export m2/soft-pool

exports/m2/soft-pool/index.html

@@ -1202,7 +1202,7 @@ training, with the spill of E1 beside them; seed mean and range over every run.
 <p>Both soft twins spilled more than their fixed arms, and the pool over slices 1 to 3, which has no fixed twin, spilled most.</p>
 <p>The ramp over slices 1 to 3 lost the anchor on one seed of eight at the fixed temperature, where the anchor-ramp ramps from slice 1 lost two or three. That fits the second anchor-ramp reading, that the earlier slices need about as much pull as the step gives them: this ramp gives slices 1 to 3 that much, and drops slice 4 instead.</p>
 <p>Its one loss came later and in another way, with the alignment partly back by the end but no longer doing the work. That is a reason to keep scoring removal, and not alignment alone, when judging a run.</p>
-<p>The soft ramp kept the anchor on every seed here without the step between slices 1 and 2, so it would be a candidate if its spill can be kept down. If the spill comes from the soft pool reaching every position, as E3 suggests, two changes would aim at it: a soft pool over the answer positions only, so the pull is shared among the example answers and the query answer and nowhere else, or a stronger anti term while the pool is soft. Turning the anti term off on slice 4 might also let the ramp keep the concept there. As it stands, <code>falloff</code> still keeps the anchor on every seed with the edit inside the criterion, which no new arm matched on both counts.</p>
+<p>The soft ramp kept the anchor on every seed here without the step between slices 1 and 2, so it would be a candidate if its spill can be kept down. If the spill comes from the soft pool reaching every position, as E3 suggests, a stronger anti term while the pool is soft would aim at it. Turning the anti term off on slice 4 might also let the ramp keep the concept there. As it stands, <code>falloff</code> still keeps the anchor on every seed with the edit inside the criterion, which no new arm matched on both counts.</p>
 <h2 id="method">Method<a class="anchor-link" href="#method" title="Link to this heading">&para;</a></h2>
 <p>The pooling temperatures are step inputs here, scheduled like the weights (<code>AnchorSpec.tau_start</code>, <code>slice_tau_start</code> and <code>tau_anneal_end</code>); before this experiment they were fixed when the training step was built. The schedule is minimum-jerk in log τ from the start value at epoch 0 to the end value at epoch 40, and holds after. Training otherwise follows the split-anti-seeds experiment unchanged. The trajectory of each run also keeps α by slice and position for each op separately, and the e₁ component of every syntax embedding, for later use. The 32 runs cost about $11.</p>
 <div class="footnote all-sidenoted">

exports/m2/soft-pool/index.md

@@ -188,7 +188,7 @@ The ramp over slices 1 to 3 lost the anchor on one seed of eight at the fixed te
 
 Its one loss came later and in another way, with the alignment partly back by the end but no longer doing the work. That is a reason to keep scoring removal, and not alignment alone, when judging a run.
 
-The soft ramp kept the anchor on every seed here without the step between slices 1 and 2, so it would be a candidate if its spill can be kept down. If the spill comes from the soft pool reaching every position, as E3 suggests, two changes would aim at it: a soft pool over the answer positions only, so the pull is shared among the example answers and the query answer and nowhere else, or a stronger anti term while the pool is soft. Turning the anti term off on slice 4 might also let the ramp keep the concept there. As it stands, `falloff` still keeps the anchor on every seed with the edit inside the criterion, which no new arm matched on both counts.
+The soft ramp kept the anchor on every seed here without the step between slices 1 and 2, so it would be a candidate if its spill can be kept down. If the spill comes from the soft pool reaching every position, as E3 suggests, a stronger anti term while the pool is soft would aim at it. Turning the anti term off on slice 4 might also let the ramp keep the concept there. As it stands, `falloff` still keeps the anchor on every seed with the edit inside the criterion, which no new arm matched on both counts.
 
 ## Method