguishable).
* The only DETECTABLE allocation effect is NEGATIVE: forcing 50/50 with KNN is
detectably WORSE on damage vs `pattern` (−11.0, CI excludes 0, p=0.042) and
detectably worse than `sel_pb` on BOTH wins and damage (−0.42 wins, p=0.017;
−16.7 dmg, p=0.008). Allocation can HURT; it cannot help here.
* **`policy` vs `gun choice` (MEASURED).** Forcing the selector's own applied
share (`share_7030_pb` ≈ `sel_pb`) is not distinguishable on either primary,
so at the same allocation the ranking-vs-forced *policy* makes no measurable
difference. `sel_pb` itself is nominally positive (+0.11 wins, +5.7 dmg) but
not distinguishable — this session's reference is a low one (48.9% wins),
consistent with the campaign's known session-level reference variance.
* **Caveat that strengthens the finding (MEASURED).** The applied share of the
*shipped* selector is not a stable policy at all: its round-boundary dwell
lock makes per-round shares bimodal (a gun can hold 100% of a round). The
forced arms, with the fix, are the first clean allocation measurement — and
they show no upside to deviating from Pattern-dominant.
**MEASURED:** the allocator code + default-OFF parity (39/48/19/25), the
round-boundary bug and its fix, all 4 forced arms' applied shares (exactly
50/50 and 70/30) and per-gun per-shot hit rates, the 270-battle session, every
per-opponent delta, CI, MDE, sign and sign-flip test, and the policy-isolation
comparison. **INFERRED:** that no share schedule whatsoever could ever help
(this batch bounds only the shares it tested, at this MDE and panel).