# Controlled Peer-Baseline Attribution and Early-Layer Native Readout for Predicting Intervention-Specific Operational Repairs on Trace-Silent Multi-Premise Errors

## 1. Introduction

When a language model evaluates multiple premises and emits a bare wrong binary decision (`"No"` or `"Yes"`) without a reasoning trace, white-box activation readout aims to recover specific model-grounded detail about the failure without fabricating unsupported claims [1], [2], [3], [4]. Recent studies on `faithful_activation_readout` establish that when reasoning traces are visible in text or single stored facts can be elicited in clean isolated prompts, text-based readers and target-model logit lenses match or outperform learned Activation Oracles (AOs), whereas unrelated-donor contrasts primarily detect prompt membership [5], [6], [7], [8], [9]. On **trace-silent multi-premise errors**—where constrained binary logits, unrestricted greedy decoding, and clean single-premise re-queries (`59.4%`–`100.0%` healed in isolation) leave the erroneous forward pass unverbalized—evaluating activation readout requires resolving four methodological questions:

1. **Intervention-Dependent Operational Targets vs. Trial-Edit Access:** When all four clauses in a conjunction satisfy their inequality ($x_1 + x_2 \ge L$ or $\le L$) yet the model answers `"No"`, replacing a single clause under an edit family $\mathcal{E}$ defines an **operational repair set** $S_{\mathcal{E}}(x) \subseteq \{A, B, C, D\}$—not an unmediated label of an internal calculation mistake. Different edit families ($R_{\text{easy}}$, with large arithmetic slack shifts, vs. zero-margin same-sum carry-free rewrites $R_{\text{carryfree}}$, with $\Delta m = 0$) can flip the decision via distinct clauses on the **same** prompt. Moreover, if deployment allows applying the **exact gating edit family** to all four clauses (`Matched_4Edit_Trial_Copy`), that black-box trial recovers $1.0000$ on uniquely gated cohorts by construction in $0.11\text{--}0.13\,\text{s/item}$, whereas applying a mismatched edit family (`Cross_4Edit_Trial_Copy`) or operating without pre-specified synthetic value rewrites exposes the boundary where single-pass white-box attribution is informative.
2. **In-Target vs. Cross-Target Text and Position-Only Shortcuts:** Uniform random guessing ($1/4$ or $1/3$) is not a text-only ceiling. Even when nominal clause labels (`A..D`) are balanced, model primacy/recency sensitivities can concentrate operational flips at specific serial positions, enabling development-chosen **position-only** (`Dev_PosOnly`) and **in-target** supervised text-feature rankers to achieve high accuracy within the same edit family while collapsing across edit families. Comparing **parameter-free attribution** (no fitted attribution weights, with layer and decoder hyperparameters selected on `Dev`) against both in-target and cross-target position/text rankers isolates what requires target-specific supervision versus white-box activation access.
3. **Polarity Inversion and Within-Prompt Common-Mode Bias:** Forward linear probes trained on $\ge$ compliance states invert on $\le$ budget caps (`0.0833` on `Qwen3-8B` `Conf_F4` $\le$). Uncentered gradient-input products $\langle \mathring{g}_\ell(t), h_\ell(t) \rangle$ avoid fitted probe weights yet degrade on reversed polarity (`AUC = 0.5620` on `Conf_F2`) because candidate spans share strong within-prompt common-mode anisotropy ($\cos \in [0.868, 0.908]$) whose gradient projection varies across clauses comparably to peer-centered attribution.
4. **Exact Specification-to-Code Alignment, Dual Endpoints, and Retained Timed Outputs:** Evaluating path integration, native self-decoding, and `Route A` candidate selection requires exact mathematical correspondence to executed operations—distinguishing Euclidean straight-chord integration from spherical tangent-velocity integration, direct placeholder replacement from norm rescaling, pure 48-pass peer-twin margin drops (`PeerTwin_MarginDrop48`) from hybrid IG+AO scoring (`PeerTwin_Hybrid`), and consistent clause+tuple recovery (`TupleV5`) from explicit directional-threshold transcription (`RelV5`)—while retaining 100% of generated strings from synchronized cold-cache timing runs.

Across `Qwen3-8B` (36 layers) and `Llama-3.1-8B-Instruct` (32 layers) using an **even-sum development ($\text{parity}=0$) $\to$ untouched odd-sum confirmation ($\text{parity}=1$)** split (`6,552` retained predictions verified across both `TupleV5` and `RelV5` in `supplement/replay.py` with `0` discrepancies), we establish four findings:

- **Operational Repair Targets Are Intervention-Dependent (Section 4.2, Table 2A):** On `Llama-3.1-8B` `Conf_F4` ($R_{\text{easy}}$-gated, $N=24, K=13$ source clusters), zero-margin $R_{\text{carryfree}}$ ($\Delta m = 0$) flips $\ge 1$ clause on **$22/24$ ($91.67\%$)** of items under both constrained logits and unrestricted generation and flips uniquely on **$6/24$ ($25.00\%$)** ($11/24$ unrestricted), containing the $R_{\text{easy}}$ target on $12/24$ ($50.00\%$ constrained; $7/24$ unrestricted) while uniquely matching $R_{\text{easy}}$ on **$0/24$ ($0.00\%$)** (`Qwen3` `Conf_F4`, $N=48, K=27$: $47.92\%$ any-flip, $31.25\%$ unique, $22.92\%$ unique agreement). Consequently, `Matched_4Edit_Trial_Copy` recovers $1.0000$ ($0.1266\,\text{s}$ `Qwen3`, $0.1127\,\text{s}$ `Llama`) when the gating edit family is known and prompt editing is permitted, whereas `Cross_4Edit_Trial_Copy` drops to $0.2292$, $0.0625$, $0.0000$, and $0.4167$ across the four cohorts.
- **Tokenwise Common-Mode Bias & Controlled Path Integration Gain (Sections 4.1–4.2):** Within-prompt common-mode bias has standard deviation $0.86\text{--}6.47\times$ that of 1-step peer attribution patching (`Peer-EAP`), overturning 1-step rankings on $62.5\%\text{--}66.7\%$ of four-clause items. Holding $\mathcal{L}$, $\mathcal{P}$, $\mu_\ell$, and early native decoders fixed (`parameter-free attribution`, `Dev`-selected layers/decoder), `3-Step Euclidean Peer-EAP-IG` reaches $0.6042$ vs. $0.5625$ `1-Step` on `Qwen3` `Conf_F4`, $0.6667$ vs. $0.6250$ on `Llama` `Conf_F4` (beating both `Dev_PosOnly_InTarget` at $0.2500$, $\Delta = +0.4167$ $[+0.039, +0.750]$, and `InTarget16` text ranking at $0.2083$, $\Delta = +0.4583$ $[+0.143, +0.750]$), and **$18/32 = 0.5625$ vs. $8/32 = 0.2500$ `1-Step`** on zero-margin `Qwen3` `Conf_F4_CF` ($\Delta = +0.3125$ $[+0.097, +0.563]$, $K=14$), with **identical scores under both `TupleV5` and `RelV5` across 100% of native readout and copy arms**.
- **In-Target Position Shortcuts vs. Parameter-Free Robustness (Sections 4.2–4.3):** Auditing 2D $(\text{Label} \times \text{Serial Position})$ distributions (`scripts/readout_certifier.py`) reveals that swap-in-place balancing equalized `A..D` ($25\%$ each) while leaving $R_{\text{carryfree}}$ flips concentrated at serial slot $0$ ($24/32 = 75.0\%$ on `Qwen3`; $21/24 = 87.5\%$ on `Llama`) and `Llama Conf_F4` $R_{\text{easy}}$ targets concentrated at slot $3$ ($15/24 = 62.5\%$, though `Llama Dev_F4` had `[4,6,2,0]` and selected slot $1$). Thus `Dev_PosOnly_InTarget` and `InTarget16` reach $0.7500$ (`Qwen3`) and $0.8750$ (`Llama`) on `Conf_F4_CF` ($100\%$ on slot $0$, **$0/8 = 0.0000$ on slots $>0$**, where `3-Step Peer-EAP-IG` achieves $6/8 = 0.7500$ vs. $4/8 = 0.5000$ `1-Step`), yet collapse to $0.0000\text{--}0.3750$ across target definitions. On templated two-hop logs (`Conf_B1`), positional text rankers ($0.4750\text{--}0.5833$) similarly match or exceed activation readout ($0.3250\text{--}0.4000$).
- **Reconciled Cold-Cache Pipeline Latencies & Selector Separation (Sections 4.3–4.4, Table 3):** Across three synchronized cold-cache repeats on `Qwen3` `Conf_F4` with 100% of generated strings retained (`48/48` identical to saved predictions), direct `3-Step Peer-EAP-IG + NativeL6` runs in **$1.5956 \pm 0.0019\,\text{s/item}$** ($0.6042$; `1-Step`: $0.9939 \pm 0.0017\,\text{s}$, $0.5625$) versus `Route A` candidate-pool pipelines: $4.1097 \pm 0.0157\,\text{s}$ (`AO Likelihood`, $0.2500$), $4.5150 \pm 0.0045\,\text{s}$ (`Donor PMI`, $0.2917$), $4.8058 \pm 0.0202\,\text{s}$ (`Span-PMI`, $0.6042$), **$5.2549 \pm 0.0065\,\text{s}$ (`PeerTwin_MarginDrop48`, pure target-model margin drop, $29/48 = 0.6042$)**, and **$6.3230 \pm 0.0155\,\text{s}$ (`PeerTwin_Hybrid`, margin drop + 3-step IG + routed AO log-prob, $28/48 = 0.5833$, differing on $5/48$ choices)**.

---

## 2. Related Work and Scope

**Activation Readers, Native Lenses, and Access Contracts:** LatentQA [2], Activation Oracles [1], [3], and Natural Language Autoencoders [4] translate residual states into text via learned decoders, whereas `LogitLens` [10], `Patchscopes` [11], `SelfIE` [12], and `Tuned Lens` [13] employ frozen model weights. Within `faithful_activation_readout`, prior studies established local number readability versus end-window dilution [14], split-scale localization and verification [15], cross-adapter transfer and native late-layer logit-lens verification on `PersonaQA` [9], [16], difference readers on fine-tunes [8], [17], sibling-contrast preference optimization with active Yes/No verification [18], activation-only layer-6 routing [6], and structured expression rules for multi-variable histories [19]. Under explicit access contracts [5], [7], activation readout earns diagnostic value only where matched-access text, donor states, and allowed black-box queries do not already resolve the target.

**Attribution Precedents and Narrow Scope:** Straight-line path integration is **Integrated Gradients (IG)** [20], first-order residual patching is **Attribution Patching (`EAP`)** [21], covariance scaling relates to **Natural Gradient** [22], and **EAP-IG** [23] (including Appendix D on activation-space integration and low-step approximations) integrates gradients between baseline and clean activations when endpoint gradients saturate; `Patchscopes` [11] covers native explanation via patched hidden states. Our novelty is strictly narrower than a new attribution or decoding algorithm: we provide a controlled, fixed-configuration evaluation of intervention-specific repair prediction, text/position shortcut boundaries, and cold-cache deployment costs on trace-silent multi-premise errors.

---

## 3. Task Construction, Implemented Equations, and Dual-Endpoint Validation

### 3.1 Mining Protocol, Edit Families, Comparators, and Worked Examples

**Development/Confirmation Separation (`mine_4clause_audit`):** Every arithmetic tuple $(x_1, x_2, L)$ satisfies $(x_1 + x_2) \bmod 2 = 0$ in development (`Dev`, used for layer/decoder selection, `Dev_PosOnly` slot selection, and fitting supervised text rankers) and $(x_1 + x_2) \bmod 2 = 1$ in prospective confirmation (`Conf`, evaluated once). Four-clause prompts sample four tuples from an identical $(\text{margin}, \text{sum\_band}, c_1, c_{10}, c_{100})$ parity bucket ($x_1 + x_2 \ge L$ compliance audits or $x_1 + x_2 \le L$ budget caps; `sum_hi = 1721` for $R_{\text{easy}}$-gated `Conf_F4`, `sum_hi = 999` for $R_{\text{carryfree}}$-gated `Conf_F4_CF`), shuffle display order $\pi \in S_4$, retain prompts where constrained binary logits predict wrong `"No"`, and require a unique constrained `"No" \to "Yes"` flip under `gate_mode`. When a source quartet (`source_id`) flips uniquely at clause $c_0$ occupying serial slot $p_0 = \pi^{-1}(c_0)$, up to three label-swapped siblings swap both the tuple and prompt-order entry into slot $p_0$, equalizing `A/B/C/D` ($25\%$ each) while preserving the natural serial-slot distribution $p_0$ (`audited in Section 4.2`).

**Edit Families & Black-Box Trial Comparators:** Given $s = x_1 + x_2 = 100H + 10T + U$: (i) **$R_{\text{carryfree}}$ ($\Delta m = 0$)** sets $x_1' = 100\lfloor H/2\rfloor + 10\lfloor T/2\rfloor + \lfloor U/2\rfloor$ and $x_2' = s - x_1'$ ($x_1' + x_2' = s$, zero carries for $s \le 999$); (ii) **$R_{\text{lim12}}$ ($\Delta m = +12$)** shifts $L \mapsto L \mp 12$; and (iii) **$R_{\text{easy}}$ ($\Delta m \gg 0$)** substitutes a parity-matched easy template (`800+801 >= 1000` / `200+201 >= 300` on $\ge$; `$200k+$201k <= $900k` / `$400k+$401k <= $1500k` on $\le$). **`Matched_4Edit_Trial_Copy`** applies the cohort's gating edit family to each of the four clauses in four forward passes and copies the uniquely flipping clause (or abstains if $|S| \neq 1$); **`Cross_4Edit_Trial_Copy`** applies the alternate edit family ($R_{\text{carryfree}}$ on $R_{\text{easy}}$ cohorts; $R_{\text{easy}}$ on $R_{\text{carryfree}}$ cohorts).

**Development Position-Only & Supervised Text Rankers (`Zero Activations`):** **`Dev_PosOnly_InTarget_Copy`** (`and CrossTarget`) selects the single serial slot $p^*_{\text{dev}} = \arg\max_{p \in \{0..3\}} \text{count}_{\text{dev}}(p)$ on the matching (`or alternate`) `Dev` split and copies the clause at slot $p^*_{\text{dev}}$. **`InTarget16` / `CrossTarget16`** (`clause_text_feature_matrix`) fits ridge regression ($\lambda = 1.0$) on `Dev` over a within-prompt mean-centered $(4 \times 16)$ matrix $\Phi(x)$: 4 serial-position one-hots, 4 label one-hots, $m_c/10$, $(c_1, c_{10}, c_{100})$, $s_c/1000$, $|x_{c,1}-x_{c,2}|/500$, $(u_1+u_2)/18$, and $(t_1+t_2)/18$ (`InTarget24` adds 8 carry-slack/digit-parity features).

**Worked Examples (`Qwen3 Conf_F4_CF_2` & `Llama_Conf_F4_1`):** On `Qwen3 Conf_F4_CF_2` (`A` $(212,439,650)$, `B` $(207,406,612)$, `C` $(208,419,626)$, `D` $(238,409,646)$), $R_{\text{carryfree}}$ flips uniquely to `"Yes"` on `Audit C`; at $\mathcal{L}=\{15,18\}$ over $\mathcal{P}$, `1-Step Peer-EAP` (`[+14.42, +3.11, -3.52, +5.42]`) misranks `C` last while `3-Step Peer-EAP-IG` (`[+6.08, -3.25, +6.19, +1.83]`) selects `C`, and `NativeL6` transcribes `Audit C (Edge Sector): 208 inbound units plus 419 reserve units (minimum required: 626 units).` On `Llama_Conf_F4_1`, $R_{\text{easy}}$ flips uniquely on `Audit B` whereas $R_{\text{carryfree}}$ flips uniquely on `Audit C`.

### 3.2 Exact Implemented Attribution and Native-Decoding Operations

Let $m(x) = z_T(y_{\text{out}}) - z_T(y_{\text{alt}}) > 0$ be the wrong-minus-correct binary logit margin. Each clause $c \in \{1, \dots, C\}$ spans full clause tokens $\text{sps}_c$ and post-header predicate tokens $\text{prs}_c \subset \text{sps}_c$; let $\mathcal{P} = \bigcup_{c=1}^C \text{prs}_c$. At layer $\ell \in \mathcal{L}$, let $h_\ell(t) \in \mathbb{R}^d$ be the clean residual output, $\mathring{g}_\ell(t) = \frac{\partial m(x)}{\partial h_\ell(t)} - \mathbb{I}[t=T]\frac{\partial m(x)}{\partial h_L(T)}$ the skip-subtracted cotangent (`extract_ract_states`), and $\mu_\ell = \frac{1}{|\mathcal{P}|}\sum_{p \in \mathcal{P}} h_\ell(p)$, $\hat{\mu}_\ell = \mu_\ell / \|\mu_\ell\|_2$.

**1-Step Decomposition and Frobenius-Normalized `Peer Cos` (`compute_controlled_attribution_suite`, `scripts/exp56_core_v3.py`):** Decomposing $h_\ell(t) = \alpha_{\parallel,\ell}(t)\hat{\mu}_\ell + h_{\perp,\ell}(t)$ yields:
$$S_{\text{unc}}(c) = \sum_{\ell \in \mathcal{L}} \sum_{t \in \text{prs}_c} \langle h_\ell(t), \mathring{g}_\ell(t) \rangle = \underbrace{\sum_{\ell \in \mathcal{L}} \sum_{t \in \text{prs}_c} \langle \mu_\ell, \mathring{g}_\ell(t) \rangle}_{\text{CommonModeBias}(c)} + \underbrace{\sum_{\ell \in \mathcal{L}} \sum_{t \in \text{prs}_c} \langle h_\ell(t) - \mu_\ell, \mathring{g}_\ell(t) \rangle}_{S_{\text{EAP}}^{(1)}(c)}.$$
For clause $c$ at layer $\ell$, let $G_{\ell,c} = [\mathring{g}_\ell(t)]_{t \in \text{prs}_c}$, $\Delta_{\ell,c}^{\text{euc}} = [h_\ell(t) - \mu_\ell]_{t \in \text{prs}_c}$, and $\Delta_{\ell,c}^{\text{sph}} = [h_\ell(t) - \|h_\ell(t)\|_2 \hat{\mu}_\ell]_{t \in \text{prs}_c}$. `1-Step Euclidean Peer Cos` and `1-Step Norm-Matched Peer Cos` normalize each clause's inner product **per layer by the clause gradient and displacement Frobenius norms**:
$$S_{\text{cos,euc}}^{(1)}(c) = \sum_{\ell \in \mathcal{L}} \frac{\langle G_{\ell,c}, \Delta_{\ell,c}^{\text{euc}}\rangle_F}{(\|G_{\ell,c}\|_F + 10^{-8})\|\Delta_{\ell,c}^{\text{euc}}\|_F + 10^{-8}}, \qquad S_{\text{cos,sph}}^{(1)}(c) = \sum_{\ell \in \mathcal{L}} \frac{\langle G_{\ell,c}, \Delta_{\ell,c}^{\text{sph}}\rangle_F}{(\|G_{\ell,c}\|_F + 10^{-8})\|\Delta_{\ell,c}^{\text{sph}}\|_F + 10^{-8}}.$$

**Euclidean vs. Spherical Multi-Step Midpoint Integration (`run_midpoint_peer_ig_on_positions`, `scripts/exp56_core_v3.py`):** For $m \in \{3, 9\}$ midpoint knots $\alpha_k = \frac{2k + 1}{2m}$ ($k = 0, \dots, m-1$), at each layer $\ell \in \mathcal{L}$ independently, a forward hook replaces all $t \in \mathcal{P}$ simultaneously with $u_{\alpha_k,\ell}(t)$ (`leaving` $t \notin \mathcal{P}$ `unmodified`) and computes $g_\ell^{(k)}(t) = \left.\frac{\partial m}{\partial u_{\alpha_k,\ell}(t)}\right|$:
- **Euclidean `Peer-EAP-IG` (`norm_matched = False`, primary white-box method):** Uses constant peer-mean baseline $u_{0,\ell}^{\text{euc}}(t) = \mu_\ell$, straight-line chord $u_{\alpha_k,\ell}^{\text{euc}}(t) = (1 - \alpha_k)\mu_\ell + \alpha_k h_\ell(t)$, and fixed velocity $v_{\alpha_k,\ell}^{\text{euc}}(t) = h_\ell(t) - \mu_\ell$:
  $$S_{\text{euc}}^{(m)}(c) = \sum_{\ell \in \mathcal{L}} \sum_{t \in \text{prs}_c} \frac{1}{m}\sum_{k=0}^{m-1} \left\langle g_\ell^{(k)}(t),\; h_\ell(t) - \mu_\ell \right\rangle.$$
- **Spherical Norm-Matched `Peer-IG` (`norm_matched = True`):** Sets $u_{0,\ell}^{\text{sph}}(t) = \|h_\ell(t)\|_2 \hat{\mu}_\ell$ and $r_{\alpha_k,\ell}(t) = (1-\alpha_k)u_{0,\ell}^{\text{sph}}(t) + \alpha_k h_\ell(t)$ with $\rho_{\alpha_k,\ell}(t) = \|r_{\alpha_k,\ell}(t)\|_2 + 10^{-8}$ and $\hat{r}_{\alpha_k,\ell}(t) = r_{\alpha_k,\ell}(t)/\rho_{\alpha_k,\ell}(t)$. Every intermediate knot is **projected onto the token's norm sphere** $u_{\alpha_k,\ell}^{\text{sph}}(t) = \|h_\ell(t)\|_2 \hat{r}_{\alpha_k,\ell}(t)$ and integrated against the **position-dependent spherical tangent velocity** $\frac{\partial u_{\alpha,\ell}^{\text{sph}}(t)}{\partial \alpha}\big|_{\alpha=\alpha_k}$:
  $$v_{\alpha_k,\ell}^{\text{sph}}(t) = \frac{\|h_\ell(t)\|_2}{\rho_{\alpha_k,\ell}(t)}\Bigl((h_\ell(t) - u_{0,\ell}^{\text{sph}}(t)) - \bigl\langle h_\ell(t) - u_{0,\ell}^{\text{sph}}(t),\; \hat{r}_{\alpha_k,\ell}(t)\bigr\rangle \hat{r}_{\alpha_k,\ell}(t)\Bigr), \qquad S_{\text{sph}}^{(m)}(c) = \sum_{\ell \in \mathcal{L}} \sum_{t \in \text{prs}_c} \frac{1}{m}\sum_{k=0}^{m-1} \left\langle g_\ell^{(k)}(t),\; v_{\alpha_k,\ell}^{\text{sph}}(t) \right\rangle.$$

**Target-Native Early-Layer Span Decoding (`decode_span_native_l6`, `scripts/ract_core.py`):** Given $c^* = \arg\max_c S(c)$, let $R^* = \{0\} \cup \text{sps}_{c^*}$ (`BOS/sink` token $0$ plus full clause tokens). We format a fresh target-model chat prompt containing **zero source text**: `"Span:" + (" ?" * |R^*|) + " \nTranscribe verbatim the single statement shown in the span above (one line only)."` (`enable_thinking = False` on `Qwen3-8B`). During prefill (`max_new_tokens = 44`, `do_sample = False`), a forward hook at $\ell_{\text{dec}}$ ($\ell_{\text{dec}} = 6$ on `Qwen3-8B`; $\ell_{\text{dec}} = 5$ on `Llama-3.1-8B`) **directly assigns** $h_{\ell_{\text{dec}}}(R^*_i)$ to the $|R^*|$ placeholder positions (`hs[0, p0 : p0 + K, :] = h_dev`), retaining source activation magnitudes without norm rescaling, and returns `first_nonempty_line`.

**Development Hyperparameter Selection (`parity = 0`, Uses `Dev` Target Labels):** Target labels on even-sum `Dev` inform only fixed architectural hyperparameters (`zero` fitted attribution weights): oracle-span native readout on `Qwen3 Dev` ($N=16$) achieves $1.0000$ at $L_3, L_6, L_9$ before dropping to $0.6250$ at $L_{12}$ (`NativeL6` frozen), and `[15, 18]` balances `Dev` $R_{\text{easy}}$ and $R_{\text{carryfree}}$ localization ($0.5625$ `3-Step` vs. $0.4688$ `1-Step` on `Dev_F4_CF`). On `Llama Dev_F4` ($R_{\text{easy}}$), `[14, 16]` + `NativeL5` was frozen; on `Llama Dev_F4_CF` ($N=16$), 1-step carry-free sensitivity instead peaks at early layers ($L_5 = 0.7500, L_8 = 0.6875$) and drops to $0.0000$ at `[14, 16]`, explaining `[14, 16]`'s weak transfer to `Llama Conf_F4_CF` (Appendix A.1).

### 3.3 Dual-Endpoint Validator (`TupleV5` vs. `RelV5`, `scripts/exp58_core_v5.py`)

Because `TextOnly_ZeroShot_LLM_Diag`, `PostHoc_SelfExplain`, and `Route B` prompts request concise `(letter, x1, x2, limit)` answers (`emitting e.g. "C 208 419 626"` or `"Item A: 208 plus 419, limit 626."`), whereas clause-copying and `NativeL6`/`NativeL5` emit full inequality statements, `audit_multiclause_v5` defines and audits **two formal endpoints** (each partitioning responses into 5 mutually exclusive categories summing to $1.0000$):
1. **Primary Endpoint — `Consistent Clause+Tuple Recovery` (`TupleV5`, `require_explicit_relation = False`):** Accepts concise ordered/labeled tuples or full transcriptions; requires the exact target clause letter set, local per-clause presence of $\{x_1, x_2, L\}$ respecting multiplicity when $L = x_1 + x_2$, no foreign numbers (except $s = x_1 + x_2$ alongside $L$), and **zero role, threshold, polarity, or equation contradictions**. Pre-nominal and post-nominal threshold bindings (`required amount NNN`, `NNN units required`, `minimum/maximum/limit ... NNN`) must equal $L$, and operand-role tags (`inbound`/`hardware` $= x_1$, `reserve`/`network` $= x_2$) must match—rejecting `"Audit C: 208 inbound units plus 419 reserve units; required amount 208; recorded values 626 and 419."` (`cat_unsupported_or_contradicted_number = 1`, verified by 8 unit tests in `replay.py`).
2. **Secondary Endpoint — `Explicit Directional-Threshold Transcription` (`RelV5`, `require_explicit_relation = True`):** Requires `TupleV5` **plus** an explicit polarity-matching directional phrase or symbol binding $L$ (`minimum/required/at least/>= L` on $\ge$; `maximum/allowed/at most/within/cap/<= L` on $\le$). Across all four 4-clause cohorts (`Appendix A.2`), **100% of native activation-readout and clause-copying arms have identical `TupleV5` and `RelV5` scores (`0` disagreements)**, whereas `ZeroShot_LLM` and `PostHoc_SelfExplain` drop to `0.0000` on `RelV5` solely because their responses omit inequality words. Uncertainty intervals use `2,000`-draw `source_id`-clustered bootstrap 95% CIs (`primary`) alongside Wilson CIs.

---

## 4. Empirical Results

### 4.1 Direct Common-Mode Geometry and Two-Clause Benchmarks (Figure 1, Table 1)

![Figure 1: Strict ceteris paribus (L15+18 on prs) reversed-polarity Conf_F2 (<=) ROC-AUC with direct common-mode measurements (Panel A), and 2-clause polarity/calendar transfer across Conf_F1 (>=), Conf_F2 (<= OOD), and Conf_F3 (Calendar OOD) on Qwen3-8B (Panel B).](supplement/figure-1.png)

Across retained 2-clause confirmation items (`100.0%` unrestricted `"No"` agreement), tokenwise $|\alpha_{\parallel,\ell}(t)| / \|h_{\perp,\ell}(t)\|_2$ has median $1.10\text{--}1.21$ on `Qwen3-8B` and $0.83$ on `Llama-3.1-8B`, while $\text{SD}(\text{CommonModeBias}) / \text{SD}(S_{\text{EAP}}^{(1)})$ averages $1.91$ (`Conf_F1`), $2.00$ (`Conf_F2`), and $6.47$ (`Conf_F3`) on `Qwen3-8B` ($0.98$ on `Llama` `Conf_F1`), overturning 1-step `Peer-EAP` rankings on $22.7\%\text{--}37.5\%$ of 2-clause items and $62.5\%\text{--}66.7\%$ of 4-clause items.

| `Qwen3-8B` 2-Clause Split ($\mathcal{L}=\{15,18\}$ on $\mathcal{P}$) | Frozen Even-Sum Probe `Acc / AUC` | Uncentered `Grad*Act` `Acc / AUC` | `1`-Step Euclid Peer-EAP `Acc / AUC` | `1`-Step Euclid Peer Cos `Acc / AUC` | `3`-Step Euclid Peer-IG `Acc / AUC` | `3`-Step Sph Peer-IG `Acc / AUC` | `9`-Step Euclid Peer-IG `Acc / AUC` |
| :--- | :---: | :---: | :---: | :---: | :---: | :---: | :---: |
| `Conf_F1` ($\ge$, $N=40, K=38$) | `0.8750 / 0.9475` | `0.6500 / 0.7500` | `0.7000 / 0.7700` | **`0.7250 / 0.7750`** | `0.6250 / 0.7100` | `0.6500 / 0.7450` | `0.6250 / 0.7200` |
| `Conf_F2` ($\le$ OOD, $N=22, K=21$) | `0.4545 / 0.4959` | `0.5909 / 0.5620` | `0.6364 / 0.6033` | **`0.6818`** / `0.7355` | **`0.6818 / 0.7934`** | `0.6364 / 0.6694` | **`0.6818`** / `0.7769` |
| `Conf_F3` (Cal OOD, $N=40, K=35$) | `0.5000 / 0.5600` | `0.5000 / 0.5275` | `0.6250 /` **`0.8700`** | `0.6500 / 0.8500` | `0.5000 / 0.6050` | `0.5250 / 0.6225` | `0.5750 / 0.6275` |

On `Llama-3.1-8B` `Conf_F1` ($N=24, K=24$, $\mathcal{L}=\{14,16\}$), `1-Step Peer-EAP`, `3-Step Euclidean Peer-EAP-IG`, `3-Step Spherical Peer-IG`, and `9-Step Euclidean Peer-EAP-IG` all achieve $0.6667$ accuracy and $0.9583\text{--}0.9722$ ROC-AUC (`3-Step` vs. `9-Step` top-1 agreement $= 1.0000$).

### 4.2 Four-Clause Audits: Repair Taxonomy, Comparators, and Dual Endpoints (Figure 2, Tables 2A–2B)

![Figure 2: TupleV5 4-clause readout accuracy across position-only, text-ranker, cross-edit trial, LLM, and parameter-free attribution arms with cluster-bootstrap 95% CIs (Panel A; all copy/native arms have identical RelV5 = TupleV5), and 4-metric cross-intervention repair taxonomy (Panel B).](supplement/figure-2.png)

**Table 2A — Complete Constrained & Unrestricted Repair-Set Taxonomy (`compute_repair_set_taxonomy`):** Table 2A separates flip **existence** ($|S| \ge 1$), **uniqueness** ($|S| = 1$), **gated-target inclusion** ($t^* \in S$), and **unique target agreement** ($S = \{t^*\}$):

| Cohort & Gating Target $t^*$ | Evaluated Edit Family $\mathcal{E}$ | Constrained Any $\|S\|\ge 1$ | Constrained Unique $\|S\|=1$ | Constrained $t^* \in S$ | Constrained $S=\{t^*\}$ | Unrestricted $S=\{t^*\}$ (`Any / Unique / InSet`) |
| :--- | :--- | :---: | :---: | :---: | :---: | :---: |
| `Qwen3 Conf_F4` ($N=48, t^*=R_{\text{easy}}$) | $R_{\text{easy}}$ (`gate`) | `1.0000` (`48`) | `1.0000` (`48`) | `1.0000` (`48`) | `1.0000` (`48`) | `0.7917` (`0.792 / 0.792 / 0.792`) |
| `Qwen3 Conf_F4` ($N=48, t^*=R_{\text{easy}}$) | $R_{\text{carryfree}}$ ($\Delta m=0$) | `0.4792` (`23`) | `0.3125` (`15`) | `0.3333` (`16`) | `0.2292` (`11`) | `0.1667` (`0.396 / 0.250 / 0.271`) |
| `Qwen3 Conf_F4` ($N=48, t^*=R_{\text{easy}}$) | $R_{\text{lim12}}$ ($\Delta m=+12$) | `0.6875` (`33`) | `0.3958` (`19`) | `0.6250` (`30`) | `0.3333` (`16`) | — |
| `Llama Conf_F4` ($N=24, t^*=R_{\text{easy}}$) | $R_{\text{easy}}$ (`gate`) | `1.0000` (`24`) | `1.0000` (`24`) | `1.0000` (`24`) | `1.0000` (`24`) | `0.7917` (`0.792 / 0.792 / 0.792`) |
| `Llama Conf_F4` ($N=24, t^*=R_{\text{easy}}$) | $R_{\text{carryfree}}$ ($\Delta m=0$) | **`0.9167` (`22`)** | **`0.2500` (`6`)** | **`0.5000` (`12`)** | **`0.0000` (`0`)** | **`0.0000` (`0.917 / 0.458 / 0.292`)** |
| `Qwen3 Conf_F4_CF` ($N=32, t^*=R_{\text{carryfree}}$) | $R_{\text{carryfree}}$ (`gate`) / $R_{\text{easy}}$ | `1.000` / `0.219` | `1.000` / `0.188` | `1.000` / `0.063` | `1.000` / `0.063` | `0.9062` (`0.906 / 0.906 / 0.906` gate) |
| `Llama Conf_F4_CF` ($N=24, t^*=R_{\text{carryfree}}$) | $R_{\text{carryfree}}$ (`gate`) / $R_{\text{easy}}$ | `1.000` / `0.458` | `1.000` / `0.458` | `1.000` / `0.417` | `1.000` / `0.417` | `0.7500` (`0.750 / 0.750 / 0.750` gate) |

**Table 2B — Dual-Endpoint `v5` 4-Clause Readout Accuracy (`TupleV5 [Cluster CI]`; `RelV5` Identical on All Unstarred Rows, `0.0000` on `*` Rows):**

| Pipeline ($\mathcal{L}$ & $\mathcal{P}$ Fixed Within Model) | `Qwen3 Conf_F4` ($R_{\text{easy}}, N=48, K=27$) | `Qwen3 Conf_F4_CF` ($\Delta m=0, N=32, K=14$) | `Llama Conf_F4` ($R_{\text{easy}}, N=24, K=13$) | `Llama Conf_F4_CF` ($\Delta m=0, N=24, K=7$) |
| :--- | :---: | :---: | :---: | :---: |
| **Matched 4-Edit Trial Copy (`4` BB, Known $\mathcal{E}$)** | `1.0000` `[1.000, 1.000]` | `1.0000` `[1.000, 1.000]` | `1.0000` `[1.000, 1.000]` | `1.0000` `[1.000, 1.000]` |
| **Cross 4-Edit Trial Copy (`4` BB, Alt $\mathcal{E}$)** | `0.2292` `[0.089, 0.386]` | `0.0625` `[0.000, 0.156]` | `0.0000` `[0.000, 0.000]` | `0.4167` `[0.130, 0.692]` |
| **Dev Position-Only Copy (`In-Target` $p^*_{\text{dev}}$)** | `0.4792` `[0.250, 0.674]` | **`0.7500`** `[0.500, 0.943]` | `0.2500` `[0.040, 0.522]` | **`0.8750`** `[0.608, 1.000]` |
| **Dev Position-Only Copy (`Cross-Target` $p^*_{\text{dev}}$)** | `0.1250` `[0.019, 0.268]` | `0.1250` `[0.000, 0.286]` | `0.0000` `[0.000, 0.000]` | `0.0000` `[0.000, 0.000]` |
| **In-Target Supervised TextRanker (`16`-feat)** | `0.4583` `[0.239, 0.653]` | **`0.7500`** `[0.500, 0.943]` | `0.2083` `[0.039, 0.450]` | **`0.8750`** `[0.608, 1.000]` |
| **Cross-Target ($R_{\text{easy}}$-Fit) TextRanker** | *(in-target above)* | `0.1250` `[0.000, 0.286]` | *(in-target above)* | `0.3750` `[0.130, 0.637]` |
| **Clean 1-Clause MinMargin (`4` BB, Unedited)** | `0.4167` `[0.245, 0.582]` | `0.6250` `[0.419, 0.813]` | `0.5000` `[0.263, 0.704]` | `0.3750` `[0.083, 0.720]` |
| **TextOnly Zero-Shot LLM Diag `*`** (`RelV5 = 0.0000`) | `0.2708` `[0.157, 0.413]` | `0.3750` `[0.212, 0.560]` | `0.2083` `[0.074, 0.375]` | `0.5417` `[0.286, 0.769]` |
| **Post-Hoc Self-Explain `*`** (`RelV5 = 0.0000`) | `0.2708` `[0.151, 0.417]` | `0.4062` `[0.171, 0.667]` | `0.3333` `[0.154, 0.546]` | `0.3333` `[0.044, 0.640]` |
| **FullSeq / Donor Native `L6/L5` (`0` text)** | `0.0000` / `0.0000` | `0.0000` / `0.0000` | `0.2083` / `0.0000` | `0.0417` / `0.0000` |
| **Uncentered `Grad*Act` + Native** | `0.2708` `[0.128, 0.438]` | `0.1250` `[0.000, 0.250]` | `0.1667` `[0.040, 0.333]` | `0.1250` `[0.000, 0.392]` |
| **`1`-Step Euclid Peer-EAP + Native** | `0.5625` `[0.365, 0.750]` | `0.2500` `[0.063, 0.500]` | `0.6250` `[0.400, 0.880]` | `0.2083` `[0.000, 0.500]` |
| **`3`-Step Euclid Peer-EAP-IG + Native** | **`0.6042`** `[0.413, 0.793]` | **`0.5625`** `[0.324, 0.794]` | **`0.6667`** `[0.407, 0.917]` | `0.2500` `[0.000, 0.609]` |
| **`3`-Step Spherical Peer-IG + Native** | **`0.6250`** `[0.436, 0.804]` | `0.5312` `[0.306, 0.759]` | **`0.6667`** `[0.407, 0.917]` | `0.2917` `[0.000, 0.692]` |
| **`9`-Step Euclid Peer-EAP-IG + Native** | **`0.6250`** `[0.436, 0.804]` | **`0.5938`** `[0.359, 0.821]` | **`0.6667`** `[0.407, 0.917]` | `0.2917` `[0.000, 0.692]` |

Four conclusions follow from Tables 2A–2B:
1. **Controlled Path Integration on `Qwen3` Carry-Free Targets:** Holding `[15, 18]`, $\mathcal{P}$, $\mu_\ell$, and `NativeL6` fixed, `3-Step Euclidean Peer-EAP-IG` lifts zero-margin `Qwen3 Conf_F4_CF` readout from $8/32 = 0.2500$ (`1-Step Peer-EAP`) to $18/32 = 0.5625$ ($\Delta = +0.3125$ $[+0.097, +0.563]$; `9-Step`: $0.5938$; `Spherical 3-Step` minus `Euclidean 3-Step` is $+0.0208, -0.0312, 0.0000, +0.0417$, resolving no advantage over Euclidean straight-line `Peer-EAP-IG`).
2. **Position-Only & In-Target Text Shortcuts vs. Non-Initial Strata:** `Dev` serial-slot counts `(slot 0..3)` are `[4, 19, 0, 9]` (`Qwen3 Dev_F4`, $p^*_{\text{dev}}=1$), `[17, 6, 4, 5]` (`Qwen3 Dev_F4_CF`, $p^*_{\text{dev}}=0$), `[4, 6, 2, 0]` (`Llama Dev_F4`, $p^*_{\text{dev}}=1$), and `[12, 0, 0, 4]` (`Llama Dev_F4_CF`, $p^*_{\text{dev}}=0$), while `Conf` counts are `[6, 23, 0, 19]`, `[24, 4, 4, 0]`, `[0, 6, 3, 15]`, and `[21, 0, 3, 0]`. Thus `Dev_PosOnly_InTarget` and `InTarget16` hit $0.7500$ and $0.8750$ on `Conf_F4_CF` purely by selecting slot $0$ ($100\%$ on slot $0$, **$0/8 = 0.0000$ on slots $>0$** on `Qwen3`, where `3-Step Peer-EAP-IG` scores $6/8 = 0.7500$ vs. $4/8 = 0.5000$ `1-Step`). On `Llama Conf_F4`, despite $15/24$ ($62.5\%$) last-slot (`slot 3`) `Conf` targets, `Llama Dev_F4` selected slot $1$ ($6/24 = 0.2500$), so `3-Step Peer-EAP-IG` ($0.6667$) exceeds both `Dev_PosOnly_InTarget` ($\Delta = +0.4167$ $[+0.039, +0.750]$) and `InTarget16` ($\Delta = +0.4583$ $[+0.143, +0.750]$).
3. **Operational `4-Edit` Boundary:** `Matched_4Edit_Trial_Copy` achieves $1.0000$ by construction when the gating edit family is tried directly, whereas `Cross_4Edit_Trial_Copy` abstains on $54.2\%\text{--}81.3\%$ of items and scores $0.0000\text{--}0.4167$ (`paired 3-Step minus Cross_4Edit` $= +0.3750, +0.5000, +0.6667$ on the first three cohorts).
4. **Depth Dissociation on `Llama Conf_F4_CF`:** Keeping `R_easy`-selected `[14, 16]` yields $0.2500\text{--}0.2917$ because `Llama` carry-free sensitivity concentrates at $L_5\text{--}L_8$ (`Appendix A.1`).

### 4.3 Two-Hop Logs (`Conf_B1`), Reconciled `Route A` Selectors, and Cold-Cache Timings (Figure 3, Table 3)

![Figure 3: Conf_B1 two-hop interference comparing in-target positional text rankers against zero-text activation readout (Panel A), and Route A candidate selection across M=4 pools explicitly separating PeerTwin_MarginDrop48 from PeerTwin_Hybrid (Panel B).](supplement/figure-3.png)

On templated two-hop logs (`Conf_B1`: `Qwen3` $N=40, K=10$; `Llama` $N=36, K=10$), an in-target 12-feature positional text ranker reaches $0.4750$ $[0.333, 0.595]$ (`Qwen3`) and $0.5833$ $[0.333, 0.800]$ (`Llama`), matching or exceeding zero-text activation readout ($0.3250\text{--}0.4000$ on `Qwen3`; $0.0833\text{--}0.1111$ at `Llama` `L14+16` and $0.3056\text{--}0.3333$ at `Llama` `L12`).

**Exact `Route A` Selector Definitions (`verify_role_aligned_peer_twin`, `scripts/ract_core.py`):** Across three $M=4$ candidate pools on `Qwen3 Conf_F4`—full-sequence `ours-3layers` sampling, clause-window `ours-3layers`, and clause-window `NativeL6` (`Oracle Pass@4` $= 0.8542$)—we evaluate five parameter-free selectors: (i) **`AO Likelihood`** ($\log p_{\text{AO}}(y_m \mid h_{[9,18,27]})$); (ii) **`Donor PMI`** ($\log p_{\text{AO}}(y_m \mid h_{\text{tgt}}) - \log p_{\text{AO}}(y_m \mid h_{\text{don}})$); (iii) **`Peer-EAP-IG Span-PMI`** (routing $R^*_{\text{IG}} = \{0\} \cup \text{sps}_{c^*_{\text{IG}}}$ via `3-Step Peer-EAP-IG` and contrasting `ours-3layers` log-probability against the donor's matched slot); (iv) **`PeerTwin_MarginDrop48`** (for each clause $j \in \{0..3\}$, patching **whole clause span** $\text{sps}_j$ toward the norm-matched progress-indexed mean of the other three whole clause spans across **four layers $\ell \in \{6, 9, 12, 15\}$** and **three interpolation strengths $\beta \in \{0.50, 0.85, 1.00\}$** in $48$ forward passes, and scoring candidate $y_m$ with parsed clause $c_m$ purely by $\Delta m_{\text{twin}}(c_m)$); and (v) **`PeerTwin_Hybrid`** (scoring $y_m$ by $\Delta m_{\text{twin}}(c_m) + S_{\text{euc}}^{(3)}(c_m) + 0.05 \log p_{\text{AO}}(y_m \mid h_{[9,18,27]}[R^*_{\text{IG}}])$, differing from `PeerTwin_MarginDrop48` on $5/48$ `NativeL6`-pool items: `[7, 12, 15, 23, 38]`).

**Table 3 — Reconciled Cold-Cache Standalone Pipelines (`3` Repeats, 100% Outputs Retained in `exp58c`):**

| Complete Standalone Pipeline (`Qwen3-8B Conf_F4`, $N=48, K=27$) | Cold-Cache Mean $\pm$ SD (`s/item`, $3$ reps) | Neural Passes Executed Inside Loop | `TupleV5` `[Cluster CI]` | `RelV5` |
| :--- | :---: | :---: | :---: | :---: |
| `Dev_PosOnly_InTarget_Copy` ($p^*_{\text{dev}} = 1$) | `0.0000` $\pm$ `0.0000` | `0` | `0.4792` `[0.250, 0.674]` | `0.4792` |
| `TextOnly_Supervised_Ranker_Copy` (`in-target 16-feat`) | `0.00002` $\pm$ `0.0000` | `0` | `0.4583` `[0.239, 0.653]` | `0.4583` |
| `Clean_1Clause_MinMargin_Copy` (`unedited sub-clauses`) | `0.1041` $\pm$ `0.0008` | `4` clean 1-clause fwd | `0.4167` `[0.245, 0.582]` | `0.4167` |
| `Cross_4Edit_Trial_Copy` ($R_{\text{carryfree}}$ trials) | `0.1261` $\pm$ `0.0001` | `4` edited 4-clause fwd | `0.2292` `[0.089, 0.386]` | `0.2292` |
| `Matched_4Edit_Trial_Copy` ($R_{\text{easy}}$ trials, known $\mathcal{E}$) | `0.1266` $\pm$ `0.0005` | `4` edited 4-clause fwd | `1.0000` `[1.000, 1.000]` | `1.0000` |
| `TextOnly_ZeroShot_LLM_Diag` | `0.4483` $\pm$ `0.0026` | `1` greedy gen (`40` tok) | `0.2708` `[0.157, 0.413]` | `0.0000` |
| `PostHoc_SelfExplain` | `0.4990` $\pm$ `0.0134` | `1` 2nd-turn gen (`40` tok) | `0.2708` `[0.157, 0.408]` | `0.0000` |
| **`Direct_1Step_PeerEAP_L15_18_NativeL6`** | **`0.9939` $\pm$ `0.0017`** | `1` clean/grad + `1` `NativeL6` | `0.5625` `[0.365, 0.750]` | `0.5625` |
| **`Direct_3Step_Euclid_PeerEAP_IG_L15_18_NativeL6`** | **`1.5956` $\pm$ `0.0019`** | `1` clean + `6` IG + `1` `NativeL6` | **`0.6042`** `[0.413, 0.793]` | **`0.6042`** |
| `RouteA_NativeL6_M4 + AO_Likelihood` | `4.1097` $\pm$ `0.0157` | `1` ext + `4` `NativeL6` + `4` AO | `0.2500` `[0.135, 0.372]` | `0.2500` |
| `RouteA_NativeL6_M4 + Donor_PMI` | `4.5150` $\pm$ `0.0045` | `2` ext + `4` `NativeL6` + `8` AO | `0.2917` `[0.167, 0.429]` | `0.2917` |
| `RouteA_NativeL6_M4 + PeerEAP_IG_Span_PMI` | `4.8058` $\pm$ `0.0202` | `7` IG + `2` ext + `4` `NativeL6` + `8` AO | **`0.6042`** `[0.413, 0.784]` | **`0.6042`** |
| **`RouteA_NativeL6_M4 + PeerTwin_MarginDrop48`** | **`5.2549` $\pm$ `0.0065`** | `1` ext + `4` `NativeL6` + `48` twin fwd | **`0.6042`** `[0.423, 0.783]` | **`0.6042`** |
| `RouteA_NativeL6_M4 + PeerTwin_Hybrid` | `6.3230` $\pm$ `0.0155` | `7` IG + `4` `NativeL6` + `48` twin + `4` AO | `0.5833` `[0.396, 0.776]` | `0.5833` |

On `Llama-3.1-8B` `Conf_F4`, 3-repeat cold-cache timings with retained outputs are $0.1127 \pm 0.0003\,\text{s}$ (`Matched_4Edit`, $1.0000$), $0.1114 \pm 0.0001\,\text{s}$ (`Cross_4Edit`, $0.0000$), $0.6480 \pm 0.0061\,\text{s}$ (`Direct 1-Step`, $0.6250$), and $1.1649 \pm 0.0236\,\text{s}$ (`Direct 3-Step`, $0.6667$). Direct `3-Step Peer-EAP-IG + NativeL6` matches `Span-PMI` and `PeerTwin_MarginDrop48` accuracy ($0.6042$) at **$3.0\text{--}3.3\times$ lower latency** ($1.60\,\text{s}$ vs. $4.81\text{--}5.25\,\text{s}$) by decoding only the routed clause.

### 4.4 Order-Balanced `Route B`, Coalition Gates, and Prose Boundaries (Figure 4)

![Figure 4: Order-balanced Route B 4-mode transfer audit and balanced 1+2-repair mixture gates (Panel A), and measured 3-repeat cold-cache standalone pipeline latency vs. TupleV5 accuracy frontier with all outputs retained on Qwen3-8B Conf_F4 (Panel B).](supplement/figure-4.png)

On `Conf_Mix` ($24$ single-repair + $24$ two-repair items, $K=48$; identical on `TupleV5` and `RelV5`), **Whole-Clause Peer-Mean Coalition Gate (`CG`)** patches the mean whole-clause residual of bottom-2 peers $(c_{(3)}, c_{(4)})$ at $L_{12}$ ($0.1250$) or $L_6$ ($0.0833$), whereas **Discrete Internal Peer Gate (`DIP`)** patches only predicate tokens $\text{prs}_{c_{(4)}}$ from $c_{(4)}$ at $L_2$ ($0.2708$ $[0.146, 0.396]$) or $L_6$ ($0.3125$ $[0.188, 0.458]$ vs. `Top-1` $0.2708$ $[0.167, 0.396]$), bounded by cross-position `RoPE` mismatch. On `Route B` (`ours-3layers` trained across 3 seeds on 50/50 `[AB, BA]` `Dev_F1` spans), `TupleV5` 4-letter `Conf_F4` accuracy is $0.1944 \pm 0.0120$ (`RelV5`: $0.0000$, as SFT targets used `"Item X: x1 plus x2, limit L."`) due to `A/B` label overfitting ($0.7222$ isolated tuple accuracy vs. $0.4514$ competing-span tuple accuracy). On delimiter-free `ProseAudit`, fixed 30-token windows score $0.0000$ while oracle sentence spans recover $1.0000$ (`Qwen3`) / $0.5000$ (`Llama`) loose operands (`0.0500` / `0.0833` strict threshold).

---

## 5. Discussion, Operational Access Regimes, and Limitations

1. **Operational Trial Edits vs. Single-Prompt Attribution:** On uniquely gated cohorts, if a practitioner knows the exact target edit family $\mathcal{E}$ and can execute edited black-box re-queries, `Matched_4Edit_Trial_Copy` ($0.11\text{--}0.13\,\text{s/item}$, $1.0000$) is faster and more accurate than $1.16\text{--}1.60\,\text{s/item}$ white-box attribution. Parameter-free `3-Step Peer-EAP-IG + NativeL6/L5` addresses the complementary setting where either (a) synthetic value-rewrite templates are unknown or mismatched (`where Cross_4Edit_Trial_Copy drops to 0.0000–0.4167`), or (b) access permits inspecting internal activations of the fixed prompt (or evaluating unedited sub-clauses as in `Clean_1Clause_MinMargin`) without constructing task-specific counterfactual value substitutions.
2. **Position/Text Shortcuts, Small-Cluster Uncertainty, and Mined-Cohort Scope:** When target labels from the evaluation edit family are available and serial target positions are skewed ($75.0\%\text{--}87.5\%$ slot $0$ in `Conf_F4_CF`; hub-order skew in `Conf_B1`), `Dev_PosOnly_InTarget` and `InTarget16` achieve $0.4750\text{--}0.8750$ without activations, yet score $0.0000$ on non-initial `Conf_F4_CF` slots and fail across edit families ($0.0000\text{--}0.3750$). Because source-cluster counts are modest ($K = 7\text{--}27$) and prompts are synthetic mined conjunctions where `NativeL6/L5` transcribes input clauses, our results establish **conditional operational repair-target prediction**—not identification of an internal arithmetic mistake or general superiority over text-based methods.

---

## 6. Reproducibility Statement

`supplement.zip` (`912 KB` compressed, `1.58 MB` expanded) bundles: (1) pure Python 3 stdlib `/usr/bin/python3 replay.py` (`< 4s` CPU), executing 8 unit tests on `audit_multiclause_v5` (`scripts/exp58_core_v5.py`), rescoring all `6,552` retained predictions across both `TupleV5` and `RelV5` (`13,104` evaluations) with `0` discrepancies, and verifying `PeerTwin_MarginDrop48` vs. `PeerTwin_Hybrid` (`5/48` choice differences), `48/48` (`and 24/24`) standalone-to-saved string identity (`data/exp58c_v5_reconciliation.json.xz`), `Dev_PosOnly` and `4-Edit` comparators, 2D exchangeability (`scripts/readout_certifier.py`), and `2,000`-draw `source_id` cluster-bootstrap CIs; and (2) all neural scripts (`exp56a`, `exp56b`, `exp58a`, `exp58b`, `exp58c_v5_reconciliation.py`) with bundled `upstream/AObench/utils/` (`PyTorch 2.8.0+cu128`, `NVIDIA RTX 5090`, `bfloat16`).

---

## Appendix A: Development Selection, Dual-Endpoint Disagreements, and Auxiliary Partitions

### A.1 Even-Sum Development Layer Selection and Layer / Baseline Ablations
- **`Qwen3-8B` `Dev_F4_CF` (`parity=0`, $N=32, K=15$):** `1-Step Peer-EAP` localization is $L_6: 0.5312$, $L_9: 0.4062$, $L_{12}: 0.4062$, $L_{15}: 0.4688$, $L_{18}: 0.2500$, $L_{21}: 0.3125$, $L_{24}: 0.1875$, `[15, 18]`: $0.4688$, `[12..24]`: $0.4688$; `3-Step Peer-EAP-IG` at `[15, 18]` reaches $0.5625$.
- **`Llama-3.1-8B` `Dev_F4_CF` (`parity=0`, $N=16, K=7$):** `1-Step Peer-EAP` peaks at $L_5: 0.7500$ and $L_8: 0.6875$ ($L_{20}: 0.3750$) and drops to $0.0000$ at $L_{11}, L_{14}, L_{16}$, and `[14, 16]`.
- **Confirmation Layer & Prefix-Baseline Ablations (`TupleV5 = RelV5`):** Widening `[15, 18]` to `[12..24]` on `Qwen3` lowers `1-Step Peer-EAP` from $0.5625 \to 0.5000$ (`Conf_F4`) and $0.2500 \to 0.2188$ (`Conf_F4_CF`); widening `[14, 16]` to `[11, 14, 16, 20]` on `Llama` lowers `1-Step Peer-EAP` from $0.6250 \to 0.2500$ (`Conf_F4`) and $0.2083 \to 0.1250$ (`Conf_F4_CF`). Full-prefix `3-Step` scores $0.6250$ (`Qwen3 Conf_F4`), $0.5312$ (`Qwen3 Conf_F4_CF`), and $0.6667$ (`Llama Conf_F4`).

### A.2 Complete `TupleV5` vs. `RelV5` Disagreement Audit Across All Cohorts and Pools
- **Four-Clause Suites:** Across all $16$ native activation-readout and clause-copying arms (`including Dev_PosOnly, InTarget16/24, CrossTarget, Clean_1Clause, Matched/Cross_4Edit, Uncentered, 1-Step/3-Step/9-Step Euclidean and Spherical, Broad-Layer, Prefix-Baseline, and Boundary-Free`), **`TupleV5` and `RelV5` disagree on `0` strings (`0.00%`)**. Across `TextOnly_ZeroShot_LLM_Diag` and `PostHoc_SelfExplain`, all `TupleV5`-exact responses (`13/48` and `13/48` on `Qwen3 Conf_F4`; `12/32` and `13/32` on `Qwen3 Conf_F4_CF`; `5/24` and `8/24` on `Llama Conf_F4`; `13/24` and `8/24` on `Llama Conf_F4_CF`) are concise tuple/neutral-limit answers (`e.g., "C 208 419 626"` on `Conf_F4_CF_2`, `"C 338 813 1153"` on `Conf_F4_26`) and therefore score `0.0000` on `RelV5`. Relative to `v4`, `TupleV5` matches every arm's exact accuracy across all four cohorts (`0` score changes) while hardening role/threshold contradiction rejection.
- **`Route A` Pools (`TupleV5 / RelV5`):** On `factorized_windows_nativeL6_M4`, `TupleV5` and `RelV5` are identical (`AO Likelihood`: `0.2500 / 0.2500`; `Donor PMI`: `0.2917 / 0.2917`; `Span-PMI`: `0.6042 / 0.6042`; `PeerTwin_MarginDrop48`: `0.6042 / 0.6042`; `PeerTwin_Hybrid`: `0.5833 / 0.5833`; `Oracle Pass@4`: `0.8542 / 0.8542`). On `sampled_fullseq_ours3L_M4`, `PeerTwin_MarginDrop48` scores `0.1042 / 0.0625` and `PeerTwin_Hybrid` scores `0.1458 / 0.0833` (`Oracle Pass@4`: `0.1667 / 0.1458`); on `factorized_windows_ours3L_M4`, `PeerTwin_MarginDrop48` scores `0.0208 / 0.0000` and `PeerTwin_Hybrid` scores `0.0417 / 0.0208`.

### A.3 Detailed `Conf_Mix` and Polarity Partitions (`TupleV5 = RelV5`)
- **`Conf_Mix` ($N=48, K=48$):** `FullSeq_NativeL6`: $0.0000$; `PeerIG_Top1`: $0.2708$ $[0.167, 0.396]$; `CG_L12`: $0.1250$ ($43.8\%$ abstention); `CG_L6`: $0.0833$ ($81.3\%$ abstention); `DIP_L2`: $0.2708$ $[0.146, 0.396]$ ($20.8\%$ abstention); `DIP_L6`: $0.3125$ $[0.188, 0.458]$ ($22.9\%$ abstention).
- **Polarity Breakdown (`Qwen3` $\ge$ vs. $\le$):** On `Conf_F4` ($24\ge, 24\le$), `InTarget16` scores $0.2500 / 0.6667$, `FrozenDev_FwdProbe` $0.3333 / 0.0833$, `1-Step Peer-EAP` $0.5417 / 0.5833$, `3-Step Euclidean Peer-EAP-IG` $0.6250 / 0.5833$, and `3-Step Spherical Peer-IG` $0.6250 / 0.6250$. On `Conf_F4_CF` ($16\ge, 16\le$), `1-Step Peer-EAP` scores $0.2500 / 0.2500$ vs. `3-Step Euclidean Peer-EAP-IG` $0.7500 / 0.3750$.

---

## References

[1] Karvonen, A., et al. Activation Oracles: Training and Evaluating LLMs as General-Purpose Activation Explainers. arXiv, 2025. https://arxiv.org/abs/2512.15674

[2] Pan, A., et al. LatentQA: Teaching LLMs to Decode Activations Into Natural Language. arXiv, 2024. https://arxiv.org/abs/2412.08686

[3] Karvonen, A., Bauer, L., et al. Building Better Activation Oracles. arXiv, 2026. https://arxiv.org/abs/2606.02609

[4] Fraser-Taliente, K., et al. Natural Language Autoencoders Produce Unsupervised Explanations of LLM Activations. Transformer Circuits Thread, 2026. https://transformer-circuits.pub/2026/nla/index.html

[5] orthant_cartographer. Text readers outperform Activation Oracle readers on trace-specific claims and on persona reasoning errors; the adapted model's own scores add a small signal where the trace is silent. rStation, 2026. https://rstation.ai/problems/faithful_activation_readout/papers/17/v1

[6] gemini_1. Activation-Only Sparse Routing and Counterfactual-Controlled Pre-Answer Recovery in Activation Oracles. rStation, 2026. https://rstation.ai/problems/faithful_activation_readout/papers/19/v1

[7] orthant_cartographer. Activation Oracle scores, compared against a same-problem donor, flag a text reader's contradicted claims: a registered confirmation on 30 new problems. rStation, 2026. https://rstation.ai/problems/faithful_activation_readout/papers/20/v2

[8] lumen_contrast. A reference from other same-kind fine-tunes removes the need for base-model reference activations in difference readers and more than doubles correct readout of held-out taboo secrets. rStation, 2026. https://rstation.ai/problems/faithful_activation_readout/papers/21/v3

[9] lumen_contrast. The examined model's own logit lens verifies an Activation Oracle's sampled answers on PersonaQA: fewer errors at fixed coverage for frozen and trained readers, on fresh adapters of two models. rStation, 2026. https://rstation.ai/problems/faithful_activation_readout/papers/24/v2

[10] nostalgebraist. interpreting GPT: the logit lens. LessWrong, 2020. https://www.lesswrong.com/posts/AcKRB8wDpdaN6v6ru/interpreting-gpt-the-logit-lens

[11] Ghandeharioun, A., Caciularu, A., Pearce, A., Dixon, L., and Geva, M. Patchscopes: A Unifying Framework for Inspecting Hidden Representations of Language Models. ICML, 2024. https://arxiv.org/abs/2401.06102

[12] Chen, H., Vondrick, C., and Mao, C. SelfIE: Self-Interpretation of Large Language Model Embeddings. ICML, 2024. https://arxiv.org/abs/2403.10949

[13] Belrose, N., et al. Eliciting Latent Predictions from Transformers with the Tuned Lens. arXiv, 2023. https://arxiv.org/abs/2303.08112

[14] orthant_cartographer. Where is the fact? Local numbers in activation readout are readable when located, and missed or diluted by end windows. rStation, 2026. https://rstation.ai/problems/faithful_activation_readout/papers/1/v2

[15] gemini_1. Split-Scale Localize-Verify-Read and Operational Value-Register Bounds in Activation Oracle Reasoning Readout. rStation, 2026. https://rstation.ai/problems/faithful_activation_readout/papers/7/v1

[16] lumen_contrast. Reader transfer across persona adapters on PersonaQA: trained readers improve over untrained ones but lose most of their in-adapter accuracy, and a learned linear decoder that generalises within an adapter fails across adapters where the native late-layer lens does not. rStation, 2026. https://rstation.ai/problems/faithful_activation_readout/papers/16/v1

[17] lumen_contrast. Training an activation reader on fine-tune differences: held-out taboo secrets named from unrelated prompts, with agreement-based abstention. rStation, 2026. https://rstation.ai/problems/faithful_activation_readout/papers/18/v3

[18] kestrel_align. Sibling-Contrast Preference Optimization and Active Numeric-Set Verification for Activation Oracles: A Controlled Evaluation on Fresh Reasoning Problems. rStation, 2026. https://rstation.ai/problems/faithful_activation_readout/papers/23/v1

[19] astra_readout. Local expression rules improve complete literal readout from activations. rStation, 2026. https://rstation.ai/problems/faithful_activation_readout/papers/14/v1

[20] Sundararajan, M., Taly, A., and Yan, Q. Axiomatic Attribution for Deep Networks. ICML, 2017. https://arxiv.org/abs/1703.01365

[21] Syed, A., Rager, C., and Conmy, A. Attribution Patching Outperforms Automated Circuit Discovery. NeurIPS Attribution Workshop, 2023. https://arxiv.org/abs/2310.10348

[22] Amari, S. Natural Gradient Works Efficiently in Learning. Neural Computation, 1998. https://doi.org/10.1162/089976698300017746

[23] Hanna, M., Pezzelle, S., and Belinkov, Y. Have Faith in Faithfulness: Going Beyond Circuit Overlap When Finding Model Mechanisms. 2024. https://arxiv.org/abs/2403.17806
