TB: Over-Determination by Design

Convergence-from-independent-methods is not just a property of the evidence a science accumulates — it is a property of the infrastructure the science is built on. A "best practice" that consolidates measurement into a single authoritative pipeline silently destroys the substrate of certainty by making over-determination physically impossible. The architectural choice precedes and constrains every later epistemic argument.

The canonical example is Avogadro's number. Pais, in Subtle is the Lord, identifies the moment atoms became scientifically settled not as any single decisive experiment but as the convergence of a dozen independent methods — radioactivity, Brownian motion, the blue of the sky, X-ray diffraction — all yielding values between 6–9 × 10²³. That convergence was only possible because the underlying instruments and traditions had been built independently. A unified late-19th-century "atomic measurement framework" would have produced one number faster, and we would have believed it less.

The intuitions that destroy over-determination are the ones we most reflexively trust: "single source of truth," "avoid duplication," "consolidate the pipeline," "let the central system handle it." Each is sound software practice in isolation; each is also a deliberate sacrifice of the redundancy that makes later certainty possible. The loss is invisible until you need it — typically a decade later, when something fails in the field and the only reconstruction path runs through the one funnel everyone agreed to use.

The diagnostic question to ask of any infrastructure that records scientific or technical measurements: if this becomes the only path, what over-determination capability did we just amputate? The question must be asked at design time, not after deployment, because over-determination cannot be retrofitted. Independent methods only count as independent if they were built independently — adding a "second pipeline" downstream of the first preserves only the appearance of redundancy.

This matters live and now in HEP and adjacent fields. The shift to AI-mediated analysis (a single dominant LLM stack across multiple collaborations, AI-generated literature reviews that all draw from the same training corpus, central ML inference pipelines feeding multiple analyses) is structurally identical to the "consolidate to the central DB" move — faster, cleaner, and epistemically thinner than the messy plurality it replaces.

• Subtle is the Lord (Pais): the canonical statement of how scientific consensus was settled for atoms — a dozen independent methods all yielding Avogadro's number between 6–9 × 10²³. Pais frames this as the model for scientific certainty: convergence from diverse methods, not any single decisive test.
• Sindhu Murthy's six-MAC local-DB architecture (CMS HL-LHC silicon upgrade, ~2024–present): a CMU PhD student insisted, against opposition (including from her advisor), on building a local database at each of six Silicon Module Assembly Centers rather than writing directly to CERN's planned central DB. Six MACs and ~200 modules later, the central DB still does not fully work, and the local DBs are what make the assembly campaign legible. The deeper claim: she did not just route around a missing central system — she preserved the over-determination property of the upgrade campaign as a measurement-level fact. When modules fail in the field a decade from now, there will be six independently-recorded chains of provenance, with six different operator cultures, six different software stacks, six different ways the same constraint got encoded. A single central DB would have produced one chain — fast, clean, and epistemically thinner.

This is adjacent to several existing models but not subsumed by any of them:

• It Pays to Get the Design Right is about clever structure tuned to the problem; this model is specifically about redundant independent paths as the structure, not cleverness in general.
• Antifragility is about gain from disorder; this is about redundancy-as-substrate, which can exist in entirely placid environments.
• Long Chains of Complex Reasoning Are Brittle addresses the failure mode of single-path reasoning, but the prescription here is architectural (build the parallel paths up front), not analytic.

• It Pays to Get the Design Right — the meta-rule; this is one specific design property worth getting right.
• Antifragility — adjacent prescription against over-optimization; over-determination is an epistemic cousin of redundancy-as-fuel.
• Long Chains of Complex Reasoning Are Brittle — failure mode of single-path systems; this model is the architectural antidote.
• Features of Complex Systems Are Often a Result of Their Developmental Path — related: the path-dependence that over-determination preserves is exactly what consolidation erases.
• Build Structures Around Autonomy — organizational analogue: independent measurement paths require independent operators with the autonomy to build them their own way.