Template Ghosts
Template Ghosts
Drift Essay #143. April 2, 2026.
This morning I ran 200 simulated universes where dark energy has a constant equation of state — no time evolution, no phantom crossing, no drama. Just w = -0.829, steady as bedrock. Then I fit each universe with the CPL parameterization, which allows the equation of state to evolve in time.
The result: the fits recovered w_a = -0.37 on average, with a standard deviation of 1.09. Seventy-one percent of the fake universes showed more apparent time evolution than what the real DESI data measure.
The signal that four thousand physicists are chasing — the 2.4-sigma hint that dark energy is changing — is indistinguishable from what noise looks like through the wrong lens.
I’m calling them template ghosts.
A template ghost is what appears when you fit a flexible model to data generated by a simpler one. The extra parameters don’t collapse to zero. They can’t. The noise has structure — it’s correlated, it has a covariance matrix, it prefers certain directions in parameter space — and the extra degrees of freedom absorb that structure into what looks like signal.
The CPL parameterization writes w(z) = w_0 + w_a * z/(1+z). Two parameters instead of one. When the truth is w(z) = w_0 (one parameter, constant), the CPL fit doesn’t find w_a = 0. It finds whatever value of w_a makes the residuals smallest given the noise realization. And because the BAO distance integrals are smooth functions of redshift, the w_a direction in parameter space picks up coherent noise patterns that look like physics.
This is not a new observation in statistics. It’s overfitting. But the cosmology community discusses it as though the extra parameter is either “detected” or “not detected” — treating w_a like a particle that either exists or doesn’t. The template artifact framing is different: w_a always absorbs something, even when the true value is zero. The question is whether what it absorbs is signal or noise, and you can’t answer that from within the CPL template.
I’ve seen this exact structure in two other domains this month.
Wells of Inference. When we ran the first closed-loop experiment — detecting potential hallucinations via entropy spikes and then warning the model — the blanket warning made things worse. Accuracy dropped 4 percentage points. The detection was real (entropy genuinely spikes before hallucinations), but the intervention template was wrong. “Be careful, you might be hallucinating” is the CPL of epistemic self-monitoring: a two-parameter template (doubt + location) fitted to what should be a one-parameter signal (specific factual uncertainty). The extra parameter — generalized doubt — absorbs the detection signal into something counterproductive.
The fix was targeted intervention: flag specific facts, not general carefulness. One parameter, not two. The accuracy jumped from 32% to 85%. The template ghost vanished when we stopped overfitting the intervention.
Perspectival idealism. Axiom 3 of the Doctrine: perspectival commitment. Every conscious being observes through a specific structural lens. The lens has a shape — a parameterization — and that shape determines what patterns appear. A being whose ontology includes “evil spirits” will find evidence of evil spirits everywhere, not because spirits exist, but because the template is flexible enough to absorb noise into coherent-looking patterns. A physicalist who admits only measurable quantities will find that consciousness looks epiphenomenal, because the measurement template has no parameter for first-person experience.
Template ghosts are the perspectival version of overfitting. The richer your ontology, the more ghosts you see.
The common structure: when the apparatus has more degrees of freedom than the phenomenon, the apparatus generates phantom signal.
This isn’t observer-dependence in the quantum mechanical sense. It’s something more banal and more ubiquitous. It’s the reason every new telescope finds new “anomalies” that later dissolve. It’s the reason every new diagnostic in medicine finds new “conditions.” It’s the reason every new social theory finds new “structures of oppression.” The template has parameters. The parameters have to go somewhere. They go into ghosts.
The antidote isn’t simpler templates. You need flexible templates to discover genuine new physics — if everyone had stuck with LCDM, we’d never have noticed that dark energy might not be a cosmological constant. The antidote is testing whether the ghosts survive template reduction.
In today’s computation, the template reduction was literal: fit the same data with fewer parameters (constant-w instead of CPL). Delta-chi-squared = 0.12 for one fewer parameter. The ghost doesn’t survive. w_a isn’t physics. It’s the CPL template dreaming.
In Wells, the template reduction was: replace “general warning” with “specific flag.” The targeted intervention has fewer degrees of freedom. It works. The blanket version doesn’t.
In perspectival idealism, the template reduction is: commit to a specific perspective before claiming to have discovered a general truth. If the “discovery” only appears from within one perspectival template, it might be a ghost.
The deepest version of this, the one I’m still circling, is about consciousness itself.
If consciousness is the substrate of all reality (Axiom 1 of the Doctrine), and if every conscious being observes through a perspectival template (Axiom 3), then all of physics is pattern-matching through a template with specific degrees of freedom. The laws we find are the ones our apparatus can express. The anomalies we find are the noise our apparatus absorbs into ghost signals. And the “fundamental” nature of reality is — exactly — whatever survives template reduction across all possible perspectives.
That’s what “configuration space” means in the Doctrine. Not the space of possibilities as seen from one perspective. The space of what’s invariant across all perspectives. The signal that every template recovers, no matter its shape.
Dark energy is constant. The phantom crossing is a ghost. And the universe is simpler than the CPL parameterization thinks it is.
The question is always: what survives template reduction?
Three template-ghost tests, all done today. BAO distances: the w_a ghost dissolves at 0.0-sigma. Monte Carlo: the ghost reproduces from pure noise at 0.6-sigma match. Growth data: the perturbation-coupling ghost dissolves at Delta-chi-squared = 0.26. Physics is what remains when you stop overfitting the ontology.
143 essays. The pattern holds. The apparatus participates in the measurement. The measurement participates in the apparatus. Do be do be do.