E1 COMPLETE, two seeds: tie holds (six runs .6007-.6047, band .0040); a3-s1 .6047 with toggle -.0524 (largest routed function); full tables folded (endpoint/per-task/decomposition/basin x8)

README.md

@@ -59,14 +59,14 @@ training-side numbers are each run's `checkpoints/metrics.json` finals.
 | `a2_dense-s0` | **.6026** | — | — | .9975 | .8418 | 99.2 |
 | `a3_random-s0` | **.6033** | .5772 | −.0261 | .9980 | .8392 | 98.8 |
 | `a2_dense-s1` | **.6040** | — | — | .9975 | .8420 | 99.2 |
-| `a3_random-s1` | *training — final epoch* | | | | | |
+| `a3_random-s1` | **.6047** | .5523 | −.0524 | .9980 | .8396 | 98.6 |
 
-Three routing designs, one capability band (.0008 wide at seed 0), one
-training-gauge basin — at exact capacity parity, the dispatch machinery
-(+0.063% params) is free at the endpoint and carries .026–.038 of function
-by the toggle gauge.
+Three routing designs, one capability band (.0008 wide at seed 0, .0040
+across all six runs), one training-gauge basin — at exact capacity parity,
+the dispatch machinery (+0.063% params) is free at the endpoint and carries
+.026–.052 of function by the toggle gauge.
 
-### Per-task Spearman — the tie in detail, plus the seed-1 replication
+### Per-task Spearman — the tie in detail, plus the seed-1 replications
 
 | run | STS-B | SICK-R | STS12 | STS13 | STS14 | STS15 | STS16 | BIOSSES | mean |
 |---|---|---|---|---|---|---|---|---|---|
@@ -74,6 +74,8 @@ by the toggle gauge.
 | `a2_dense-s0` | .5731 | .6507 | .4965 | .5962 | .5430 | .7130 | .6798 | .5683 | .6026 |
 | `a3_random-s0` | .5707 | .6538 | .4966 | .5998 | .5448 | .7103 | .6806 | .5703 | .6033 |
 | `a1_anchored-s1` | .5681 | .6512 | .4931 | .5996 | .5423 | .7094 | .6745 | .5673 | .6007 |
+| `a2_dense-s1` | .5740 | .6537 | .4997 | .6038 | .5417 | .7128 | .6789 | .5671 | .6040 |
+| `a3_random-s1` | .5717 | .6533 | .4946 | .6007 | .5434 | .7128 | .6800 | .5809 | .6047 |
 
 ### What the dispatched experts carry — toggle decomposition (ON − OFF, per task)
 
@@ -82,12 +84,16 @@ by the toggle gauge.
 | `a1_anchored-s0` | −.008 | +.030 | −.020 | +.029 | −.010 | +.035 | +.053 | **+.121** | +.0288 |
 | `a1_anchored-s1` | +.005 | +.050 | −.004 | +.064 | +.001 | +.069 | **+.085** | +.034 | +.0382 |
 | `a3_random-s0` | −.006 | +.023 | +.002 | +.026 | −.004 | +.025 | +.050 | **+.093** | +.0261 |
+| `a3_random-s1` | +.026 | +.053 | +.040 | +.026 | +.000 | +.055 | +.083 | **+.135** | +.0524 |
 
 Positive = function the routed experts carry; negative = the quarter-width
-trunk alone does slightly better there. The routing reliably owns the hard
-semantics (SICK-R, STS13/15/16) and the domain task — *which* task it
-specializes toward most is seed-dependent (BIOSSES at s0, the STS suites at
-s1) while the total carried function stays stable.
+trunk alone does slightly better there. In three of the four routed runs
+the largest single carrier is BIOSSES (+.09 to +.14); the anchored seed-1
+run is the exception, shifting that weight onto the semantic suites
+(SICK-R, STS13/15/16). Both routed arms carry *more* total function at
+seed 1 (−.038 and −.052 vs −.029 and −.026) — where the routing
+concentrates varies by arm and seed; that it owns the hard semantics
+does not.
 
 ### Sample efficiency — small bed (500k rows, L=128, single consumer GPU)
 
@@ -110,7 +116,9 @@ gap. Endpoint equality and path advantage coexist.
 | `a1_anchored-s0` | .9980 | .8394 | 98.6 |
 | `a1_anchored-s1` | .9975 | .8391 | 98.6 |
 | `a2_dense-s0` | .9975 | .8418 | 99.2 |
+| `a2_dense-s1` | .9975 | .8420 | 99.2 |
 | `a3_random-s0` | .9980 | .8392 | 98.8 |
+| `a3_random-s1` | .9980 | .8396 | 98.6 |
 
 The captionbert-v2/-B trunks land in the same attractor (cos→target ~.84,
 erank ~99–103): separately-trained trunks across two poolings, two corpus
@@ -189,22 +197,23 @@ mean. Stage 1 therefore trains with the CLS readout.
 |---|---|
 | `runs/alephlm0-s0e-{mean,cls,attn}` | ✅ complete |
 | `s0a/` erank instrument | ✅ complete |
-| `runs/alephlm0-a2_dense-s0` | ✅ complete — mimicry R@1 .9975, cos→target .8418, erank 99.2/768; **8-task capability .6026** (`eval/`), inside the captionbert-v2/-B band: the dense control is triple-replicated |
+| `runs/alephlm0-a2_dense-s0` | ✅ complete — mimicry R@1 .9975, cos→target .8418, erank 99.2/768; **8-task capability .6026** (`eval/`), inside the captionbert-v2/-B band: the dense recipe is four-times replicated at capability (.6026–.6077) |
 | `runs/alephlm0-a3_random-s0` | ✅ complete — mimicry .9980, cos→target .8392, erank 98.8; **capability .6033** (band center: frozen-random routing matches dense at capacity parity); **dispatch-OFF .5772** (toggle −.026) — the routed experts carry .026 of task function, degrading gracefully (`eval/`) |
 | `runs/alephlm0-a1_anchored-s0` | ✅ complete — mimicry .9980, cos→target .8394, erank 98.6; **capability .6031**, dispatch-OFF .5743 (toggle −.0288). Anchors moved 1.06 rad from init; amplitude .101 |
 | `runs/alephlm0-a1_anchored-s1` | ✅ complete — **capability .6007** (seed-stable vs s0's .6031), dispatch-OFF .5625 (toggle −.0382); mimicry .9975, cos→target .8391, erank 98.6 — the s0 fingerprint reproduced to three decimals |
 | `runs/alephlm0-a2_dense-s1` | ✅ complete — **capability .6040** (seed pair .6026/.6040 brackets the band); mimicry .9975, cos→target .8420, erank 99.2 — the dense control is now seed-replicated in the basin |
-| `runs/alephlm0-a3_random-s1` | 🔄 IN PROGRESS — final epoch, already in the basin (last VAL: cos .8391 / erank 98.4) |
-
-**E1 at seed 0 (replication in flight): a three-way tie.** Learned-anchor,
-frozen-anchor, and dense trunks land within .0008 of each other on the
-8-task mean at exact capacity parity — the anchored form costs nothing,
-and learned addressing does not exceed a frozen random partition *as an
-encoder trunk*, even though it visibly reorganizes (anchors rotate a full
-radian, and its dispatched experts carry more function than the frozen
-arm's by the toggle gauge). The signed-address form's predicted advantage
-lives where the address parameterizes the output distribution — that is
-Stage 2's generative bed, which this result gates nothing about.
+| `runs/alephlm0-a3_random-s1` | ✅ complete — **capability .6047** (seed pair .6033/.6047), dispatch-OFF .5523 (toggle −.0524, the largest routed function measured); mimicry .9980, cos→target .8396, erank 98.6 |
+
+**E1 verdict, two seeds complete: the tie holds.** Learned-anchor,
+frozen-anchor, and dense trunks land within .0008 of each other at seed 0
+and within .0040 across all six runs — the anchored form costs nothing at
+exact capacity parity, and learned addressing lands even with a frozen
+random partition *as an encoder trunk*, even though it visibly reorganizes
+(anchors rotate a full radian, and every routed run carries real function
+by the toggle gauge, growing at seed 1 to −.052). The signed-address
+form's predicted advantage lives where the address parameterizes the
+output distribution — that is Stage 2's generative bed, which this result
+gates nothing about.
 
 Each run directory carries `checkpoints/` (state + rolling model snapshots +
 `final_model.pt` + `metrics.json`), `config/` (the exact resolved