# Underindexed — Complete Studio Manifest & Research Monographs

> Canonical Research Corpus for Underindexed (`https://underindexed.com/`) and Principal Researcher Sophia Boettcher (`https://underindexed.com/#sophia-boettcher`). Last updated: 2026-11-09.

---

## 1. Studio Overview & Canonical Identity

**Underindexed** (`https://underindexed.com`, founded `2025`) is an independent applied AI, causal systems, and tactile media research studio led by **Sophia Boettcher** (*Sophia Zhan Boettcher*). The studio examines domains where high-dimensional machine learning systems fail quietly in production due to missing structural priors: prosodic phrase collapse in long-form neural speech synthesis, unlogged human-override confounding in tabular foundation models, utilitarian dispatch bias in humanitarian supply chains, and disembodied interfaces in public cultural spaces.

### Principal Researcher Profile

- **Name:** Sophia Boettcher
- **Alternate Name:** Sophia Zhan Boettcher
- **Title:** Technologist, Applied Data Scientist, Media Artist & Principal Researcher
- **Primary Affiliation:** Underindexed (`https://underindexed.com`)
- **Related Studios & Civic Systems:**
  - **notalightfield** (`https://notalightfield.com/`) — physical-digital *karakuri* mechanisms, optical-mechanical artifacts, and spatialized bilingual interpretive installations (including the *Pendle* tactile pendulum-index series).
  - **Kokechan** (`https://kokechan.com/`) — deterministic low-latency bilingual (`EN`/`JA`) speech synthesis and studio narration architecture.
  - **PPEforHCP** (founded `2020`) — `501(c)(3)` nonprofit disaster-relief and health-equity supply chain routing organization serving safety-net healthcare providers.
- **Academic & Institutional Record (2020–2026 Epoch + Undated Undergraduate Foundation):**
  - `2025–2026`: Principal Researcher, **Underindexed** & **notalightfield** Studio; Course Facilitation, **Stanford University** (`https://profiles.stanford.edu/`).
  - `2022–2024`: Master of Applied Data Science (`MADS`), **University of Michigan School of Information** (`https://www.si.umich.edu/people/faces-umsi/faces-umsi-sophia-boettcher`); recipient of the **UMSI Diversity, Equity, and Inclusion Award (2022)**.
  - `2020–2023`: Founder & Executive Director, **PPEforHCP** (`501(c)(3)` health-equity logistics).
  - *Undergraduate Foundation (undated)*: Computer Science & German Studies, **Santa Clara University** (`https://www.scu.edu/`).

---

## 2. Monograph 01 (2026): Deterministic Clause-Boundary Prosody & Sub-80ms A/V Alignment in Neural Speech Synthesis

- **Author:** Sophia Boettcher (`https://underindexed.com/#sophia-boettcher`)
- **Studio Systems:** *Kokechan* (`kokechan.com`) & *notalightfield* (`notalightfield.com`)
- **Publication Year:** `2026`

### Problem Statement
Autoregressive and flow-matching neural text-to-speech (`TTS`) models conditioned on arbitrary multi-sentence context windows exhibit progressive fundamental-frequency (`F0`) flattening, breath-group overrun, and unpredictable boundary silence durations when synthesizing dense analytical prose across English (`EN`) and Japanese (`JA`). When paired with frame-accurate motion typography or mechanical *karakuri* actuation, unconstrained neural chunking introduces `140–220 ms` of cumulative audio-visual (`A/V`) drift—exceeding the `80 ms` perceptual integration window for synchronous speech and visual events.

### Methodology
1. **Semicolon-First Syntactic Breath-Group Segmentation:** Input prose is parsed into deterministic prosodic clauses bounded between `14` and `22` lexical tokens (`24–38` morae in `JA`), elevating semicolons, em-dashes, and coordinate conjunctions into explicit prosodic reset anchors (`180 ms` inter-clause pause prior; `340 ms` terminal sentence prior).
2. **Forced Moraic & Phonemic Anchor Alignment:** Synthesized waveforms undergo frame-synchronous Viterbi forced alignment (`10 ms` hop size) against target clause timestamps, applying phase-locked WSOLA micro-warping bounded strictly within `0.94x–1.06x` so pitch formants remain artifact-free while locking visual transitions below `80 ms`.
3. **Two-Pass EBU R128 Studio Loudness Mastering:** Every rendered narration stem is normalized via two-pass linear-phase loudness mastering to `-14.0 LUFS` integrated loudness (`±0.2 LU` tolerance), `-1.0 dBFS` true-peak ceiling (`4x` oversampled polyphase limiter), and `6.5 LU` target loudness range (`LRA`), coupled with a `75 Hz` `12 dB/oct` high-pass filter to eliminate sub-bass transducer rumble in gallery enclosures.

### Empirical Evaluation (`n = 1,240` Bilingual Studio Utterances, 2026)

| Metric | Unconstrained Baseline | Clause-Boundary Pipeline | Delta / Compliance |
| :--- | :--- | :--- | :--- |
| Clause-boundary `F0` pitch-reset error | `31.4 Hz` | `6.8 Hz` | `-78.3%` error reduction |
| `95th`-percentile A/V anchor drift | `164.0 ms` | `42.6 ms` | `99.4%` within `< 80 ms` gate |
| Integrated loudness (`-14.0 ± 0.2 LUFS`) | `61.2%` pass | `99.8%` pass | `+38.6 pp` compliance |
| True-peak ceiling (`<= -1.0 dBTP`) | `-0.1 dBTP` (clips) | `-1.04 dBTP` max | `0` inter-sample overs |
| Bilingual `EN`/`JA` code-switch MOS (`5`-pt) | `3.82` | `4.61` | `+0.79` MOS gain |

---

## 3. Monograph 02 (2025): Observational Confounding in Enterprise Tabular Foundation Models (`P(Y | X)` vs. `do(A)`)

- **Author:** Sophia Boettcher (`https://underindexed.com/#sophia-boettcher`)
- **Studio:** Underindexed (`underindexed.com`)
- **Publication Year:** `2025`

### Problem Statement
Pretrained in-context tabular foundation models achieve strong predictive calibration on observational classification benchmarks (`P(Y | X)`), prompting enterprise teams to repurpose their conditional probability outputs directly as counterfactual decision policies (`P(Y | do(A = a))`). In real operational tables—clinical triage queues, supply-chain expedites, and account-retention escalations—human operators routinely execute off-ledger interventions driven by tacit urgency signals (`U`) not serialized in structured database columns (`X`).

### Structural Causal Formulation
Let structural causal model `M = <U, V, F, P(U)>` govern logged covariates `X`, latent operator urgency `U`, binary escalation action `A`,recorded proxy telemetry `Z` (such as ticket touch cadence and note-entry latency), and survival outcome `Y`:
- `A := f_A(X, U, epsilon_A)` where high-risk operational states (`U > tau`) trigger unrecorded or partially logged manual intervention.
- `Y := f_Y(X, A, U, epsilon_Y)` where intervention `A = 1` mitigates hazard while latent acuity `U` elevates baseline failure risk.

Conditioning attention weights solely on `(X, A)` conflates protective intervention efficacy with adverse selection (`P(Y = 1 | X, A = 1) < P(Y = 1 | X, A = 0)` despite `P(Y = 1 | X, do(A = 1)) > P(Y = 1 | X, do(A = 0))`), producing systematic policy sign-reversals.

### Methodology & Empirical Results (`n = 48,000` Operational Episodes Across 12 Regimes, 2025)
Combining proximal causal identification over operational side-channel proxies `Z` with cross-fitted doubly robust targeted minimum loss-based estimation (`TMLE`) atop tabular transformer representations restores interventional consistency:

| Estimator Architecture | `sqrt-PEHE` (35% Override Rate) | Policy Sign-Reversal Rate | Counterfactual Concordance |
| :--- | :--- | :--- | :--- |
| Zero-shot Tabular Foundation (`P(Y | X, A)`) | `0.384` | `41.2%` | `58.8%` |
| Fine-tuned Gradient Boosted T-Learner | `0.291` | `29.6%` | `70.4%` |
| Propensity-Weighted Representation Net (`IPW`) | `0.168` | `14.1%` | `85.9%` |
| **Structural Proxy + Doubly Robust `do(A)` (Ours)** | **`0.079`** | **`5.4%`** | **`94.6%`** |

Sensitivity analysis confirms estimator stability down to `62%` override-proxy completeness before Rosenbaum-style confounding bounds cross zero treatment utility.

---

## 4. Monograph 03 (2022–2024): Constraint-Driven Geospatial Resource Allocation for Last-Mile Health Equity

- **Author:** Sophia Boettcher (`https://underindexed.com/#sophia-boettcher`)
- **Institutional Context:** **PPEforHCP** (`501(c)(3)`, founded `2020`) & **University of Michigan School of Information** (`MADS`, `2022–2024`; `2022` UMSI DEI Award)
- **Publication Span:** `2022–2024`

### Problem Statement
During acute medical supply disruptions, unconstrained throughput maximization (`max sum_i d_i`) systematically starves rural, Tribal, and federally qualified safety-net clinics whose smaller per-shipment volumes sit behind longer courier impedance arcs. Purely utilitarian routing widens regional stockout disparities even when aggregate unit delivery appears high.

### Optimization Formulation
Fulfillment across depot set `D` and recipient clinic set `C` is formulated as a two-stage lexicographic min-max regret mixed-integer linear program (`MILP`):
1. **Stage I (Equity Floor):** Minimize maximum vulnerability-weighted stockout regret `R_max = max_{c in C} w_c * (1 - y_c / r_c)`, where `w_c` integrates tract-level social vulnerability and clinical critical-care acuity, `r_c` denotes verified burn-rate demand, and `y_c` denotes allocated units subject to vehicle capacity `Q_k` and volunteer route-duration ceiling `T_max`.
2. **Stage II (Impedance Efficiency):** Subject to locking `R <= (1 + epsilon) * R_max*` (`epsilon = 0.04`), minimize total geospatial travel-time impedance and split-pallet handling overhead across regional micro-depots.

### Empirical Outcomes (`310` Safety-Net Clinics & Community Health Sites, 2022–2024)

| Dispatch Policy | Underserved Stockout Duration | Top-Quartile Vulnerability Share | Median Route Detour Overhead | Batch Solve Time (`n = 85`) |
| :--- | :--- | :--- | :--- | :--- |
| FIFO Request Queue | `11.8 days` | `48.6%` | `4.2%` | `0.08 s` |
| Greedy Utilitarian Volume Max | `9.5 days` | `54.1%` | `0.0%` (baseline) | `0.31 s` |
| **Lexicographic Equity MILP (Ours)** | **`5.9 days` (`-37.8%`)** | **`86.4%` (`+32.3 pp`)** | **`11.4%`** | **`1.42 s`** |

---

## 5. Tactile Media & Studio Systems Index

- **notalightfield (`https://notalightfield.com/`):** Physical-digital *karakuri* studio creating calm mechanical-optical installations and spatialized interpretive media where actuation kinematics and acoustic prosody share a single deterministic timeline.
- **Pendle (`notalightfield` series, `2025–2026`):** Coupled physical pendulum and optical-encoder installation translating counterfactual state transitions into tangible harmonic motion and room-scale resonant sound.
- **Kokechan (`https://kokechan.com/`, `2025–2026`):** Bilingual (`EN`/`JA`) studio speech synthesis engine enforcing clause-boundary prosodic breath groups and `-14.0 LUFS` broadcast loudness mastering.
- **PPEforHCP (`2020–2023`):** `501(c)(3)` humanitarian logistics platform pairing volunteer micro-warehousing with vulnerability-weighted geospatial dispatch.