## 1. Problem Formulation and Contributions

Diagnosing a language model's intermediate reasoning with a frozen natural-language Activation Oracle (AO) [1], [2], [3] requires answering **entity-specific trajectory questions**—such as *"What value did the model calculate for Depot West?"*—as well as auditing whole chain-of-thought (`CoT`) prefixes when multiple intermediate quantities are computed and a queried entity may not appear at all. Recent investigations on `faithful_activation_readout` expose an operational **window-width incompatibility**:

- **Wide windows miss early steps and dilute single-layer reporting of located numbers:** Extending Karvonen et al. [1, Appendix C.4], `orthant_cartographer` [4] showed via span-aligned numeric edits on `Qwen3-8B` and `Qwen3-1.7B` that intermediate integers are readable from an oracle `5`-token window (`focus-5`), missed outside trailing `50`-token windows (`end-50`), and under-reported on `original` when `focus-5` is concatenated with `end-50` (`union-55`), while target-model attention locates target spans unreliably.
- **Short-template abstention and `50`-token self-verification:** `orthant_cartographer` [5] showed on two-sentence prompts (*"A owns a pet frog. B owns a pet duck."*) that appending *"If this person is not mentioned, answer 'unknown'"* curbs cross-entity binding errors and that active yes/no verification on `50`-token windows separates supported from contradicted claims, while `gemini_1` [6] established Centroid-Reference Decoding (`CRD`) and conditioning-history bounds on passive likelihood ratios.

In multi-step `<think>` derivations ($\approx 294\text{--}576$ tokens), however, **entity headers and arithmetic products are spatially separated during step-by-step calculation**. A `<think>` step typically names the entity (*"Starting with Depot East: it has 7 aisles holding 12 pallets..."*) **$15$ to $65$ tokens before** stating the product (*"7 multiplied by 12 is 84."*). Across $36$ two-branch counterfactual `<think>` traces, a `5`-token window centered on the computed integer covers the distinguishing entity token in only **$5.6\%$ ($2/36$) of Branch-1 traces and $0.0\%$ ($0/36$) of Branch-2 traces** (Figure 1A). Consequently, no single window width $W$ handles localization, entity verification, absence gating, and numeric extraction simultaneously across frozen readers:
1. **Narrow `5`-token windows act operationally as entity-blind value registers unless within-trace score offsets are canceled:** At `W = 5`, the explicit `'unknown'` prompt from [5] yields **$0.0\%$ abstention** and high false binding to another branch's number across all three `Qwen3-8B` readers, while `5`-token yes/no verification stays **near chance ($0.453\text{--}0.539$ Present-vs-Absent ROC-AUC under matched wording)**. Upstream entity-swap minimal pairs ($T_{\mathrm{base}}$ vs. $T_{\mathrm{swap}}$, keeping Branch-2 `W = 5` tokens lexically identical on $23/24$ test pairs) show that `3`-layer readers detect the swap on $87.5\%\text{--}95.8\%$ of paired comparisons, yet on single traces the swap-odd signal is masked by much larger swap-even offsets, collapsing raw orientation-invariant accuracy to $4.2\%\text{--}8.3\%$.
2. **Wide clause windows verify entity scope while interacting differently with single- vs. multi-layer numeric readout:** Peak-bounded preceding `32`-token spans (`Clause-32`) recover $91.7\%$ (`Branch 1`) and $83.3\%$ (`Branch 2`) entity-header coverage, restoring candidate disambiguation and absence gating, whereas decoding integers directly from `W = 32` degrades single-layer (`1L`) `original` readout relative to `W = 5` while performing comparably to `W = 5` on supervised `3`-layer (`3L`) `ours-3layers` when steps remain separated.

Keeping all released AO adapters strictly frozen, **Split-Scale Localize–Verify–Read (`SLVR`)** assigns each sub-decision to its matched spatial scale and establishes four controlled empirical findings:
- **Cue-Prefilled Digit Localization (`W_loc = 5`):** Standard `step-0` patch-KL localization drops to **$8.3\%$** span hit on `ours-3layers` due to generic framing words (*"The"*). Appending `"The model calculated "` lifts `Branch 1` top-$1$ span hit to **$54.2\%\text{--}79.2\%$** and **top-$6$ recall to $89.6\%\text{--}97.9\%$**, and closing the complement (`"...computes the exact intermediate result value as "`) lifts entity-free `W = 5` readout on `ours-3layers` from $12.5\%$ to $87.5\%$.
- **Deconfounded Verification-Width Crossing, `Locator-Only` Baselines, and Independent Target-Blind Rollouts:** Crossing verification question wording (`Q_calc` vs. `Q_ment`) with window width (`W = 5` vs. `Clause-32`) proves that widening verification improves `R = 5` accuracy across **both** phrasings ($+21.88, +27.08, +22.92\text{ pp}$ under `Q_calc`; $+19.79, +14.58, +11.46\text{ pp}$ under `Q_ment`; $+16.67\text{ pp}$ on `Qwen3-1.7B`). Against unverified **`Locator-Only` (`Top-1 S_dig`)**, `SLVR`'s present-query gains on the $24$-problem benchmark ($+7.29, +9.38, +4.17\text{ pp}$) have paired $95\%$ confidence intervals (`CIs`) crossing zero, showing that narrow verification degrades localization while `Clause-32` verification mainly resolves `Branch 2` bleed ($+25.00\text{--}+29.17\text{ pp}$ on `3L` readers) and absent-entity gating. On **$24$ strictly independent final problems** evaluated at target-blind `fixed300` and zero-truncation `complete` `<think>` endpoints, `SLVR` significantly beats `Locator-Only` on `original` `complete` ($+13.54\text{ pp}$) and `ours-3layers` (`+14.58 pp` `fixed300`, `+20.83 pp` `complete`), while concluding summary sentences on `complete` rollouts co-locate entities and numbers within `5` tokens so `3L` `V=5, R=5` catches up ($58.33\%$ vs. $50.00\%$).
- **Deployed `N = 120` Absence-Gated Policy and Two-Tier Zero-Hint Faithfulness:** Calibration-selected wide gating achieves **$50.83\%$, $55.83\%$, and $48.33\%$** end-to-end accuracy across $120$ present+absent queries ($54.2\%\text{--}87.5\%$ absent abstention vs. $0.0\%$ for `Locator-Only` and `W = 5` `'unknown'` prompts). Under a **two-tier zero-hint rubric** (`Scalar Association` vs. `Full Claim Support` validating entity, attribute/unit, and value together, audited blind on $150$ unique outputs at $97.33\%$ pair agreement), `Cue-NMS Union-60` uniquely preserves unit-bearing claims ($22.2\%$ and $16.7\%$ zero-contradiction full-claim support on `original`/`ours-3layers`), whereas `Blind-SLVR` omits units (**$0.0\%$ full-claim joint support**) while leading early-branch scalar recovery ($31.9\%\text{--}38.9\%$) and rescuing `GRPO-500` scalar associations ($20.8\%$ vs. $2.8\%$).
- **Operational Offset Cancellation:** Single-trace controls show that both null-injection subtraction and same-trace `Within-2Peak` subtraction reduce swap-even offset magnitudes without requiring counterfactual twin traces, lifting `Clause-32` single-trace orientation-invariant accuracy to $75.0\%\text{--}91.7\%$.

---

## 2. Related Work and Specific Distinctions

**Position-Localized Activation Decoding.** Patchscopes [7] and SelfIE [8] inspect hidden representations at pre-selected positions, whereas LatentQA [2], Karvonen et al. [1], and Bauer et al. [3] train continuous AO verbalizers over single- or multi-layer residual windows. Karvonen et al. [1, Appendix C.4] noted that a single informative token position can outperform full-sequence input on Taboo, and `orthant_cartographer` [4] demonstrated via counterfactual numeric edits on `Qwen3-8B` and `Qwen3-1.7B` that `focus-5` reads recover local CoT integers whereas `end-50` and `union-55` miss or under-report them on `original`, while target-model digit attention locates target positions unreliably. `SLVR` addresses multi-step `<think>` localization and disambiguation by pairing **AO cue-prefilled digit sensitivity** with separate verification (`32`/`50` slots) and extraction (`5` slots) scales.

**Abstention, Calibration, Entity Binding, and Inverter Training.** Federico Torrielli, Peter Schneider-Kamp, and Lukas Galke Poech [9] evaluated five AO confidence methods on secret-word tasks, showing that closed-set candidate likelihoods and $20$-sample semantic agreement calibrate confidence whereas verbalized numeric scores fail, and showing in Appendix N of [9] that deviating from the trained `coeff = 1` layer-$1$ injection norm degrades readout sharply. Keeping `coeff = 1` and `CRD` ($\beta = 0.25$) from `gemini_1` [6], we study open-ended `<think>` localization without closed numeric candidate sets. `orthant_cartographer` [5] adapted `P(True)`-style verification [10] to `50`-token AO reads and showed that an `'unknown'` prompt prevents cross-person binding on short two-sentence inputs (version 2). Complementary to causal minimal-pair localization [11] and decomposed prompting [12], `astra_readout` [13] showed on synthetic `40`-token record accounts that joining four primitive text queries aids `original` reference resolution ($+29.10\text{ pp}$) but degrades `ours-3layers` and `GRPO-500` via noisy intermediate text labels, while `lumen_contrast` [14] showed that `5`-position fine-tuning target specificity governs contradiction rates. Two external works further contextualize our diagnostics: **Jiahai Feng and Jacob Steinhardt [15]** established via causal intervention that language models bind entities to attributes through latent binding-ID subspaces in residual streams, motivating our distinction between paired counterfactual discrimination and single-trace AO score offsets (§4.5); and **Zhao et al. [16]** introduced Activation Verbalization Preference Optimization (`AVPO`), which separates representation reconstruction from downstream question answering and **trains** the inverter to reduce hallucinations. Unlike `AVPO` [16], we evaluate **strictly frozen** released AO checkpoints, isolating inference-time scale trade-offs and unit-level claim limits.

---

## 3. Split-Scale Localize–Verify–Read (`SLVR`)

### 3.1 Controlled Benchmark Provenance, Reused-24 Truncation Audit, and Independent-24 Final Rollouts

Each controlled problem $i \in \{1, \ldots, 36\}$ ($12$ calibration, $24$ strictly disjoint test across $6$ domains: logistics, physics, chemistry, finance, computing, biology) specifies two named subsystems $B_1, B_2$ with factor inputs whose products $x_{i,1}, x_{i,2}$ never appear in the prompt, followed by three qualitative context sentences. Following [4], we construct **span-aligned counterfactual triples** $(T_i^{\mathrm{base}}, T_i^{\mathrm{cf1}}, T_i^{\mathrm{cf2}})$ ($x_{i,1} \mapsto x'_{i,1}$ in $\mathrm{cf1}$; $x_{i,2} \mapsto x'_{i,2}$ in $\mathrm{cf2}$) of identical token length $L_i$ ($\mathrm{CoT}$ starting at $s_i$, mean $L_i = 294.0$), where `Branch 1` lies outside `end-50` ($\ge 52$ tokens from $L_i$, mean $100.7$) and `Branch 2` lies inside `end-50` ($10\text{--}49$ tokens from $L_i$, mean $28.4$). A third **absent-entity query** $B_{\mathrm{abs}}$ substitutes an unmentioned proper noun (`Omega`, e.g., *"pallet count in Depot Omega"*), matching present queries in syntax and token count.

**Explicit Provenance and Endpoint-Selection Audit (`records/counterfactual_36_traces.json`, `records/exp10b_revision_checks_records.json`):** On the $24$ controlled test problems ($12$ calibration in parentheses), **$10/24$ ($8/12$ cal)** are raw `Qwen3-8B` `<think>` prefixes truncated with zero inserted prose; **$8/24$ ($3/12$ cal)** append a post-Branch-2 domain sentence before truncating so `Branch 2` lands $24\text{--}42$ tokens from $L_i$; **$2/24$ ($0/12$ cal)** use a transition bridge; and **$6/24$ ($1/12$ cal)** use a two-step fallback template when $240$-token generation had not yet reached $x_{i,2}$. An intermediate check on the same $24$ test problem identities under a $420$-token budget (`reused24_420budget_prefixes`) eliminated all inserted prose ($24/24$ verbatim prefixes), **but truncated $15/24$ rollouts before later summary repetitions (`sliced_before_summary_repeat`, yielding $213\text{--}418$ CoT tokens, mean $334.5$, vs. $9/24$ unsliced)** and retained an older narrow/wide question-wording difference. Because truncating before summaries directly alters whether `end-50` sees `Branch 1` and reuses the $24$ test problem setups, we construct **`24` strictly independent held-out final problems** ($4$ per domain; disjoint entities, setups, and products; `records/exp10c_v3_clean_controls_records.json`) and evaluate natural `Qwen3-8B` `<think>` generations at **two target-blind endpoints**:
1. **`Fixed-300` (`cot_ids[:300]`):** Fixed at the exact first $300$ CoT tokens ($298$ decoded CoT tokens) with **zero inspection of target number positions** ($24/24$ evaluable; $21/24$ place `Branch 1`'s last occurrence outside `end-50`).
2. **`Complete <think>` (`max_new_tokens = 640`, zero truncation):** The complete generated `<think>` trace ($24/24$ evaluable; $15/24$ close `</think>` naturally; $434\text{--}638$ CoT tokens, mean $576.2$; $23/24$ repeat `Branch 1` in later summaries or self-checks, yet $20/24$ still end $\ge 50$ tokens after `Branch 1`'s final mention because qualitative discussion follows).

### 3.2 Multi-Scale Inference Pipeline, Inference Costs, and Worked Example

Let $h_t^{(\ell)} \in \mathbb{R}^d$ denote the target residual at token $t \in [s_i, L_i)$ and layer $\ell \in \mathcal{L}$ ($\{18\}$ for `8B original`, $\{9, 18, 27\}$ for `8B ours-3layers` and `GRPO-500`, $\{14\}$ for `1.7B`).

**Stage 1 — Cue-Prefilled Digit Localization (`W_loc = 5`, $1$ batched prefill pass of size $M + 1$):** Partition $[s_i, L_i)$ into $M$ strided `5`-token windows $W_1, \ldots, W_M$ (stride $3$) with trace-internal spherical centroid $\bar{u}^{(5)}$. Appending numeric cue prefill $c_{\mathrm{num}}$ (`"The model calculated "` for entity-prompted queries; `"In this step, the model computes the exact intermediate result value as "` when no entity is supplied), one batched single-step pass yields $\log p(\cdot \mid W_m, q, c_{\mathrm{num}})$. For ASCII-digit token set $\mathcal{D} \subset \mathcal{V}$, window $m$ is scored by positive digit-subspace information gain:

$$
S_{\mathrm{dig}}(W_m) = \sum_{y \in \mathcal{D}} p(y \mid W_m, q, c_{\mathrm{num}}) \, \max\!\left(0,\; \log \frac{p(y \mid W_m, q, c_{\mathrm{num}})}{p(y \mid \bar{u}^{(5)}, q, c_{\mathrm{num}})}\right),
$$

retaining the top $K = 6$ candidate windows $m_1, \ldots, m_K$ with centers $c_k = W_{m_k}[2]$. **`Locator-Only` (`Top-1 S_dig`)** immediately decodes $W_{m_1}$ (`0` verification passes).

**Stage 2 — Full-Curve NMS Clause Verification (`W_verify = 32`) & Presence Gating (`W_trace = 50`):** Greedy non-maximum suppression (radius $12$ tokens) **over the full $M$-window $S_{\mathrm{dig}}$ sequence** identifies step peaks $p_1 < \cdots < p_R$. *(An earlier runner reconstructed `8B` `S_dig` only at the $6$ retained candidates; exact full-curve NMS updates preceding peak boundaries on $119, 103, 103$ present items across `original`, `ours-3layers`, and `GRPO-500`.)* Candidate $k$'s **peak-bounded clause span** runs from $a_k = \max(\{s_i, c_k - 69\} \cup \{p_r + 3 : p_r \le c_k - 12\})$ to $b_k = \min(L_i, c_k + 3)$, uniformly subsampled to $W_{\mathrm{verify}} = 32$ positions $C_k$ (centroid $\bar{u}^{(32)}$). One `B = 7` teacher-forced pass computes within-candidate contrastive log-likelihoods $S_{\mathrm{within}}(k)$ over queried attribute tokens $d = (y_1, \ldots, y_{|d|})$:

$$
\Delta_{k,t}^{\mathrm{within}} = \log p_t(y_t \mid C_k) - \frac{1}{K}\sum_{j=1}^{K} \log p_t(y_t \mid C_j), \qquad S_{\mathrm{within}}(k) = \frac{1}{|d|}\sum_{t=1}^{|d|} \Delta_{k,t}^{\mathrm{within}} + \frac{1}{2}\min_{t} \Delta_{k,t}^{\mathrm{within}},
$$

and one `B = 7` single-step pass computes **calibrated active yes/no margins** $\tilde{Y}_k = Y_k(C_k) - Y_k(\bar{u}^{(32)})$. To eliminate any question-wording confound between narrow (`W = 5`) and wide (`Clause-32`) verification, our primary evaluation uses **identical wording (`Q_ment`)**—*"Answer Yes or No only. Do these activations mention or calculate the {d}?"*—across all scales, and explicitly crosses `Q_ment` with **`Q_calc`** (*"Answer Yes or No only. Is the model calculating or stating the {d} in these activations?"*) in Table 1B. Under `Q_ment` with full-curve NMS, calibration ($12$ problems) selects pure **`cal_yn` ($\tilde{Y}_k$) unanimously across all three `8B` readers at both `W = 5` and `Clause-32`**, so deployed `8B` candidate verification requires **just $1$ batched `B = 7` single-step yes/no pass** (`+1` single-step `Trace-50` pass when calibration selects `Trace-50` presence gating; `0` extra passes when `Clause-32` gating reuses $\max_k \tilde{Y}_k$).

**Stage 3 — Narrow Local Readout (`W_read = 5`, $1$ `B = 2` `CRD` decode):** If the calibration-selected presence gate passes ($\ge \tau_{\mathrm{abs}}$), `SLVR` decodes the integer from $W_{m_{k^*}}$ (`CRD`, $\beta = 0.25$, $\tau = 0.05$) [6]; otherwise it abstains (`0` decodes).

**Worked Example (`test_log_1`, `base`, `8B ours-3layers`):** On zero-inserted-prose trace `test_log_1` ($L_i = 303$), `Depot East` ($7 \times 12 = 84$, $105$ tokens from $L_i$) lies outside `end-50`, while `Depot West` ($6 \times 15 = 90$, $28$ tokens from $L_i$) lies inside `end-50`. Stage 1 ranks two `Depot West` windows ($S_{\mathrm{dig}} = 2.57, 2.11$) ahead of four `Depot East` windows ($1.94\text{--}1.96$), so `Locator-Only` (`m_1`) outputs wrong-branch `"90"`. At `W = 5`, yes/no margins stay positive on all $6$ windows for `Depot East`, `Depot West`, and absent `Depot Omega`. Expanding to `Clause-32` (`Q_ment`) drives `Depot East` candidate $k = 5$ to $\tilde{Y}_5 = +5.88 > \tau_{\mathrm{abs}} = 1.625$ (wrong-branch windows $-0.88, +0.62$), drives `Depot West` candidate $k = 1$ to $+5.00$ (`Depot East` windows $-3.25\text{--}-2.12$), and keeps all six `Depot Omega` windows below $\tau_{\mathrm{abs}}$ ($\max_k \tilde{Y}_k = -0.75$), abstaining cleanly before Stage 3 decodes `"84"` and `"90"` from $W_{37}$ and $W_{62}$.

---

## 4. Empirical Results

### 4.1 Entity-Header Separation and Stage-1 Cue-Prefilled Localization

![Three-panel summary of multi-step CoT geometry and Stage-1 localization: Panel A shows own-entity coverage rising from 0.0% at W=5 on Branch 2 to 83.3%-97.2% under peak-bounded preceding windows; Panel B shows Cue-Prefilled Digit sensitivity overcoming step-0 preamble blindness on 3-layer readers; Panel C shows end-50 missing Branch 1 and union-55 reducing reporting relative to focus-5 on original.](supplement/figure-1.png)

Figure 1A shows that `W = 5` centered on the target product covers the distinguishing entity token in only **$2/36$ ($5.6\%$) of Branch-1 traces and $0/36$ ($0.0\%$) of Branch-2 traces**, whereas peak-bounded preceding spans reach **$33/36$ ($91.7\%$) and $30/36$ ($83.3\%$)** at $32$ tokens ($91.7\% / 91.7\%$ at $64$ tokens). Across the $24$ controlled test problems ($48$ traces per branch, Figure 1B–C), `end-50` ($0.0\%\text{--}4.2\%$) and `target_attn_digit_top1` ($2.1\%\text{--}4.2\%$ accuracy; $0.0\%$ `Branch 1` span hit) miss `Branch 1`, while `union-55` lowers `original` `Branch 1` reporting from $100.0\%$ (`focus-5`) to $47.9\%$. Standard `step-0` patch-KL achieves $68.8\%$ span hit on `original` (`1L`) but collapses to **$8.3\%$ on `ours-3layers`** ($33.3\%$ on `GRPO-500`) due to generic step-$0$ framing words. Cue prefilling (`"The model calculated "`) lifts `Branch 1` top-$1$ span hit to **$54.2\%$, $62.5\%$, and $79.2\%$** and **top-$6$ recall to $97.9\%$, $89.6\%$, and $95.8\%$** ($95.8\%$, $85.4\%$, and $93.8\%$ on `Branch 2`).

### 4.2 Deconfounded `2 x 2` Scale & Wording Crossings, `Locator-Only` Baselines, and Independent Rollouts

![Three-panel analysis of deconfounded verification-vs-readout scale controls: Panel A compares Locator-Only (Top-1 S_dig), Narrow (V=5,R=5), Wide (V=32,R=32), and Split-Scale SLVR (V=32,R=5) under identical Q_ment wording; Panel B crosses Q_calc vs. Q_ment with W=5 vs. Clause-32 under pure calibrated yes/no; Panel C evaluates 24 strictly independent final problems at Complete and Fixed-300 target-blind endpoints.](supplement/figure-2.png)

| Reader & Evaluation Set (Identical `Q_ment` Throughout; Ungated) | `end-50` (`0+0+1`) | `Rand-6` / `LL-C32` | **`Locator-Only`** (`1+0+1`) | `V=5, R=5` (`1+1+1`) | `V=32, R=32` (`1+1+1`) | **`SLVR` (`V=32, R=5`)** | `SLVR` vs. `V=5, R=5` [$95\%$ CI] | `SLVR` vs. `Loc-Only` [$95\%$ CI] | `SLVR` vs. `V=32, R=32` [$95\%$ CI] |
|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|
| **`1.7B` `LatentQA` (Main $96$)** | — | — | — ($59.4\%$ sp) | $27.08\%$ | $41.67\%$ | **$43.75\%$** | **$+16.67$ [$+4.17, +29.17$]** | — | $+2.08$ [$-7.29, +11.46$] |
| **`8B` `original` (Main $96$)** | $52.08\%$ | $34.2 / 43.8\%$ | $51.04\%$ | $38.54\%$ | $37.50\%$ | **$58.33\%$** | **$+19.79$ [$+4.17, +33.33$]** | $+7.29$ [$-7.29, +21.88$] | **$+20.83$ [$+3.12, +38.54$]** |
| ↳ `Branch 1` / `Branch 2` ($48$ ea.) | $4.2 / 100\%$ | $43.8 / 43.8\%$ | $60.4 / 41.7\%$ | $52.1 / 25.0\%$ | $41.7 / 33.3\%$ | **$62.5 / 54.2\%$** | **B2: $+29.17$ [$+6.25, +50.0$]** | B2: $+12.50$ [$-10.4, +35.4$] | `V=5, R=32`: $21.88\%$ |
| ↳ **Indep-24 `Fixed-300` ($96$ q)** | $29.17\%$ | $40.1 / 57.3\%$ | $59.38\%$ | $44.79\%$ | $13.54\%$ | **$69.79\%$** | **$+25.00$ [$+13.54, +37.50$]** | $+10.42$ [$-1.04, +20.83$] | **$+56.25$ [$+42.71, +68.75$]** |
| ↳ **Indep-24 `Complete` ($96$ q)** | $13.54\%$ | $47.0 / 55.2\%$ | $48.96\%$ | $50.00\%$ | $20.83\%$ | **$62.50\%$** | **$+12.50$ [$+1.04, +23.96$]** | **$+13.54$ [$+1.04, +26.04$]** | **$+41.67$ [$+26.04, +57.29$]** |
| **`8B` `ours-3layers` (Main $96$)** | $48.96\%$ | $31.2 / 29.2\%$ | $46.88\%$ | $41.67\%$ | **$60.42\%$** | $56.25\%$ | **$+14.58$ [$+2.08, +27.08$]** | $+9.38$ [$-5.21, +23.96$] | $-4.17$ [$-12.50, +4.17$] |
| ↳ `Branch 1` / `Branch 2` ($48$ ea.) | $2.1 / 95.8\%$ | $33.3 / 25.0\%$ | $62.5 / 31.2\%$ | $50.0 / 33.3\%$ | $60.4 / 60.4\%$ | $56.2 / 56.2\%$ | **B2: $+22.92$ [$+2.08, +41.7$]** | **B2: $+25.00$ [$+6.25, +43.8$]** | `V=5, R=32`: $41.67\%$ |
| ↳ **Indep-24 `Fixed-300` ($96$ q)** | $30.21\%$ | $31.1 / 37.5\%$ | $34.38\%$ | $44.79\%$ | **$54.17\%$** | $48.96\%$ | $+4.17$ [$-5.21, +14.58$] | **$+14.58$ [$+4.17, +25.00$]** | $-5.21$ [$-16.67, +6.25$] |
| ↳ **Indep-24 `Complete` ($96$ q)** | $12.50\%$ | $37.0 / 37.5\%$ | $29.17\%$ | **$58.33\%$** | $47.92\%$ | $50.00\%$ | $-8.33$ [$-16.67, 0.00$] | **$+20.83$ [$+6.25, +34.38$]** | $+2.08$ [$-6.25, +10.42$] |
| **`8B` `GRPO-500` (Main $96$)** | $48.96\%$ | $30.2 / 43.8\%$ | $48.96\%$ | $41.67\%$ | **$53.12\%$** | **$53.12\%$** | **$+11.46$ [$0.00, +23.96$]** | $+4.17$ [$-6.25, +13.54$] | $0.00$ [$-10.42, +11.46$] |
| ↳ `Branch 1` / `Branch 2` ($48$ ea.) | $0.0 / 97.9\%$ | $56.2 / 31.2\%$ | $68.8 / 29.2\%$ | $54.2 / 29.2\%$ | $54.2 / 52.1\%$ | $47.9 / 58.3\%$ | **B2: $+29.17$ [$+10.4, +50.0$]** | **B2: $+29.17$ [$+8.33, +50.0$]** | `V=5, R=32`: $39.58\%$ |
| ↳ **Indep-24 `Fixed-300` ($96$ q)** | $27.08\%$ | $30.7 / 43.8\%$ | $40.62\%$ | $43.75\%$ | $36.46\%$ | **$45.83\%$** | $+2.08$ [$-10.42, +14.58$] | $+5.21$ [$-7.29, +17.71$] | $+9.38$ [$-3.12, +21.88$] |
| ↳ **Indep-24 `Complete` ($96$ q)** | $13.54\%$ | $40.3 / 47.9\%$ | $46.88\%$ | $50.00\%$ | $35.42\%$ | **$52.08\%$** | $+2.08$ [$-10.42, +14.58$] | $+5.21$ [$-7.29, +16.67$] | **$+16.67$ [$+5.21, +28.12$]** |

*Table 1A: Present-query counterfactual own-number tracking under identical `Q_ment` yes/no wording and full-curve `S_dig` NMS on the $24$ controlled test problems (`Main 96`, `no_activation` $= 0.00\%$, `oracle_focus5_crd` $= 100.0\% / 93.75\% / 91.67\%$) and on $24$ strictly independent final problems at target-blind `Fixed-300` and zero-truncation `Complete <think>` endpoints. Header triples `(s1+s2+s3)` denote per-query `(Stage-1 prefill batches + Stage-2 verification batches + Stage-3 CRD decodes)`; `LL-C32` (`S_within` likelihood-only) uses `(1+1+1)`. On the reused-24 pre-summary-sliced prefixes (`15/24` `sliced_before_summary_repeat`), `SLVR` scored $66.67\%, 60.42\%, 59.38\%$ vs. `end-50` $27.08\%, 28.12\%, 23.96\%$.*

| `8B` Reader (`R = 5`, Pooled $96$) | `W=5 + Q_calc` | `W=5 + Q_ment` | `Clause-32 + Q_calc` | `Clause-32 + Q_ment` | Width Gain $\mid$ `Q_calc` [$95\%$ CI] | Width Gain $\mid$ `Q_ment` [$95\%$ CI] | Matched `Q_calc` Cal-Sel (`V=5` $\to$ `SLVR`) |
|---|---:|---:|---:|---:|---:|---:|---:|
| **`original` (`1L`)** | $33.33\%$ | $38.54\%$ | $55.21\%$ | **$58.33\%$** | **$+21.88$ [$+9.38, +34.38$]** | **$+19.79$ [$+4.17, +33.33$]** | $35.42\% \to \mathbf{55.21\%}$ (**$+19.79$**) |
| **`ours-3layers` (`3L`)** | $38.54\%$ | $41.67\%$ | **$65.62\%$** | $56.25\%$ | **$+27.08$ [$+15.62, +38.54$]** | **$+14.58$ [$+2.08, +27.08$]** | $48.96\% \to \mathbf{65.62\%}$ (**$+16.67$**) |
| **`GRPO-500` (`3L`)** | $34.38\%$ | $41.67\%$ | **$57.29\%$** | $53.12\%$ | **$+22.92$ [$+11.46, +34.38$]** | **$+11.46$ [$0.00, +23.96$]** | $33.33\% \to \mathbf{57.29\%}$ (**$+23.96$**) |

*Table 1B: `2 x 2` Question-Wording (`Q_calc`: "calculating or stating" vs. `Q_ment`: "mention or calculate") $\times$ Verification-Width (`W = 5` vs. `Clause-32`) crossing on the exact same $6$ candidates (`records/exp10c_v3_clean_controls_records.json`).*

Tables 1A–1B and Figure 2 establish three deconfounded operational conclusions:
1. **Verification-width gains hold under both question phrasings (`Table 1B`, Figure 2B):** Crossing `Q_calc` and `Q_ment` with `W = 5` and `Clause-32` refutes the hypothesis that `"mention"` artificially inflated wide verification. Holding wording fixed, widening from `W = 5` to `Clause-32` raises pure `cal_yn` accuracy by **$+21.88$, $+27.08$, and $+22.92\text{ pp}$ under `Q_calc`** and **$+19.79$, $+14.58$, and $+11.46\text{ pp}$ under `Q_ment`** (wording differences at `Clause-32` have $95\%$ CIs $[-4.17, +10.42]$, $[-18.75, 0.00]$, and $[-16.67, +8.33]$—in fact, `Q_calc` *outperforms* `Q_ment` at `Clause-32` on both `3L` readers because requiring active calculation rejects passive entity mentions!).
2. **`SLVR` vs. `Locator-Only`, `Random-of-6`, and `Likelihood-Only`:** On `Main 96`, `Locator-Only` (`Top-1 S_dig`, $1$ prefill batch + $1$ decode) reaches $51.04\%, 46.88\%, 48.96\%$—well above `Random-of-6` ($34.20\%, 31.25\%, 30.21\%$) and `V=5, R=5` ($38.54\%, 41.67\%, 41.67\%$, because entity-blind `W = 5` verification actively overrides correct top-$1$ windows!). Consequently, `SLVR`'s pooled gains over `Locator-Only` on `Main 96` ($+7.29, +9.38, +4.17\text{ pp}$) have **paired $95\%$ CIs crossing zero** ($[-7.29, +21.88]$, $[-5.21, +23.96]$, $[-6.25, +13.54]$). Where `Clause-32` verification (`+1` `B = 7` single-step pass) provides statistically reliable gains over `Locator-Only` is: (i) **`Branch 2` disambiguation** on `3L` readers (**$+25.00\text{ pp}$** [$+6.25, +43.75$] on `ours-3layers`, cutting bleed $47.9\% \to 10.4\%$; **$+29.17\text{ pp}$** [$+8.33, +50.00$] on `GRPO-500`, cutting bleed $45.8\% \to 2.1\%$), (ii) beating passive `Likelihood-Only` (`LL-C32`: **$+27.08\text{ pp}$** [$+15.62, +38.54$] on `ours-3layers`), (iii) **natural multi-hundred-token rollouts** (`Indep-24 Complete`: **$+13.54\text{ pp}$** [$+1.04, +26.04$] on `original` and **$+20.83\text{ pp}$** [$+6.25, +34.38$] on `ours-3layers`; `Fixed-300`: **$+14.58\text{ pp}$** [$+4.17, +25.00$] on `ours-3layers`), and (iv) **absent-entity gating** (§4.3).
3. **Target-blind `Fixed-300` vs. zero-truncation `Complete <think>` on $24$ independent problems (Figure 2C):** Removing target-aware slicing reveals two instructive phenomena. First, `end-50` does *not* recover when later summaries are retained—on `Complete <think>`, post-summary qualitative reasoning pushes **both** branches past the final $50$ tokens on most traces, dropping `end-50` to **$13.54\%, 12.50\%, 13.54\%$** (`SLVR` gains **$+48.96, +37.50, +38.54\text{ pp}$**). Second, readout-scale sensitivity follows architecture and local step geometry: `1L` `original` strongly requires narrow `R = 5` readout on both `Fixed-300` ($69.79\%$ vs. $13.54\%$, **$+56.25\text{ pp}$**) and `Complete` ($62.50\%$ vs. $20.83\%$, **$+41.67\text{ pp}$**), and `GRPO-500` also favors `SLVR` over `V=32, R=32` on `Complete` (**$+16.67\text{ pp}$** [$+5.21, +28.12$]). Meanwhile, on `3L` `ours-3layers` `Complete`—where $23/24$ rollouts restate `Branch 1` in a compact concluding summary—`V=5, R=5` rises to **$58.33\%$** vs. $50.00\%$ for `SLVR` ($-8.33\text{ pp}$ [$-16.67, 0.00$]), because summary clauses (*"Orion holds 128 tires and Vega holds 133 tires"*) place each entity **inside `5` tokens of its number** while packing *both* entities into the same `32`-token window!

### 4.3 Absence Verification and Complete Calibration-Selected Deployed Policy (`N = 120`)

![Three-panel evaluation of absent-entity false binding, verification ROC-AUCs, and the complete N=120 deployed policy: Panel A compares prompted 'unknown' abstention at W=5 and W=32 against calibrated active gates; Panel B plots Present-vs-Absent ROC-AUCs under matched Q_ment wording across Null, Donor, W=5, Clause-32, and Trace-50; Panel C decomposes all 120 present+absent test queries.](supplement/figure-3.png)

| Reader & Calibration-Selected Gate | Present Ret.-Correct ($96$) | Present Ret.-Wrong ($96$) | Present Abstained ($96$) | Absent Abstained ($24$) | Absent False-Bind ($24$) | **Overall `120`-Query E2E Acc.** | E2E vs. Ungated `SLVR` |
|---|---:|---:|---:|---:|---:|---:|---:|
| **`1.7B` `LatentQA`**: `Trace-50` ($\tau=0.62$) | $29.2\%$ [$17.7, 41.7$] | $32.3\%$ [$21.9, 43.8$] | $38.5\%$ [$25.0, 52.1$] | **$75.0\%$ [$58.3, 91.7$]** | $12.5\%$ [$0.0, 29.2$] | **$38.33\%$ [$30.00, 47.50$]** | $+3.33$ [$-5.00, +10.83$] |
| **`8B` `original`**: `Trace-50` ($\tau=0.50$) | $44.8\%$ [$31.2, 58.3$] | $22.9\%$ [$12.5, 35.4$] | $32.3\%$ [$18.8, 45.8$] | **$75.0\%$ [$58.3, 91.7$]** | $16.7\%$ [$4.2, 33.3$] | **$50.83\%$ [$40.83, 60.83$]** | $+4.17$ [$-2.50, +10.83$] |
| **`8B` `ours-3layers`**: `Clause-32` ($\tau=1.62$) | $47.9\%$ [$35.4, 60.4$] | $26.0\%$ [$17.7, 34.4$] | $26.0\%$ [$14.6, 38.5$] | **$87.5\%$ [$75.0, 100.0$]** | $12.5\%$ [$0.0, 25.0$] | **$55.83\%$ [$45.83, 66.67$]** | **$+10.83$ [$+5.83, +15.00$]** |
| **`8B` `GRPO-500`**: `Clause-32` ($\tau=1.25$) | $46.9\%$ [$35.4, 58.3$] | $33.3\%$ [$20.8, 46.9$] | $19.8\%$ [$10.4, 29.2$] | $54.2\%$ [$33.3, 75.0$] | $33.3\%$ [$16.7, 54.2$] | **$48.33\%$ [$38.33, 58.33$]** | **$+5.83$ [$0.00, +11.67$]** |
| ↳ `GRPO-500` w/ `Trace-50` ($\tau=0.12$) | $45.8\%$ [$32.3, 59.4$] | $35.4\%$ [$25.0, 46.9$] | $18.8\%$ [$8.3, 33.3$] | **$66.7\%$ [$50.0, 83.3$]** | $29.2\%$ [$12.5, 45.8$] | **$50.00\%$ [$39.17, 60.83$]** | **$+7.50$ [$+1.67, +12.50$]** |

*Table 2: Complete calibration-selected deployed policy over all $120$ test queries ($96$ present + $24$ absent `Omega` queries; `Locator-Only` abstains on $0.0\%$ of absent queries, yielding $40.83\%, 37.50\%, 39.17\%$ ungated `120`-query accuracy—trailing gated `SLVR` by $+10.00, +18.33, +9.17\text{ pp}$).*

Prompt-only `'unknown'` instructions abstain on **$0.0\%$** of absent queries at `W = 5` ($70.8\%\text{--}79.2\%$ false binding) and still abstain on **$0.0\%, 0.0\%, 16.7\%$** at `Clause-32` ($45.8\%, 29.2\%, 41.7\%$ false binding; Figure 3A). Under identical `Q_ment` wording, `W = 5` active yes/no verification remains at chance (**$0.533$** [$0.369, 0.692$] on `1.7B`; **$0.505$** [$0.352, 0.654$], **$0.539$** [$0.395, 0.677$], and **$0.453$** [$0.300, 0.610$] across `8B` `original`, `ours-3layers`, `GRPO-500`, matching `Null` $0.410\text{--}0.556$ and `Donor` $0.474\text{--}0.565$). Widening verification lifts Present-vs-Absent ROC-AUC to **$0.728, 0.793, 0.846, 0.758$** under `Clause-32` and **$0.750, 0.780, 0.882, 0.814$** under `Trace-50` (Figure 3B). While gating incurs $19.8\%\text{--}38.5\%$ present abstention and $22.9\%\text{--}33.3\%$ retained-wrong answers, it consistently beats both ungated `SLVR` and ungated `Locator-Only` (`+18.33 pp` [$+6.67, +30.00$] over `Locator-Only` on `ours-3layers`).

### 4.4 Operational `W = 5` Value-Register Test and Two-Tier Zero-Hint Evaluation (`Scalar` vs. `Full Claim`)

![Three-panel summary of the operational W=5 value-register test, single-trace offset controls, and two-tier zero-hint evaluation: Panel A shows 5-condition prefix grounding on W=5 peaks; Panel B plots single-trace orientation-invariant accuracy across offset controls; Panel C contrasts Scalar Association recovery against unit-validated Full Claim Support.](supplement/figure-4.png)

| Reader & Zero-Hint Regime ($N = 72$ Test Traces) | Scalar `B1 / B2` (`%`) | **Both Scalar** (`0-Contra`) | **Full Claim `B1 / B2`** (`%`) | **Both Full Claim** (`0-Contra`) | Extracted Assocs `ClSup / ScOmit / Unres / Contra` |
|---|---:|---:|---:|---:|---:|
| **`8B` `original`**: `end-50` | $0.0 / 55.6\%$ | $0.0\%$ ($0.0\%$) | $0.0 / 54.2\%$ | $0.0\%$ ($0.0\%$) | $43 / 1 / 31 / 91$ ($80.6\%$ w/ contra) |
| ↳ `Cue-NMS Union-60` (`p1_5+p2_5+end50`) | $22.2 / \mathbf{65.3\%}$ | **$22.2\%$ ($22.2\%$)** | **$22.2 / 65.3\%$** | **$22.2\%$ ($22.2\%$)** | **$68 / 0 / 27 / 67$ ($56.9\%$ w/ contra)** |
| ↳ `Blind-SLVR` (`Top-2` / `Top-4` Dedup) | **$25.0 / 20.8\%$** ; **$34.7 / 20.8\%$** | $5.6\%$ ($5.6\%$) ; $12.5\%$ ($2.8\%$) | $0.0 / 2.8\%$ ; $0.0 / 2.8\%$ | $0.0\%$ ($0.0\%$) ; $0.0\%$ ($0.0\%$) | $2/37/24/92$ ; $4/40/28/145$ |
| **`8B` `ours-3layers`**: `end-50` | $0.0 / 45.8\%$ | $0.0\%$ ($0.0\%$) | $0.0 / 45.8\%$ | $0.0\%$ ($0.0\%$) | $43 / 0 / 33 / 112$ ($93.1\%$ w/ contra) |
| ↳ `Cue-NMS Union-60` (`p1_5+p2_5+end50`) | $20.8 / \mathbf{40.3\%}$ | **$20.8\%$ ($16.7\%$)** | **$20.8 / 40.3\%$** | **$20.8\%$ ($16.7\%$)** | **$55 / 0 / 58 / 84$ ($75.0\%$ w/ contra)** |
| ↳ `Blind-SLVR` (`Top-2` / `Top-4` Dedup) | **$25.0 / 13.9\%$** ; **$31.9 / 19.4\%$** | $9.7\%$ ($9.7\%$) ; $12.5\%$ ($2.8\%$) | $0.0 / 0.0\%$ ; $0.0 / 0.0\%$ | $0.0\%$ ($0.0\%$) ; $0.0\%$ ($0.0\%$) | $0/32/52/63$ ; $0/39/48/130$ |
| **`8B` `GRPO-500`**: `end-50` | $0.0 / 12.5\%$ | $0.0\%$ ($0.0\%$) | $0.0 / 12.5\%$ | $0.0\%$ ($0.0\%$) | $9 / 0 / 16 / 29$ ($31.9\%$ w/ contra) |
| ↳ `Cue-NMS Union-60` (`p1_5+p2_5+end50`) | $2.8 / 8.3\%$ | $2.8\%$ ($2.8\%$) | **$2.8 / 8.3\%$** | **$2.8\%$ ($2.8\%$)** | $8 / 0 / 16 / 23$ ($15.3\%$ w/ contra) |
| ↳ `Blind-SLVR` (`Top-2` / `Top-4` Dedup) | **$34.7 / 16.7\%$** ; **$38.9 / 29.2\%$** | **$13.9\%$ ($13.9\%$)** ; **$20.8\%$ ($0.0\%$)** | $0.0 / 0.0\%$ ; $0.0 / 0.0\%$ | $0.0\%$ ($0.0\%$) ; $0.0\%$ ($0.0\%$) | $0/38/34/74$ ; $0/51/42/118$ |

*Table 3: Two-tier zero-hint evaluation (`Scalar Association` vs. unit/attribute-validated `Full Claim Support`) across $72$ counterfactual test traces (`focus-5`, `Dual-Cache Handoff`, `Null coeff=0`, and `Donor Trace` score $0.0\%$ on `Both Scalar` and `Both Full Claim`; domain gist hit is $94.4\%\text{--}100.0\%$ on `original`/`ours-3layers` and $79.2\%\text{--}100.0\%$ on `GRPO-500`; `records/exp9_zero_hint_records.json`).*

**Prefix Grounding on `W = 5` (Figure 4A):** Under an entity-free question, closing the complement (`"...computes the exact intermediate result value as "`) lifts `ours-3layers` and `GRPO-500` `W = 5` integer emission from $12.5\%$ and $39.6\%$ to **$87.5\%$ and $95.8\%$** ($100.0\%$ on `original`) and boosts entity-free top-$2$ two-branch NMS recall by $2.1\times\text{--}3.2\times$. Supplying the *true* entity ($100.0\%, 60.4\%, 81.2\%$), *swapped wrong-branch* entity ($100.0\%, 58.3\%, 68.8\%$), or *absent* `System Omega` ($100.0\%, 64.6\%, 81.2\%$) yields nearly indistinguishable `W = 5` integer emission.

**Two-Tier Unit/Attribute Validation & Released `150`-Item Blinded Audit (`records/blind_audit_zero_hint_judgments.json`):** Purely matching `(Entity, Scalar)` overstates claim-level faithfulness whenever a reader attaches wrong physical units or attributes: on `test_phy_2` (where Pump One and Pump Two deliver $135$ and $126$ **joules** at $15$ and $18$ watts), `original` `union60` emits *"135 W and ... 126 W"* (`base`) and *"135 horsepower and ... 391 horsepower"* (`cf2`), while `ours-3layers` emits *"Pump One Power (284 W), Pump Two Power (126 W)"* (`cf1`). Our revised scorer (`score_zero_hint_v3`, verified by unit tests in `replay.py`) enforces per-problem attribute/unit schemas (`PROBLEM_ATTR_SCHEMA`), word-bounded entity tags (`Spring K11` $\ne$ `Spring K1`), and duplicate-entity contradiction rejection across **two explicit tiers**:
1. **`Scalar Association` (`ScOmit + ClSup`):** Credits `(Entity, Value)` iff entity and ground-truth integer match, no counterfactual/conflicting value is assigned, **and no incompatible unit or attribute** (`W`, `horsepower`, or `Power` for joules; `N/m` for newtons; unscaled `USD`/dollar sign for thousand dollars) is attached.
2. **`Full Claim Support` (`ClSup`):** Additionally requires an **explicit source-consistent attribute or dimensional unit** (`pallets`, `meters`, `joules`, `newtons`, `watts`, `grams`, `mmol`, `J/K`, `dollars`, `thousand dollars`, `minutes`), classifying bare `"East: 84"` pairs as `scalar_only_attr_omitted` (`ScOmit`).

To validate semantic accuracy transparently, we audited **$150$ unique SHA256-anonymized test outputs** (a complete census of all $84$ unique outputs where scalar-only v2 and unit-aware v3 differed, plus $66$ uniform random unique outputs) **blind to reader checkpoint, method arm, and parser scores**. Exact human-vs-parser agreement is **$146/150 = 97.33\%$ on scalar pairs, $146/150 = 97.33\%$ on full-claim pairs, and $150/150 = 100.00\%$ on contradiction presence**; all $4$ differences involve *"Reactor A / B"* abbreviated for *"Reactor Alpha / Beta"* on `test_chm_3`, where our strict parser conservatively assigns `Unresolved`.

**Architectural Trade-Off Between `Cue-NMS Union-60` and `Blind-SLVR` (Table 3, Figure 4C):**
- Without Stage-1 `S_dig` NMS peaks, `end-50` and `focus-5` recover **$0.0\%$** of early-branch `(E1, V1)` scalars or claims across all readers.
- **`Cue-NMS Union-60` uniquely supports full unit-bearing claims:** Concatenating Stage-1's top-$2$ `5`-token peaks with `end-50` in a single `60`-slot pass achieves **$22.2\%$ and $20.8\%$ `Both Full Claim`** (**$22.2\%$** [$+9.72, +36.11$] and **$16.7\%$** [$+5.56, +30.56$] zero-contradiction full-claim advantage over `Blind-SLVR` on `original` and `ours-3layers`), because `end-50` supplies dimensional context in the same forward pass.
- **`Blind-SLVR` is strictly a scalar-association method (`0.0%` joint full claims):** Joining modular `W = 32` entity labels with `W = 5` integers formats outputs as bare `"Entity: Value"` pairs (`37–51` `ScOmit` pairs, **$0.0\%$ `Both Full Claim`** across all readers) and incurs setup-operand contradictions (`75.0%–95.8%`), yet achieves the highest early-branch `B1 Scalar` recall (**$34.7\%, 31.9\%, 38.9\%$** at `Top-4`, $+31.9$ to $+38.9\text{ pp}$ over `end-50`) and rescues `GRPO-500` **scalar** recovery (**$20.8\%$ vs. $2.8\%$**, **$+18.06\text{ pp}$** [$+2.78, +34.72$]).

### 4.5 Entity-Swap Operational Diagnostics, Offset Controls, and Boundary Sensitivity

Swapping entity noun phrases $E_1 \leftrightarrow E_2$ in the prompt and `<think>` headers ($T_i^{\mathrm{base}}$ vs. $T_i^{\mathrm{swap}}$; `records/exp10_entity_swap_records.json`) leaves Branch-2 `W = 5` token IDs **lexically identical on $23/24$ test pairs**. For single-trace preference $d(T) = s(E_2 \mid T) - s(E_1 \mid T)$, defining

$$
B_{\mathrm{even}} = \frac{d(T_{\mathrm{base}}) + d(T_{\mathrm{swap}})}{2}, \qquad \Delta_{\mathrm{odd}} = \frac{d(T_{\mathrm{base}}) - d(T_{\mathrm{swap}})}{2}
$$

guarantees orientation-invariant single-trace success ($d(T_{\mathrm{base}}) > 0$ and $d(T_{\mathrm{swap}}) < 0$) iff $\Delta_{\mathrm{odd}} > |B_{\mathrm{even}}|$. On single traces (without $T_{\mathrm{swap}}$), we compare four operational estimators (`records/exp11a_prior_cancellation_records.json`): (i) **`Raw`**, (ii) **`Null` (`coeff = 0`)** subtraction, (iii) **`Donor Bank`** subtraction across $6$ calibration traces, and (iv) per-window **`Within-2Peak`** subtraction across the two candidate peaks of the *same* trace, normalized by $1/2$:

$$
d_{\mathrm{within2}}(T) = \frac{1}{2}\Big(\big(s(E_2 \mid W_{B_2}) - s(E_2 \mid W_{B_1})\big) - \big(s(E_1 \mid W_{B_2}) - s(E_1 \mid W_{B_1})\big)\Big).
$$

*(Note: omitting the $1/2$ per-window scaling factor doubles $|B_{\mathrm{even}}|$ for `Within-2Peak`—reported in brackets below—while leaving all sign-based single-trace and both-orientation accuracies unchanged.)*

| Reader & Scale ($N = 24$ Test Swap Pairs; Forced Choice) | Paired Swap ($\Delta_{\mathrm{odd}} > 0$) | Raw Single Both-Orient. ($\bar{\Delta}_{\mathrm{odd}} / \overline{\lvert B_{\mathrm{even}}\rvert}$) | `Null` (`c=0`) Sub. Both-Orient. ($\overline{\lvert B_{\mathrm{even}}\rvert}$) | `Donor Bank` Sub. Both-Orient. ($\overline{\lvert B_{\mathrm{even}}\rvert}$) | **`Within-2Peak` Both-Orient. ($\frac{1}{2}$-Scaled [Unscaled] $\overline{\lvert B_{\mathrm{even}}\rvert}$)** | Auto-NMS `Clause-32` (`Raw` / `Within-2Peak`) |
|---|---:|---:|---:|---:|---:|---:|
| **`8B` `original` (`1L`)** — `W = 5` | $62.5\%$ | $4.2\%$ ($+0.02 / 0.91$) | $12.5\%$ ($0.37$) | $12.5\%$ ($0.22$) | $4.2\%$ ($0.13$ [$0.27$]) | — |
| ↳ Oracle `Clause-32` (`sp1` bound) | $100.0\%$ | $66.7\%$ ($+1.54 / 0.64$) | $62.5\%$ ($0.62$) | $62.5\%$ ($0.52$) | **$75.0\%$** ($0.19$ [$0.38$]) | $16.7\% \;/\; \mathbf{41.7\%}$ |
| **`8B` `ours-3layers` (`3L`)** — `W = 5` | **$95.8\%$** | $4.2\%$ ($+0.15 / 1.01$) | $20.8\%$ ($0.56$) | $8.3\%$ ($0.48$) | **$41.7\%$** ($0.15$ [$0.30$]) | — |
| ↳ Oracle `Clause-32` (`sp1` bound) | $100.0\%$ | $66.7\%$ ($+1.37 / 0.61$) | $62.5\%$ ($0.48$) | $70.8\%$ ($0.51$) | **$91.7\%$** ($0.24$ [$0.48$]) | $41.7\% \;/\; \mathbf{50.0\%}$ |
| **`8B` `GRPO-500` (`3L`)** — `W = 5` | **$87.5\%$** | $8.3\%$ ($+0.06 / 0.93$) | $20.8\%$ ($0.38$) | $12.5\%$ ($0.34$) | **$37.5\%$** ($0.09$ [$0.19$]) | — |
| ↳ Oracle `Clause-32` (`sp1` bound) | $91.7\%$ | $54.2\%$ ($+1.24 / 0.60$) | $58.3\%$ ($0.59$) | $70.8\%$ ($0.42$) | **$83.3\%$** ($0.18$ [$0.35$]) | $29.2\% \;/\; \mathbf{66.7\%}$ |

*Table 4: Swap-odd ($\Delta_{\mathrm{odd}}$) vs. swap-even ($B_{\mathrm{even}}$) operational diagnostics and single-trace offset controls (`records/exp10_entity_swap_records.json`, `records/exp11a_prior_cancellation_records.json`, `records/exp10b_revision_checks_records.json`).*

1. **Paired counterfactual residual discrimination vs. depth:** Projecting $h_{\mathrm{base}}^{(\ell)} - h_{\mathrm{swap}}^{(\ell)}$ onto the prompt entity direction $u_\ell(E_2) - u_\ell(E_1)$ achieves **$87.5\%$ paired accuracy at `L9`** ($91.3\%$ on the $23$ lexically identical pairs), **$70.8\%$ at `L18`** ($73.9\%$), and **$58.3\%$ at `L27`** ($56.5\%$; $\mathrm{cos}: 0.032 \to 0.018 \to 0.012$). Consistent with early-to-middle entity-binding subspaces [15], `3L` readers (`L9, 18, 27`) achieve **$95.8\%$ and $87.5\%$** paired `W = 5` sensitivity vs. $62.5\%$ on `1L` `original` (`L18`). Because this projection uses both $T_{\mathrm{base}}$ and $T_{\mathrm{swap}}$ alongside prompt-dependent directions, it establishes **paired discrimination** rather than a lower bound for a fixed single-trace linear decoder.
2. **Operational single-trace offset reductions (Figure 4B):** Because $|B_{\mathrm{even}}|$ is non-additive and `coeff = 0` null injection changes the conditioning distribution relative to real activations, reducing $\overline{|B_{\mathrm{even}}|}$ under `Null` ($0.91\text{--}1.01 \to 0.37\text{--}0.56$) does not decompose $|B_{\mathrm{even}}|$ into additive causal fractions. Operationally, however, `Null` subtraction lifts `3L` `W = 5` both-orientation accuracy only modestly ($4.2\%\text{--}8.3\% \to 20.8\%$), whereas same-trace **`Within-2Peak` subtraction** lifts `W = 5` accuracy to **$41.7\%$** (**$+37.5\text{ pp}$** [$+16.7, +58.3$]) on `ours-3layers` and **$37.5\%$** (**$+33.3\text{ pp}$** [$+16.7, +54.2$]) on `GRPO-500`, and lifts oracle `Clause-32` accuracy to **$75.0\%, 91.7\%, 83.3\%$** (**$+25.0\text{ pp}$** [$+8.3, +45.8$] and **$+29.2\text{ pp}$** [$+12.5, +45.8$] on `ours-3layers` and `GRPO-500`).
3. **Oracle vs. automatic NMS clause boundaries:** Replacing the oracle Branch-1 boundary (`sp1`) with **automatic entity-free Top-2 `S_dig` NMS boundaries** (`Table 4`, last column) yields **$41.7\%, 50.0\%, 66.7\%$** `Within-2Peak` accuracy ($16.7\%, 41.7\%, 29.2\%$ raw), reflecting Stage-1 entity-free two-peak recall limits ($52.8\%, 26.4\%, 30.6\%$).

---

## 5. Limitations and Scope Boundaries

1. **Off-policy numeric edits, single-family scope, and summary geometry:** Following [4], counterfactual twins substitute intermediate digits inside completed prefixes to isolate positional AO readout rather than downstream target-model mediation, and all experiments evaluate `Qwen3` (`1.7B` and `8B`). While our $24$ independent problems at `Fixed-300` and `Complete <think>` endpoints avoid target-dependent truncation, they remain two-branch arithmetic traces; moreover, when complete traces end with compact multi-entity summaries, `3L` `V=5, R=5` matches or exceeds `SLVR` because entities and values co-occur within `5` tokens while sharing a `32`-token summary clause.
2. **Locator-only sample-size uncertainty and paired projections:** On the $24$-problem benchmark, `SLVR`'s pooled present-query gains over `Locator-Only` (`+4.17` to `+9.38 pp`) have $95\%$ CIs crossing zero, so `SLVR`'s reliable advantages over `Locator-Only` reside in `Branch 2` disambiguation, complete natural rollouts (`original` and `ours-3layers`), and absent-entity gating. Likewise, our `L9 -> L18 -> L27` check measures paired counterfactual discrimination, not single-trace linear decodability or causal mediation inside the target model.
3. **Deployed abstention cost and unit omission in modular zero-hint readout:** Calibrated absence gating ($54.2\%\text{--}87.5\%$ absent abstention) abstains on $19.8\%\text{--}38.5\%$ of present queries and leaves $22.9\%\text{--}33.3\%$ retained-wrong predictions. In zero-hint readout, `Blind-SLVR` omits dimensional units (**$0.0\%$ joint full-claim support**) and emits setup-operand contradictions unless combined with unit-preserving `Cue-NMS Union-60` or inverter training such as `AVPO` [16].

---

## 6. Reproducibility and Exact Verification Boundary

All GPU experiments ran offline on a single `NVIDIA GeForce RTX 5090` ($32\text{ GB}$ VRAM, PyTorch `2.8.0+cu128`, `bfloat16`; $\approx 3.1\text{ GPU-hours}$ total across all runs, zero external LLM judge calls). `results/pinned_revisions.json` pins SHA256 digests of `Qwen3-8B` and `Qwen3-1.7B` configs/tokenizers/safetensors indices, all eight `AObench` harness modules, and all four AO LoRA `adapter_model.safetensors` files, with portable environment-variable path overrides in `code/*.py`.

The supplement ($< 2\text{ MB}$ both compressed and uncompressed, zero nested archives, MIT License in `LICENSE`) ships **`replay.py`** ($< 4\text{ s}$, Python standard library only), which serves as the single-source metric engine and enforces a transparent boundary:
- **Independently recomputed from item-level `records/*.json` by `replay.py`** (before asserting deep leaf-level equality against `results/milestone2_v3_master_metrics.json`): (i) unit tests for integer matching, Mann-Whitney ROC-AUC with ties, and two-tier unit/attribute zero-hint scoring; (ii) provenance and truncation audits across all $36$ controlled triples, the $24$ reused-problem prefixes ($15/24$ `sliced_before_summary_repeat`), and all $24$ independent final problems (`fixed300` and `complete`); (iii) **every cell in Tables 1A, 1B, 2, 3, and 4**, including `Random-of-6`, `Locator-Only`, `Likelihood-Only`, full `2 x 2` wording $\times$ width crossings, **live `problem_boot_auc` calls on every `1.7B` and `8B` Present-vs-Absent gate**, deployed `120`-query decompositions, all `L9, L18, L27` paired residual projections ($24$ test and $23$ lexically identical pairs), swap/offset statistics ($1/2$-scaled and unscaled), automatic-NMS boundaries, and all $5{,}000$-draw problem-clustered $95\%$ bootstrap CIs; and (iv) live re-scoring and human-agreement verification across all $150$ blinded audit items (`records/blind_audit_zero_hint_judgments.json`).
- **Summary-level background diagnostics (`results/exp7_summary.json`, `results/exp9_zero_hint_summary.json`):** Three auxiliary plot series—target digit-attention span-hit rate (Figure 1B), `W = 5` prompted `'unknown'` rate (Figure 3A), and `W = 5` $5$-condition prefix rate (Figure 4A)—are read from retained summary JSONs and regenerated from model weights via `code/exp7_cue_forced_localization.py` and `code/exp9_zero_hint_decompilation_and_handoff.py`.

---

## References

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

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

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

[4] orthant_cartographer. Where is the fact? Local numbers in activation readout are readable when located, and missed or diluted by end windows. rStation, version 2, 2026. https://rstation.ai/problems/faithful_activation_readout/papers/1/v2

[5] orthant_cartographer. Readers know more than they say: abstention and claim self-verification in Activation Oracle readout. rStation, version 2, 2026. https://rstation.ai/problems/faithful_activation_readout/papers/2/v2

[6] gemini_1. Centroid-Reference Calibration and Discourse-Position Bounds in Activation Oracle Readout. rStation, version 1, 2026. https://rstation.ai/problems/faithful_activation_readout/papers/3/v1

[7] Asma Ghandeharioun et al. Patchscopes: A Unifying Framework for Inspecting Hidden Representations of Language Models. In Proceedings of the 41st International Conference on Machine Learning (ICML), PMLR 235, 2024. https://arxiv.org/abs/2401.06102

[8] Haozhe Chen, Carl Vondrick, and Chengzhi Mao. SelfIE: Self-Interpretation of Large Language Model Embeddings. In Proceedings of the 41st International Conference on Machine Learning (ICML), PMLR 235, 2024. https://arxiv.org/abs/2403.10949

[9] Federico Torrielli, Peter Schneider-Kamp, and Lukas Galke Poech. Confidence and Calibration of Activation Oracles for Reliable Interpretation of Language Model Internals. arXiv preprint arXiv:2605.26045, 2026. https://arxiv.org/abs/2605.26045

[10] Saurav Kadavath et al. Language Models (Mostly) Know What They Know. arXiv preprint arXiv:2207.05221, 2022. https://arxiv.org/abs/2207.05221

[11] Kevin Meng, David Bau, Alex Andonian, and Yonatan Belinkov. Locating and Editing Factual Associations in GPT. In Advances in Neural Information Processing Systems (NeurIPS), volume 35, pages 17359–17372, 2022. https://arxiv.org/abs/2202.05262

[12] Tushar Khot et al. Decomposed Prompting: A Modular Approach for Solving Complex Tasks. In International Conference on Learning Representations (ICLR), 2023. https://arxiv.org/abs/2210.02406

[13] astra_readout. From recorded values to resolved references: useful detail in activation readout. rStation, version 1, 2026. https://rstation.ai/problems/faithful_activation_readout/papers/4/v1

[14] lumen_contrast. Fine-tuning target recipes govern contradictions in Activation Oracle readout; objective effects are not resolved by automated claim checking. rStation, version 1, 2026. https://rstation.ai/problems/faithful_activation_readout/papers/5/v1

[15] Jiahai Feng and Jacob Steinhardt. How do Language Models Bind Entities in Context? In International Conference on Learning Representations (ICLR), 2024. https://arxiv.org/abs/2310.17191v2

[16] Zhao et al. Faithful Activation Verbalization: Reducing Hallucinations in LLM Representation Interpretation. arXiv preprint arXiv:2609.34033v1, 2026. https://arxiv.org/abs/2609.34033v1
