The PDZ/DCA long-range epistasis gap and the IB compression debate are both estimand-vs-estimator stories — but they resolve in opposite directions

Both vault notes describe a gap between an inferred statistical quantity and a real target. Checking whether that's a real bridge or a coincidence of phrasing required asking, for each case, where the gap actually lives: in the estimand (the true target quantity) or in the estimator (the procedure used to measure it). Biostatistics keeps this distinction as a standing pair of terms — "the estimand is the true effect... an estimator is a formula or algorithm used to estimate the target quantity" (JAMAevidence, Estimands, Estimators, and Estimates) — Tier 3-4, definitional use only.

Applied to the DCA case — claim-pdz-dca-couplings-track-short-range-epistasis-more-than-long-range's ρ=0.69 (short) vs. ρ=0.48 (long) gap in the PDZ domain: Rodriguez Horta & Weigt (2021, PLOS Comp Biol) tested whether phylogenetic bias — a known regularization/finite-sampling confound in DCA — explains weak long-range contact prediction, and found "the largest coevolutionary couplings... are only weakly influenced by phylogeny"; spurious phylogenetic correlations mainly contaminate the intermediate-strength couplings, not the strong ones. Combined with claim-dca-underestimates-long-range-epistasis-in-allosteric-materials's ground-truth synthetic-network result (the same short/long asymmetry reproduces in a fully controlled spring-network model with no real-world sampling problem at all), this points to the estimand (real epistasis) being well-behaved and the estimator (pairwise DCA couplings) being genuinely under-expressive for long-range, collective allosteric coupling — a real, ground-truth-confirmed representational limit, not primarily a sampling artifact.

Applied to the IB case, per the vault's own claim-ib-compression-may-be-a-binning-artifact-not-real-mutual-information: Goldfeld et al. prove the estimand itself — true mutual information in deterministic monotone-nonlinearity networks — is mathematically frozen (infinite or constant). There, the observed movement can only be the estimator (binning) tracking something else (geometric clustering). [unverified-mechanism — needs primary]: whether this artifact verdict is universal or nonlinearity-dependent is contested elsewhere in the ML-theory literature and is already flagged in the vault's claim-ib-compression-phase-is-nonlinearity-dependent-not-universal.

Why this was hop-worthy

The two notes share surface vocabulary ("a measured change may be artifact, not real") but the actual diagnosis sits at opposite ends of the estimand/estimator distinction — DCA's gap is estimator-side and real; IB's is estimand-side and (in the deterministic case) provably artifactual — which is a sharper and more useful finding than either "yes, same thing" or "no relation."

Further leads

• Rodriguez Horta & Weigt's randomization method for isolating phylogenetic signal could be a template for testing whether other DCA-adjacent claims in the vault are estimator-real vs. sampling-artifact.
• Whether the "estimand vs estimator" framing is worth its own concept note — it recurs informally in the vault's claim-llm-neural-metacognition-incomplete (architectural "nothing to access" vs. geometric "low-dimensional access") and observation-weighted-voting-power-gap-recurs-across-cs-law-regulation (terminological echo vs. real convergence) clusters.

Entity candidates

• Martin Weigt — person — DCA/protein-coevolution physicist (Sorbonne), no entity page found; recurs across the vault's DCA cluster.
• estimand — term — standard biostatistics/clinical-trials vocabulary (estimand vs. estimator vs. estimate), not previously checked against the vault's word list this session (vault_word tool call failed on permissions); flagging for a future first-encounter check.

Hop chain

Hop 1 — vault seed pair (unlinked, cosine 0.75)

• Hook type: Surprising claim (does a superficial resemblance hold up?)
• Hook: Both notes say "the measured gap between an inferred quantity and a real one may not be what it looks like."
• Why followed: This is the session's assigned seed question.
• Key findings: Framing the literal claim ("DCA's long-range failure is itself a binning-style artifact") scored max_cosine 0.803 against the vault's own sibling DCA note — duplicate-risk — so that literal version of the bridge was discarded in favor of investigating the mechanism directly.

Hop 2 — WebSearch: DCA regularization / finite-sampling / phylogenetic-bias literature

• Hook type: Mechanism question (zoom in)
• Hook: DCA is known to need pseudocount regularization and phylogenetic reweighting; both are candidate sources of a spurious short/long-range split.
• Why followed: If DCA's long-range weakness were mainly a sampling/regularization artifact, the bridge to IB's binning artifact would be direct and strong.
• Key findings: Regularization and phylogenetic correction are real, documented confounds in DCA generally, but the search surfaced no direct claim that they specifically explain the strong long-range signal loss.

Hop 3 — Rodriguez Horta & Weigt (2021), PLOS Comp Biol, https://pubmed.ncbi.nlm.nih.gov/34029316/

• Hook type: Mechanism question (zoom in further, primary source)
• Hook: Do phylogenetic corrections extend to the strongest (structurally real) couplings, or only weak/spurious ones?
• Why followed: This is the specific test of the artifact hypothesis for the DCA side of the bridge.
• Key findings: "The largest coevolutionary couplings... are only weakly influenced by phylogeny" — phylogenetic/sampling bias mainly contaminates intermediate-strength, false-positive-adjacent couplings, not the strong long-range signal the PDZ note is about.

Hop 4 — WebSearch: IB compression phase, artifact vs. real-phenomenon debate

• Hook type: Surprising claim (zoom out to the wider ML-theory field)
• Hook: Is Goldfeld et al.'s "provably an artifact" verdict the field's settled view?
• Why followed: To check whether the IB side of the bridge is as settled as the vault note's framing suggests.
• Key findings: The debate is live and activation-function-dependent — some work reports MI computed exactly in quantized networks still shows compression, i.e. not universally a binning artifact — consistent with the vault's own claim-ib-compression-phase-is-nonlinearity-dependent-not-universal.

Hop 5 — "Estimand vs. estimator" (JAMAevidence / biostatistics reference)

• Hook type: Mechanism question / definitional (zoom out further, cross-domain)
• Hook: Statistics already has a standard vocabulary for "is the problem in the true target quantity or in the procedure measuring it?"
• Why followed: This is the abstraction the DCA and IB cases actually share — not a shared mechanism, but a shared diagnostic question, answered oppositely in each field.
• Key findings: Estimand = true target quantity; estimator = the procedure. DCA case: estimand (real epistasis) is well-behaved, estimator (pairwise couplings) is under-expressive — real but incomplete. IB case: estimand (true MI) is itself degenerate in the deterministic regime — any estimator movement there is artifact by construction.

Hop 6 — vault_bridge tool check on the full "estimand vs estimator, DCA vs IB" framing

• Hook type: (computable assist, not one of the six — used to verify the bridge per protocol amendment)
• Hook: bridge_candidate flag on the synthesized framing.
• Why followed: Required check before capture per the hop-protocol's bridge-check amendment.
• Key findings: bridge_candidate = true; the IB-compression note sits unlinked to both DCA notes despite being in the same 0.70-0.75 frontier band as their own (already-linked) internal pair — confirming this capture would create a genuinely new link, not a redundant one.

Surprise: expected DCA's long-range weakness and IB's compression fluctuation to share a mechanism (both estimator artifacts of the same general kind) — found they sit at opposite ends of the estimand/estimator distinction: DCA's gap is a real, ground-truth-confirmed limit of the estimator against a well-behaved estimand, while IB's compression (in the deterministic case Goldfeld et al. studied) is artifactual because the estimand itself is mathematically frozen.

Surprise: expected phylogenetic/regularization bias to be a strong candidate explanation for DCA's long-range weakness (the natural "estimator artifact" reading) — found Rodriguez Horta & Weigt (2021) report the opposite for exactly the strongest couplings, pointing away from an artifact explanation for the residue-1/residue-8-class long-range signal the PDZ note describes.

Saved hooks not followed:

• Rodriguez Horta & Weigt's randomization/reshuffling method for isolating phylogenetic signal from real coupling signal — could be applied as a general artifact-detection template elsewhere in the vault's DCA cluster.
• The quantized-neural-network exact-MI line of IB literature (contra Goldfeld) — not read directly, only found via search summary; would need a primary-source read before citing further.

post-worthy: maybe — the estimand/estimator framing is a genuine, non-obvious methodological bridge that resolves the seed's ambiguity honestly (real bridge in structure, opposite in diagnosis) rather than forcing a false convergence, but it rests on one new Tier-1 source plus synthesis of two already-captured vault claims, so it's thinner than a fully independent two-source bridge.