# Phase 2: Inter-Brain Coupling and the Dyadic-Coherence Claim **Does a pre-session linguistic proxy predict real neural coupling during a real conversation?** Diego Rincón · Phronesis · Draft protocol, v1 --- ## 1. Background & Rationale *From Grammar to Coherence* (Rincón, 2026) makes a claim stronger than "grammar tracks individual psychological state": that consciousness itself is a property of the *coupled* eigenspectrum of two nervous systems, not a single brain in isolation. The Clarity Validation Pilot (this repository, `clarity-validation-protocol.md`) tests the narrower, cheaper claim — does the `displacementIndex` computed from one person's writing predict their own PHQ-9/PCL-5 scores. That is a single-person test. It cannot, on its own, support the dyadic claim. This protocol is the test that actually can: does a person's Clarity `displacementIndex`, measured *before* a real conversation, predict the degree of inter-brain coupling measured *during* that conversation with another specific person? This is expensive, access-constrained, and should only be attempted if Phase 1 clears its pre-specified threshold (r ≥ .30). It is written now, in full, so that gate is a real decision point rather than something improvised later. **Two things this protocol is not.** It is not a claim that this design is fully specified and ready to run tomorrow — recruitment access and equipment are open, named problems below, not glossed over. And it is not the same test as the ds007471 analysis already run and published (`ds007471-hyperscanning-note.md`) — that was a naive first pass on someone else's open dataset, on a musical joint-action task, using one coupling metric that turned out null. This protocol is designed to not repeat that specific mistake, and says exactly how. ## 2. What the ds007471 null result already taught this design A simple alpha-band Phase Lag Index, averaged across all 32×32 sensor-pair combinations, did not predict joint agency or synchronization performance in 29 real dyads doing a musical coordination task. That result is real and should inform this protocol, not be ignored by it: - **A single coupling metric is not enough.** PLI is one specific, defensible choice, but it is not the only one, and a null result on one metric does not mean the underlying phenomenon is absent. This protocol pre-registers multiple candidate metrics (§5) rather than betting everything on one, the way the earlier prototype did by necessity. - **The theoretically motivated metric was never actually tested.** ds007471 tested a generic literature metric (PLI). It did not test the metric this research program's own theory predicts should matter most: the eigenspectrum of the *joint* graph Laplacian across both people's channels combined, treating the two brains as one coupled system rather than computing pairwise channel statistics and averaging. That is the metric this protocol tests first and treats as primary, not an afterthought. - **A musical task may be the wrong task.** Joint agency during structured musical coordination is one specific kind of dyadic coupling. Therapeutic rapport, or any conversation where one person is trying to be understood by another, is a different kind of coupling — arguably closer to what the theory is actually about. Task choice matters and is addressed directly in §4. ## 3. Objectives / Hypotheses **H1 (primary).** A participant's pre-session Clarity `displacementIndex` (from a free-write completed alone, before the conversation) correlates negatively with the joint-graph-Laplacian coupling metric (§5) computed from the two participants' EEG during the conversation that follows. **H2 (primary).** Joint-graph-Laplacian coupling correlates positively with post-conversation self-reported connection (a short scale adapted from the UCLA-3 used in Phase 1, plus a direct "how understood did you feel" item). **H3 (secondary, direct replication check).** Does alpha-band PLI, the same metric tested in ds007471, show a relationship here that it didn't show in the musical-task dataset? If it does, that's informative about task-dependence. If it doesn't here either, that's a second independent null on the same metric, which is stronger evidence against it specifically (not against the broader claim). **H4 (exploratory).** Does coupling strength during the conversation predict a same-day post-conversation Clarity `displacementIndex`, i.e. does the coupling event itself measurably shift linguistic coherence afterward, not just before? **Pre-specified thresholds**, matching Phase 1's convention: r ≥ .30 on H1 or H2 is signal worth expanding; r < .20 is treated as null for this design and reported as such, not quietly reframed. ## 4. Design and the phased access problem Recruiting real therapist-patient dyads for simultaneous dual-EEG during an actual clinical session is a genuinely hard ask for an independent researcher: it requires a clinic partnership, a full clinical IRB (not just an expedited one), liability considerations around recording real therapy sessions, and equipment logistics most solo researchers don't have. Proposing that as the first step would be setting this protocol up to never run. Instead: **Phase 2a — proof of concept, lower barrier.** Recruited pairs (friend dyads, or strangers matched for a structured "active listening" task — one person shares something real for 10 minutes, the other listens without interrupting, then they switch) with simultaneous EEG. Each participant completes a Clarity free-write alone, immediately before the paired session. This is the same population-access difficulty as any hyperscanning study (need two people and two EEG setups in one room at once) but does not require a clinical population, a treating clinician's cooperation, or a clinical IRB — a standard human-subjects IRB covers it. **Phase 2b — the actual target, gated on 2a.** If 2a clears r ≥ .30 on H1, take the validated pipeline (task, metric, analysis code) to a real therapist-patient population, via a clinic partnership. This is the study that actually tests the claim this research program cares about. It should not be attempted first. ## 5. Coupling metrics (pre-registered, plural, ranked) 1. **Joint graph Laplacian eigenspectrum (primary).** Treat both participants' EEG channels (e.g., 32 + 32 = 64, matching the ds007471 electrode count) as one combined graph. Build the adjacency matrix from cross-participant channel coherence in a chosen band, weight within-participant edges separately from between-participant edges, take the graph Laplacian, and extract the Fiedler value — the same core mathematical object Clarity computes from a sentence graph, applied here to a neural graph instead of a discourse graph. This is the metric the theory actually predicts should carry signal, and it has never been tested in this research program before. It is untested and could easily fail; that's the point of testing it properly instead of assuming it. 2. **Imaginary part of coherency**, a standard hyperscanning-literature metric robust to volume conduction (same rationale as the PLI robustness argument in the ds007471 write-up, different specific measure). 3. **Alpha-band PLI**, the same metric from ds007471, included specifically as the direct replication check (H3). All three are computed on the same data, for every session, regardless of which hypothesis they map to — deciding post hoc which one to report would defeat the purpose of pre-registering this list. ## 6. Sample size Dyadic hyperscanning correlational designs in the published literature (per the interpersonal-synchrony research cited in `breath-as-the-missing-channel.md`) commonly run N = 20–55 dyads. Given the ds007471 experience — a real, non-trivial effect can still wash out under a naive metric even with 29 dyads and over 1,000 trials — this design should not assume a smaller N is adequate just because it's a hyperscanning study. Using the same Fisher z approach as Phase 1, detecting r = .35 at α = .05, power = .80 requires **N ≈ 62 dyads**. That is the target for Phase 2a, not a smaller convenience number picked because dyadic recruitment is harder than individual recruitment — the statistics don't get easier just because the logistics do. ## 7. Equipment and infrastructure this actually requires Stated plainly rather than assumed: this needs two synchronized EEG systems capable of recording simultaneously in one room, with a shared or synchronized clock (clock drift between two independent amplifiers is a real, well-documented hyperscanning problem, not a detail — session data is unusable without solving it), a quiet room for two people, and either owned equipment or a university/lab partnership that has it. This is not a Cloudflare-and-a-laptop project the way Phase 1 is. It is the reason this protocol explicitly recommends institutional partnership (matching the "open to IRB and collabs" posture already adopted for this research program) rather than attempting to self-fund research-grade dual EEG hardware. ## 8. Analysis Plan - **Primary:** Pearson/Spearman correlation, pre-session `displacementIndex` vs. joint-Laplacian coupling metric (H1); coupling metric vs. post-conversation connection scale (H2). - **Mixed-effects consideration:** if each participant contributes data as both listener and sharer (within-dyad role switch), role should be entered as a fixed effect and dyad as a random intercept — the same non-independence lesson from ds007471 applies here even at N=62 dyads, since each dyad contributes two directional data points. - **H3 replication check:** same PLI computation as ds007471, same reporting standard — point estimate and confidence interval regardless of significance, not just a pass/fail on p < .05. - **Multiple-metric correction:** Holm correction applied across the three coupling metrics tested against each hypothesis, since testing three metrics against the same outcome is exactly the kind of multiple-comparisons problem that needs correcting, not quietly ignoring in favor of whichever one comes back significant. ## 9. Risks & Data Privacy Minimal physical risk (EEG is non-invasive). The active-listening task in Phase 2a asks participants to share something real with a partner — lower intensity than a trauma-focused task, but real disclosure is still real disclosure; consent materials should say plainly that participants choose what they share and can stop at any time, same crisis-resource floor as Phase 1. Phase 2b, involving actual clinical sessions, carries materially higher privacy stakes (recorded EEG timed to real clinical disclosure) and needs its own dedicated privacy and clinical-IRB review — not inherited wholesale from this document. ## 10. Feasibility & Gating - **This does not proceed until Phase 1 clears r ≥ .30.** That is not a formality — Phase 1 is the cheap, fast, fully-online gate specifically so that this expensive, access-constrained study is not attempted on a hunch. - **Phase 2a requirement:** access to a two-person EEG-capable space — most realistically via a university lab partnership, not solo equipment purchase. - **Phase 2b requirement:** a clinic partnership and a clinical IRB, pursued only after 2a shows real signal. - **What's actually written and ready now:** the theoretical rationale, the metric selection (including the primary metric — joint graph Laplacian — that directly tests this research program's own claim rather than borrowing an off-the-shelf one), the sample size, and the analysis plan. What's missing is access, not design.