Having actualized the quanta and the ⊕∘▷⊗ operator algebra, the graph's future state is its own forward operator made real: a receding-horizon planner that, per future frame f = (domain, quality target, budget), composes a model M_f minimizing active-quanta cost subject to quality ≥ target — caching the fixed point on stable frames and replanning on drift. The planner forecasts the frame's domain from a near-horizon belief window (it cannot see the future); the quality target is a known per-frame input. This is optimization in expectation with regret, never a per-frame omniscient optimum.
83%
Quality hit rate
50%
Fixed-point cache rate
0
Cost regret vs oracle
2
Quality misses
3
Genuine surprises
Measured quality table — perplexity by composition × domain
Lower is better; bold = best on that domain. Cost = active LoRA quanta (sparsity footprint). Note ⊗omni is one cost-1 adapter carrying all three; ⊕ mixtures are cost-2.
Composition
Cost
code ppl
biomed ppl
general ppl
base(dense-limit)
0
169.1
89.7
69.7
▷code
1
91.9
88.0
67.4
▷biomed
1
220.0
63.8
74.2
▷shared
1
191.3
127.0
48.1
⊗omni
1
120.8
73.2
55.8
⊕code+shared
2
109.8
99.5
53.1
⊕biomed+shared
2
180.6
69.3
54.1
Online planner trajectory over the modeled future frames
t
regime
true
belief (forecast)
target
plan M_f
cost
real ppl
quality
mode
oracle
1
morning code sprint
code
code
95
▷code
1
91.9
✓
replan
▷code
2
morning code sprint
code
code
95
▷code
1
91.9
✓
cache
▷code
3
morning code sprint
code
code
95
▷code
1
91.9
✓
cache
▷code
4
code, relaxed budget
code
code
125
▷code
1
91.9
✓
cache
▷code
5
research reading (drift)
biomed
code
70
▷code
1
88.0
✗ miss
cache
⚠ surprise
▷biomed
6
research reading
biomed
biomed, code
70
▷biomed
1
63.8
✓
replan
▷biomed
7
research reading
biomed
biomed, code
70
▷biomed
1
63.8
✓
cache
▷biomed
8
standup notes (loose bar)
general
biomed
75
▷biomed
1
74.2
✓
drift
⚠ surprise
base(dense-limit)
9
doc polish (tight bar)
general
biomed, general
52
▷shared
1
48.1
✓
replan
▷shared
10
doc polish (tight bar)
general
biomed, general
52
▷shared
1
48.1
✓
cache
▷shared
11
evening bugfix (drift)
code
general
95
base(dense-limit)
0
169.1
✗ miss
drift
⚠ surprise
▷code
12
evening bugfix
code
code, general
95
▷code
1
91.9
✓
replan
▷code
Cost trajectory — planner vs hindsight oracle
━ planner (online, forecast-driven) ╌ oracle (clairvoyant, per-frame optimal) ● miss = quality target missed at an unforeseen surprise
What the realization actually found (measured, not asserted)
Routing dominates when the domain is known. Each ▷ specialist is best on its own domain; the planner's job is to pick and cache the right cost-1 specialist per frame. It matched the clairvoyant oracle's total cost exactly (regret 0).
The dense base is a cost-0 sweet spot on easy frames. Base distilgpt2 is competitive on general text at zero adapter cost — the oracle uses it on the loose standup frame. Aggressive cost-minimization is why the planner also reached for base at t=11 and got burned when code arrived.
Fixed-point caching is real. 6 of 12 frames reused the prior plan across stable runs — no recomputation until the belief drifted.
Merged/mixture compositions are dominated here. With these tiny adapters the code specialist already generalizes (biomed 88 vs its own 91.9), so ⊗omni is a middling generalist and ⊕ mixtures rarely earn their cost-2. This is an empirical property of the current quanta — stronger, more specialized adapters would shift the frontier and make the robust omni hedge pay off. That is the concrete next vector.
Tarski residual (measured, honest). Quality misses occur exactly at unforeseen domain shifts (drift the persistence forecaster could not predict). Over the modeled frame distribution the expected regret is bounded and measured here, but no online policy achieves zero regret against genuinely unknowable future frames — no free lunch. This is the graph's UNDECIDABLE TarskiResidual, measured rather than hand-waved. The two misses (t=5, t=11) are exactly the frames where an unheralded domain shift outran the belief window — no online policy avoids them against genuinely unknowable futures.
Reproduce: python realize_future.py regenerates this trajectory from the trained quanta. Graph 1c0cc1b9-d95a-4448-b7d7-a08e1e347820 · snapshot 2026-07-06