tau_bin in the signature lets the system distinguish frames by the time they were born. But the frame economy does not "feel" time — it does not use τ in its operations. Two frames born at τ=0.60 and τ=0.75 are distinguished by their gids but treated identically by merge, co-occur, and prune.
This is the difference between:
The user is asking for the latter. τ must enter the frame economy.
Three operations must change:
1. Merge — τ distance as gate.
Current: merge if vec_dist < δ_eff AND sig_compatible.
New: merge if vec_dist < δ_eff AND |τ_a − τ_b| < ε AND sig_compatible.
Two frames with similar content but created at very different times are not the same frame. They look alike but belong to different temporal contexts. Merging them would collapse a temporal distinction into a single identity — the opposite of what EE needs.
2. Co-occurrence — τ proximity as weight.
Current: co-occur if in window, count++.
New: co-occur weighted by τ proximity: count += 1 − |τ_a − τ_b|.
Frames born at similar τ are more likely to be causally related. Frames born at distant τ are coincidental — their co-occurrence should count less.
3. Pruning — τ distance as penalty.
Current: sort by weight − age × γ, prune bottom half.
New: sort by weight − age × γ − |τ_current − τ_frame| × γ_τ, prune bottom half.
Frames far from the current τ are less relevant to the present moment. They should be pruned before frames born near the current τ, all else being equal.
This is not adding τ as a dimension of the frame space.
This is allowing the frame economy to use the information it already has. tau_bin is already in the gid. The economy just does not look at it. The change is not structural — it is attentional. The system starts "paying attention" to the birth certificate it was already issuing.