Explain why continuous-frequency V2 failed without discarding its signal

[jamiepratt]

Goal

Explain why continuous-pair-frequency-logistic-v2 failed its promotion gate without discarding the useful frequency signal it captured. Rebuild the executable article as a concise, standalone causal explanation backed by new, explicitly post-hoc diagnostics.

Decisions

• Keep V2, its historical gate, and immutable evidence unchanged; V1 remains the current target.
• Target 2,500–3,200 always-visible words in six causal sections.
• Freeze a post-hoc diagnostic protocol before generating new results.
• Diagnose correctly specified logistic scenarios first, then retain realistic stress evidence as technical depth.
• Add a paired failure microscope, supported-cell pattern views, and the coverage-versus-length trade-off.
• Retain derivations, simulation protocol, provenance, stress evidence, and code behind progressive disclosure.
• Add active desktop and mobile contents navigation.
• Keep model, software, and publication validation distinct.

Scope

The publication submodule owns executable article source, model code, fixtures, diagnostic evidence, tests, interactive assets, and rendered publication inputs. This workflow repository owns orchestration, the current contract, and this issue hierarchy.

Non-goals

• Retune or retrospectively promote V2.
• Rewrite historical evidence.
• Claim that simulations validate real learners or calibrated Lexibench items.
• Change V1's current-target status.
• Select a V3 design within this article.
• Resolve the separate global Quarto resource-ownership decision in Decide ownership of the global Quarto *.cljc resource rule #12.

Acceptance Criteria

• Historical V2 evidence remains byte-for-byte unchanged and replay/provenance behavior is explicit.
• A frozen post-hoc protocol isolates oracle, inference, prior, numerical approximation, sample-size, and stopping effects.
• Primary mechanisms are checked across all 45 supported cells.
• The article presents a 2,500–3,200-word causal narrative with technical depth preserved.
• The failure microscope and supporting views distinguish single examples from repeated-run evidence.
• Responsive contents navigation reports the current section accessibly without obstructing content.
• Configured model, software, and publication checks pass and are reported separately.
• Publication and workflow pull requests preserve repository ownership and commit order.

Orchestration

Issue kind: parent PRD

Children:

• Make V2 simulation replay and cell provenance explicit #14
• Diagnose matched-model V2 failure with paired component contrasts #17
• Confirm V2 failure mechanisms across all supported cells #18
• Restructure the V2 article around its causal explanation #19
• Add responsive active contents navigation to the V2 article #20
• Review, validate, and publish the revised V2 explanation #21

Related: #13 tracks dependency-aware preview freshness for interactive assets. It is relevant to verification but is not a hard implementation dependency when the focused preview is restarted.