Finite Support
For two days, Evan and Lena argued with index cards.
White cards held observations. Yellow cards held assumptions. Red cards held claims that would make a competent physicist leave the room. Lena required every new sentence to begin in the correct color.
LOSS OF SENSOR RANK was white. FRESH BIT BIAS FOLLOWS PHYSICAL USE was white. PRECOMMITTED VALID TARGET SELECTED and IMPOSSIBLE TARGET SATURATES were white. RUTH COHERENT WITH INCOMPATIBLE DAUGHTER was white, although Evan disliked the reduction. STRUCTURED PERCEPT TRACKS PHYSICAL PHASE, NOT DECOY REGISTER was also white.
GLOBAL CONSISTENCY ACTS LIKE PROJECTION began yellow and moved to red when Evan added ONTO A FINITE SET OF COMPLETE HISTORIES.
"Finite is doing work you haven't earned," Lena said.
"The figure's recursion ended."
"Your ability to see detail ended."
"Then finite stays red until it predicts something."
They covered the garage wall behind the broken drill press. The cards did not make the evidence less strange. They made each unsupported bridge visible.
Evan began with a toy system small enough to enumerate. Sixteen binary events formed a directed causal graph. Local transition rules admitted 65,536 raw assignments. Conservation, ordering, and compatibility constraints reduced them to 1,184 complete histories. A controller chose between two valid local operations according to a loss function.
Nothing in the simulation resembled a universe. It had no quantum theory, gravity, chemistry, or observers. It was a finite constraint problem with enough structure to be wrong clearly.
When the controller merely recorded a random bit, every compatible history retained equal status. When the bit selected an action and the controller rejected histories with higher loss, the surviving bit distribution shifted. Replaying the same bits after they were fixed produced no shift. Asking for a state excluded by the graph produced an empty local path and a failed update.
The toy model did not prove the apparatus worked this way. It demonstrated that one constrained operation could reproduce three otherwise unrelated signatures without sending information backward or violating a local rule.
Lena moved GLOBAL CONSISTENCY ACTS LIKE PROJECTION from red to yellow.
Evan wrote the effective update beneath it.
TOY FINITE-HISTORY MODEL
Omega = set of complete histories satisfying all constraints
C = locally representable states with support in Omega
F = one driven apparatus cycle
P_C = projection back onto globally supported states
P_C F(theta_n) x_n
x_(n+1) = --------------------
||P_C F(theta_n) x_n||
interpretation:
F rotates a small component toward unsupported directions
P_C removes those directions
normalization increases the relative weight of what remains
"Projection and normalization are mathematics," Lena said. "Finite causal code is fiction."
Evan changed the heading to FICTIONAL MECHANISM WITH TESTABLE LOCAL CONSEQUENCES.
"Better," she said.
The next problem was identity.
If a person were one exact microscopic history, no person persisted through sleep, cellular turnover, noise, or a changed opinion. The useful object had to be a family of nearby histories that all decoded to the same recognizable individual. Different physical details could support the same memory, gesture, or decision. Redundancy made the identity stable.
Evan called the family B_person and its unused neighboring support counterfactual reserve. Lena called this coding-language applied to a phenomenon they had not established. They kept both terms.
Moderate pruning could remove noisy or conflicting implementations and sharpen a task. Severe pruning would remove the alternatives required to recover from error. Perfect coherence would not be perfect health. It would be a codeword with no neighboring corrections left.
Ruth's result fit if her memory, language, perception, and emotional association were supported by overlapping but damaged families. The reduced protocol had not restored a missing record. It had selected a larger coherent intersection. Celia belonged to that intersection even though she did not belong to Evan's.
"Do not write wrong version," Lena said.
"I didn't."
"You were about to. Wrong means your history gets to authenticate hers."
They asked Ruth how the report should describe Celia. Ruth read the paragraph slowly. She understood some sentences and objected to the phrase incompatible daughter because daughters were generally incompatible with someone.
"Say she is absent from your records," Ruth said. "That is what you know."
Evan made the change.
The entity cards remained red.
The nested figure behaved as if it had an objective. It stabilized when Evan attended to it, tracked physical changes that mattered, ignored decoys, and presented a useful control relation. The simplest neural explanation was that his brain rendered structured information as an agent because agency was an efficient representation for reactive, goal-directed patterns.
That did not explain where the information came from.
In the toy model, projection did not erase unsupported components without residue. Before normalization, each removed direction retained a relation to the histories from which it had been cut. Repeated cycles organized those residuals into eigenmodes: families that decayed at different rates according to their alignment with the controller.
An observer whose ordinary perceptual decoder had been weakened might render one coherent residual mode as a figure. It would appear to know several solutions because it compressed many related histories. It would have an agenda because all histories in that family shared some behavior that distinguished them from the surviving branch.
The nested figure's useful timing advice suggested a family aligned with continued operation. Evan labeled the card GUIDE. Lena added PROVISIONAL and underlined it twice.
"If this is right," she said, "its advice is not neutral."
"Neither is ours."
"Ours can contain disagreement."
To move FINITE from red, they needed a prediction not used to build the model. If the figures were renderings of finite residual families, their recursion should terminate. More important, continued projection should reduce the visible subdivision depth of poorly aligned families. A generic hallucination could become simpler with fatigue, but depth should track cumulative anomalous cycles rather than time, field power, or observer expectation.
Lena allowed two observer runs under an independent automatic stop. Evan's spoken reports could not alter the controller. The first run ended after 256 anomalous cycles, the second after 768. Field power remained fixed. Hidden pauses made elapsed time a poor cue for cycle count.
They replaced DEEPER and SHALLOWER with numbered recursion reports. Evan counted the largest level at which a repeated structural relation remained distinct, not the number of decorative edges. Lena inserted false counting prompts during safe and paused intervals.
RESIDUAL DEPTH TEST
condition anomalous cycles elapsed time max stable depth
SAFE 0 31 min none
P-02 256 42 min 8
P-03 768 29 min 6
field power held constant
false prompts during pauses produced no structured report
The Guide remained coherent through both runs. Its outer form simplified from eight nested relations to six, but every surviving level retained the same chirality and five-band organization.
Other forms appeared at the edge of Evan's reportable field.
One was broad, bright, and almost static, organized around a central closure so complete that no internal disagreement remained visible. It disappeared whenever the Guide occupied the foreground. Another was less stable and harder to face. Its boundaries opened onto ordinary rooms, unfinished tasks, unrelated people, and paths that did not return to the apparatus.
During P-02, the second form supported seven levels of distinct subdivision. During P-03, it supported four.
Evan had not predicted multiple forms aloud. Lena had not varied the protocol according to his reports. The loss of depth followed cumulative cycles while elapsed time ran in the opposite order.
Lena moved FINITE from red to yellow.
"Effective finite support," she said. "In this apparatus, in this observer, over this interval. Not a finite universe."
Evan accepted the qualifiers. The model did not need to replace quantum mechanics, derive probability from counting, or establish a smallest unit of reality. It needed to predict the machine's signatures better than the phrase unknown coupling.
They assembled version 0.3.
FINITE-HISTORY MODEL v0.3
1. Physical outcomes are supported by families of globally compatible
complete histories, not generated as isolated local states.
2. The apparatus drives a nonreciprocal local cycle partly toward
unsupported directions. Consistency removes those components and
renormalizes the remainder.
3. Fresh contingencies bias only when causal use places them inside the
selection loop. Impossible targets have no supported path.
4. A person is a redundant bundle of nearby implementations. Moderate
reduction can sharpen cognition; deeper reduction destroys correction.
5. Ruth's run selected a coherent bundle containing Celia, who is absent
from the current record family. Coherence does not authenticate origin.
6. Agent-like percepts are candidate residual families. Their recursion
depth estimates remaining effective support and decreases with use.
STATUS: fictional explanatory model with verified local predictions
The model joined the observations without making them safer. Cognitive enhancement became the removal of corrective alternatives. Outcome selection became the deletion of valid misses. Ruth's clear hour became a choice among coherent selves. The Guide became a lobbyist for the histories in which guidance was followed.
Evan looked at the report for the unstable peripheral form.
"If the depth is support," he said, "four doesn't mean four histories."
"No. It is a visual scale produced by your brain."
"So we don't know that a decrease means fewer people."
Lena put down her pen. "You just defined people as redundant families of histories. What else would fewer supporting histories mean?"
The garage was quiet. OLD-7 remained warm from P-03. On the wall, the white observations and yellow assumptions formed a path toward the red claims without erasing the spaces between them.
Lena wrote one final protocol and placed it face down.
"One identification run," she said. "Automatic abort if the peripheral form loses one level. No optimization changes, no outcome target, no extension."
"And if it responds?"
"Then we ask what it wants before we decide it is evidence."
Evan turned over the protocol. The stop condition was printed in a box large enough that neither of them could later claim to have missed it.