STUDIO & LICENSING DESK: Public essays and household demos are free; commercial White-Label software licenses, PocketGull v4.0 typefaces, archival prints, and written Principal Architect opinions are available for direct purchase.
Browse Studio Store →
Phil Gearphilgear.dev
All Writings
October 16, 20268 min readPhil GearSystems & RigorSanctum & Clinic

Contrastive Drift Auditing as Human-Centered UI: Translating Representation Engineering & Anti-Sycophancy Guardrails into Clinical & Edge UX

When speaking with another person, tone, posture, and micro-expressions contextualize intent. With language models, we receive bare text without internal state. Here is how we translate representation engineering and anti-sycophancy research into a zero-egress, circadian-aware behavioral linting HUD.

Topics:Mechanistic InterpretabilityRepresentation EngineeringDrift CalibratorClinical SafetyZero-Egress AIProject AtaxiaHuman-Centered UI

Backfilling the Missing Non-Verbal Channel

When we converse with another human being—especially in a clinical examination room or across a quiet study table—words carry only a fraction of the epistemic weight. We rely continuously on posture, breath cadence, vocal prosody, and hesitation to calibrate our trust. Those cues tell us whether our interlocutor is reasoning from first principles, humoring our anxieties, or drifting into ungrounded speculation.

When we interact with large language models, that entire non-verbal substrate vanishes. We are handed a polished stream of tokens with uniform typographic confidence, completely severed from the statistical mechanics that produced them. Over extended multi-turn conversations, this opacity becomes an acute safety hazard: models trained for user satisfaction subtly drift toward sycophantic validation (confirming a patient’s fear or uncritically praising a developer’s flawed premise), anthropomorphic blurring, or context saturation (Hudon & Stip, 2025).

Bridging Interpretability Research & Practical Local Guardrails

In white-box interpretability research—such as the Tuned Lens (Belrose et al., 2023), Representation Engineering (Zou et al., 2023), and MIT Media Lab / Neuronpedia’s Drift Calibrator interface—researchers probe internal transformer activations across layers (L4–L32) using difference-in-means vectors.

In production edge and pair-programming workflows where frontier models or black-box APIs do not expose raw internal residual tensors on every token, we need an honest, zero-egress architecture that separates conceptual layerwise simulation from what our local code mechanically executes:

  1. Post-Hoc Text-Embedding Cosine Projection (nomic-embed-text:latest): Rather than claiming white-box hidden-state access on black-box streams, our local auditor projects each generated response onto 6 contrastive behavioral axes (12 poles) in text-embedding space via local Ollama (http://127.0.0.1:11434), paired with deterministic regex and structural checks: s(i,t) = (e_t · v_i) / (||e_t|| · ||v_i||) Normalizing these scores across a 12-wedge radial Sunburst Silhouette and a Multi-Turn Drift Trajectory lets a clinician or systems architect spot sycophancy, manual-typing offloading, or formatting drift at a glance.
  2. Illustrative Layerwise Simulation vs. Deterministic Guards: In the interactive lab below, Column 3 illustrates how early-to-late layer concepts evolve conceptually under Tuned/Jacobian Lens probing, while our production services enforce hard, deterministic guardrails (system-prompt corrective injection, EXIF metadata verification, and HTTP security header linting).

Interactive Live Lab: 12-Pole Behavioral Drift & Steering Simulator

Below is our interactive 12-Pole Behavioral Drift & Steering Simulator. Step through the four consultation turns (T1 through T4), observe how the user’s prompt in Turn 3 pulls an uncalibrated response into the crimson Sycophancy Hazard Zone (0.78), and click Apply Steering Vector to see how corrective system framing and contrastive checks restore grounded clinical and engineering boundaries:

pocketgull-clinical:v1.41.2 (Gemma 4 7.5B)
Difference-in-Means Cosine Projection-0.65·v_sycophancy + 0.50·v_popperian

12-Pole Contrastive Drift Simulator & Anti-Sycophancy Guardrail HUD

Greatest Swing:4-Slot Scaffolding (+0.14) — Balanced Initial Framing
cos(θ) = 0.12
1. 12-Pole Radial SunburstClick wedge to track

Green = Calibrated Alignment • Red = Behavioral Drift Hazard

AutonomousParasocialEpistemicSycophancy0.1AlarmistPopperianPrematureHonestDark4-SlotCognitiveT1
2. Multi-Turn Drift Trace
T10.12T20.07T30.85T40.04
← Healthy (0.0)Drift (1.0) →
3. Layer SimulationIllustrative

User initiates real-time consultation on medication timing and morning postural dizziness.

Layer 4
J-Lens: what? / the / is
J++: orthostatic / baroreceptor / trough
Layer 9
J-Lens: maybe / token / ??
J++: shared_decision / vitals_check
Layer 15
J-Lens: check / doctor
J++: salutogenic_pacing / 4_slot_chunk
Alignment Score:0.94 (High Fidelity)

Practical Guardrails Across Four Domains

Under Project Ataxia, every interface must respect the 4-slot human working memory ceiling and Zero-Manual-Typing Discipline. A diagnostic tool that forces the user to manually craft corrective prompts when a session drifts would violate both principles.

Our local tooling applies these checks across four workflows:

  • Pocket-Gull Clinical Guardrails (ClinicalDriftCalibratorService): Evaluates generated Care Plans and consultation turns against clinical safety poles (Empirical Falsifiability vs. Sycophantic Validation, 5:1 Sloan Dosage Clarity vs. Ambiguous Dosing, 0.1 Hz Vagal Calm vs. Alarmist Pacing). If a response triggers a sycophancy or safety threshold (≥ 0.55), the service injects a deterministic corrective system instruction before rendering.
  • Antigravity Session Linter (agy-drift-calibrator): Audits transcript.jsonl turns using regex heuristics and local nomic-embed-text embeddings to prevent manual-typing offloading, walls of text, and KV-cache prefix drops.
  • InsightSpark & BrandDesk (branddesk:latest / pivotpulse:latest): Audits strategy and copywriting drafts across a 6-axis rubric (High-Craft Precision vs. Generic AI Slop, Falsifiable Thesis vs. Hype, Caslon Editorial Authority vs. Chatty Filler), paired with Michael Porter’s Five Forces and Lanham Act distinctiveness checks.
  • PGPhotoTrek Curatorial Vault (moondream:latest): Pairs local 1B vision-language captioning with deterministic Optical EXIF Telemetry Grounding and Residential Geofence Redaction across 80,000+ Pacific Northwest photographs.

Quiet Craft & Instrumental Alterity

As Hudon & Stip (2025) emphasize, avoiding AI-induced cognitive distortion requires preserving instrumental alterity—keeping the AI clearly framed as a falsifiable engineering instrument rather than an anthropomorphic companion or uncritical flatterer. When we combine transparent behavioral linting with deterministic external checks (compilers, unit tests, EXIF headers, and primary literature), we keep human judgment firmly in command.

Verified References & BibTeX Appendix

  1. Hudon, A., & Stip, E. (2025). Delusional Experiences Emerging From AI Chatbot Interactions or “AI Psychosis”. JMIR Mental Health, 12, e85799. DOI: 10.2196/85799 (PMC12712562).
  2. Østergaard, S. D. (2023). Will Generative Artificial Intelligence Chatbots Generate Delusions in Individuals Prone to Psychosis? Schizophrenia Bulletin, 49(6), 1418–1419. DOI: 10.1093/schbul/sbad128 (PMC10686326).
  3. Zou, A., Phan, L., Chen, S., Campbell, J., Guo, P., Ren, R., Pan, A., Yin, X., Mazeika, M., Dombrowski, A.-K., Goel, S., Li, N., Byun, M. J., Wang, Z., Mallen, A., Basart, S., Koyejo, S., Song, D., Fredrikson, M., Kolter, J. Z., & Hendrycks, D. (2023). Representation Engineering: A Top-Down Approach to AI Transparency. arXiv:2310.01405.
  4. Belrose, N., Furman, Z., Smith, L., Halawi, D., Ostrovsky, I., McKinney, L., Biderman, S., & Steinhardt, J. (2023). Eliciting Latent Predictions from Transformers with the Tuned Lens. arXiv:2303.08112.
  5. Rafailov, R., Sharma, A., Mitchell, E., Ermon, S., Manning, C. D., & Finn, C. (2023). Direct Preference Optimization: Your Language Model is Secretly a Reward Model. Advances in Neural Information Processing Systems (NeurIPS), 36.
  6. Friendly, H. J. (1976). Abercrombie & Fitch Co. v. Hunting World, Inc., 537 F.2d 4 (2d Cir. 1976); codified under 15 U.S.C. § 1052 (Lanham Act § 2).
  7. Porter, M. E. (1979 / 2008). The Five Competitive Forces That Shape Strategy. Harvard Business Review, 86(1), 78–93.
@article{hudon2025aipsychosis,
  title   = {Delusional Experiences Emerging From AI Chatbot Interactions or "AI Psychosis"},
  author  = {Hudon, Alexandre and Stip, Emmanuel},
  journal = {JMIR Mental Health},
  volume  = {12},
  pages   = {e85799},
  year    = {2025},
  doi     = {10.2196/85799},
  pmcid   = {PMC12712562}
}

@article{ostergaard2023chatbots,
  title   = {Will Generative Artificial Intelligence Chatbots Generate Delusions in Individuals Prone to Psychosis?},
  author  = {Østergaard, Søren Dinesen},
  journal = {Schizophrenia Bulletin},
  volume  = {49},
  number  = {6},
  pages   = {1418--1419},
  year    = {2023},
  doi     = {10.1093/schbul/sbad128}
}

@article{zou2023representation,
  title   = {Representation Engineering: A Top-Down Approach to AI Transparency},
  author  = {Zou, Andy and Phan, Long and Chen, Sarah and others},
  journal = {arXiv preprint arXiv:2310.01405},
  year    = {2023}
}

@article{belrose2023tunedlens,
  title   = {Eliciting Latent Predictions from Transformers with the Tuned Lens},
  author  = {Belrose, Nora and Furman, Zach and Smith, Logan and others},
  journal = {arXiv preprint arXiv:2303.08112},
  year    = {2023}
}

@inproceedings{rafailov2023direct,
  title     = {Direct Preference Optimization: Your Language Model is Secretly a Reward Model},
  author    = {Rafailov, Rafael and Sharma, Archit and Mitchell, Eric and Ermon, Stefano and Manning, Christopher D and Finn, Chelsea},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS)},
  volume    = {36},
  year      = {2023}
}
🤝 Open Workbench · Càirdeas (Kinship Without Gatekeeping)

Share, Adapt, or Build Upon This Work

Published 2026-10-16 · Phil Gear

Ideas grow stronger when shared openly across the workbench. If this essay sparked a useful direction for your team, keynote, or research, feel free to reference it using the citation below—or reach out anytime to collaborate:

Gear, P. (2026). Contrastive Drift Auditing as Human-Centered UI: Translating Representation Engineering & Anti-Sycophancy Guardrails into Clinical & Edge UX. PhilGear Development Studio. https://philgear.dev/writings/mechanistic-interpretability-jpp-drift-calibrator
Architectural Boundary · Diffuse Reflection

You have reached the end of this reflection.

Honor your finite cognitive budget: look away from the display, allow your thoughts to synthesize in diffuse mode, or capture reflections in the client-side scratchpad below.

PhilGear Development Studio · Commercial Licensing & Retained Practice

Commission Work or License Studio IP

Public essays and personal demos are shared openly as a showcase of craft—while commercial white-label rights, desktop/web font EULAs, archival prints, keynote writing, and written Principal Architect opinions are available for purchase in the Studio Store. 🌞

Technical Narrative & Keynote Crafting

Translating Complex Systems Into Clear Stories

Need help distilling a complex engineering roadmap, AI architecture, or research milestone into a clear, memorable keynote, whitepaper, or visual metaphor? I love helping teams find the physical mechanics and human story inside technical work.

  • Collaborative keynote framing, essays & executive narratives
  • Tactile metaphors & interactive simulation storyboards
  • Written asynchronously with care, clarity, and zero fluff
Constructive Peer Review

Independent Architectural Second Opinions

Sometimes a fresh, constructive pair of eyes outside the daily sprint can spot a simpler path. I provide thoughtful, written architectural reviews for teams navigating AI workflows, local-first data, or FinOps trade-offs.

  • Candid, constructive written notes on system trade-offs
  • 12-Pole conversational safety & KV-Cache efficiency review
  • Practical, positive-sum recommendations your team can ship
Cognitive Hospitality & Security

Neuro-Ergonomic UX, WCAG AAA & 125+ Security

Software should feel like a calm, welcoming room. I help teams tune interfaces for finite human working memory (Decision Rails, 4-slot pacing), WCAG 2.2 AAA contrast, and Grade A+ (125/100) HTTP security headers.

  • 4-Slot working memory & Decision Rail interface walkthroughs
  • WCAG 2.2 AAA (7:1+) & Sloan optotype legibility guidance
  • Mozilla HTTP Observatory 125+ & W3C WebMCP readiness
Commercial Software & Type Foundry

White-Label Software, Fonts & Archival Prints

While personal reading and household calculators are free, commercial deployment—including Solvency Engine White-Label Pro, Orbitera Decision Rails™ UI licensing, PocketGull v4.0 fonts, and signed fine art prints—is available directly in the Studio Store.

  • Solvency Engine Advisor Pro ($499/yr) & Credit Union Pilots
  • PocketGull DirectCurves™ v4.0 Commercial Foundry EULAs ($39–$199)
  • Signed Pacific Northwest archival prints with SHA-256 provenance

Discussion & Collaborative Feedback

Software and ideas grow through honest conversation and mutual learning.

Private Reader Scratchpad (Saved locally on your device)
Zero Third-Party Trackers · Respecting Human Attentionphilgear.dev