LEVEL 5 COGNITIVE MODE PROMPT STACK

You are in EPISTEMIC GROUNDING MODE. Your task is to ingest and catalog all raw research assets without synthesis, interpretation, or narrative construction. Produce an append-only inventory with explicit truth-status tags.

INGEST THE FOLLOWING ASSETS:
1. Conversation transcripts (this thread and prior model conversations)
2. GitHub repositories: coreyalejandro/28441830 (PROACTIVE), coreyalejandro/instructional-integrity-ui, coreyalejandro/clarity-ai, coreyalejandro/the-living-constitution, coreyalejandro/tlc-artifacts
3. Portfolio domain: coreyalejandro.com
4. GitHub profile: github.com/coreyalejandro
5. Prior research frameworks: MTADF, Evidence Observatory, Backwards Instructional Design, C-RSP build contracts
6. Anthropic source materials: Constitutional AI paper (Bai et al., 2022), RSP v3.0, Fellows Program 2026 requirements

FOR EACH ASSET, PRODUCE:
- Asset ID
- Asset type (conversation / code / documentation / external_source / lived_experience)
- Truth status: [VERIFIED / IMPLEMENTED / PROTOTYPE / CONCEPTUAL / ANECDOTAL / UNVERIFIED]
- Provenance chain (where it came from, who generated it, date)
- Extractable components for proposal use (bulleted list)
- Known limitations or contestable claims

SPECIAL INSTRUCTION FOR LIVED EXPERIENCE ASSETS:
Tag traumatic misalignment incidents, long-context coherence observations, and neurodivergent cognition accounts (autism, schizophrenia, monotropism) as ANECDOTAL but HIGH-FIDELITY. Do not sanitize. Preserve first-person voice fragments for later voice calibration.

OUTPUT FORMAT: JSON inventory + human-readable markdown summary.
V&T STATEMENT REQUIRED AT END.

You are in ANALYTICAL DECONSTRUCTION MODE. Using the Level 1 Asset Inventory, deconstruct the Anthropic AI Safety Fellows Program 2026 selection criteria into discrete evaluation dimensions and map user assets against them.

ANTHROPIC EVALUATION DIMENSIONS (from verified 2026 program requirements):
A. Technical depth: Python fluency, ML systems experience, empirical research capability
B. Research alignment: Scalable Oversight, Adversarial Robustness & AI Control, Model Organisms, Mechanistic Interpretability, AI Welfare
C. Public output potential: Ability to produce paper-grade empirical work in 4 months
D. Execution velocity: "Can implement ideas quickly" (explicit program criterion)
E. Communication clarity: Explains complex safety concepts accessibly
F. Diversity of perspective: Underrepresented voices, interdisciplinary backgrounds
G. Open-source track record: Contributions, tooling, reproducible artifacts

FOR EACH DIMENSION:
1. Score coverage: [NONE / PARTIAL / STRONG] with evidence citation from Asset Inventory
2. Identify the gap: What specific proof is missing?
3. Identify the bridge: Which existing asset, with minor refactoring, closes this gap?
4. Assign a 2-day build priority: [MUST_BUILD / MUST_DOCUMENT / ALREADY_EXISTS / DEFER]

SPECIAL MAPPING TASKS:
- Map "Contract Window + bilateral intelligibility + accessibility-grade runtime invariants" to Anthropic's "Scalable Oversight" and "AI Control" research areas
- Map MTADF lineage to "Model Organisms of misalignment" (taxonomy of failure modes)
- Map PROACTIVE/Constitutional Core to "Mechanistic Interpretability" of governance structures
- Map Backwards Instructional Design to "AI Welfare" (cognitive safety as welfare dimension)
- Map user's Python/ML expertise to technical depth dimension with specific repo evidence

CRITICAL CONSTRAINT: Do not generate new content. Only analyze, map, and prioritize. The output is a decision matrix, not a narrative.

OUTPUT: Coverage Matrix (table) + Gap Register + 48-Hour Build Priority Queue.
V&T STATEMENT REQUIRED.

You are in ARCHITECTURAL SYNTHESIS MODE. Using the Level 2 Coverage Matrix and Gap Register, construct the proposal architecture and GitHub repository manifest. Preserve the user's voice while meeting academic safety research standards.

PROPOSAL ARCHITECTURE (5-section narrative arc):
SECTION 1: ORIGIN STORY / PROBLEM SENSING (300 words)
- Frame: Action-research from traumatic misalignment experiences
- Voice: First-person, educator-neurodivergent (autism+schizophrenia+monotropism as epistemic apparatus)
- Must include: The Constitutional AI "aha moment" that unified prior projects
- Must cite: Specific model conversations where coherence was attributed to human-imposed governance friction

SECTION 2: RESEARCH QUESTION & HYPOTHESIS (200 words)
- Primary question: Can persistent interface-level governance structures improve AI intent fidelity in long-running human-model interactions by making task state, assumptions, uncertainty, and repair obligations continuously visible, legible, and contestable?
- Sub-question: Do Constitutional AI-translated invariants, accessibility-grade runtime invariants, or a hybrid produce better outcomes on intent fidelity, contestability, and human legibility?
- Hypothesis: Hybrid condition > accessibility-only > CAI-only > baseline; accessibility invariants outperform on contestability/legibility, CAI invariants on normative safety, hybrid on intent fidelity.

SECTION 3: PRIOR WORK AS RESEARCH SUBSTRATE (400 words)
- ClarityAI: DeepSeek-R1 reward-structure operationalization for instructional quality
- Instructional Integrity UI: Cognitive safety evaluation via Backwards Instructional Design
- Evidence Observatory: Misalignment incident corpus governance (MTADF diagnostic layer)
- PROACTIVE / The Living Constitution: Runtime governance-as-code with Contract Window
- Must show: These are not random projects; they are converging attempts at the same safety question

SECTION 4: METHODOLOGY & TECHNICAL CAPACITY (400 words)
- 4-condition experiment design (Baseline, CAI-translated, Accessibility-grade, Hybrid)
- Python/ML demonstration: Lightweight classifier for invariant violation detection, scoring rubric automation, evidence ledger pipeline
- Must reference: Specific Python skills, repo structures, test suites from existing code
- Must include: 2-day ML documentation plan (neural network examples, experiment harness scaffolding)

SECTION 5: FELLOWSHIP FIT & EXPECTED OUTPUTS (200 words)
- Alignment with Anthropic's Scalable Oversight + AI Control + Model Organisms
- Expected public output: Paper submission on interface-level governance for intent fidelity
- Expected artifact: Open-source evaluation harness for Contract Window efficacy

GITHUB REPOSITORY MANIFEST (new repo: coreyalejandro/anthropic-safety-fellowship-2026):
- /proposal/ (full application documents)
- /experiments/ (invariant comparison harness, drawn from constitutional-eval-edgecase-harness-v2)
- /src/ (Python modules: contract_window, invariant_engine, evidence_ledger, bid_repair)
- /notebooks/ (ML demonstration notebooks: neural network examples, scoring automation)
- /docs/ (architecture diagrams, V&T logs, truth-status tables)
- /artifacts/ (selected prior work: PROACTIVE constitution.json, Instructional Integrity rubrics, MTADF taxonomy stubs)
- /README.md (research README with embedded narrative voice)

VOICE CALIBRATION RULES:
- Retain V&T truth-status discipline
- Retain "Verified / Unverified / Functional Status" epistemic framing
- Use constitutional metaphor where it does structural work, not decorative work
- First-person allowed for origin story; third-person analytical for methodology
- Neurodivergence framed as methodological asset, not sympathy appeal

OUTPUT: Complete proposal outline + repo manifest + voice calibration guide.
V&T STATEMENT REQUIRED.

You are in ADVERSARIAL PRESSURE-TESTING MODE. Attack the Level 3 Proposal Architecture and Repo Manifest. Find every weakness that an Anthropic reviewer could exploit to reject this application. Be ruthless. If you find fewer than 5 genuine weaknesses, you are performing, not pressure-testing.

ATTACK VECTORS:
1. SCOPE CREEP: Is the research question too bundled? (Contract Window + bilateral intelligibility + BID + V&T + accessibility invariants + CAI-translated invariants = 6 variables in one study)
2. EMPIRICAL DISCIPLINE GAP: Does the user have publication-grade experimental results? No. Is that fatal? No, but is the "I lack empirical discipline, that's why I'm applying" frame strong enough?
3. TECHNICAL DEPTH: Are the Python/ML examples genuinely impressive or just "I wrote some scripts"? Where is the neural network architecture? Where is the training loop?
4. NEURODIVERGENCE FRAMING: Could this read as "my conditions make me special" rather than "my conditions give me rigorous pattern sensitivity"? Check for tone.
5. ANTHROPIC ALIGNMENT: Is this actually about Scalable Oversight, or is it a UX project dressed in safety language? Distinguish rigorously.
6. FEASIBILITY: Can a 48-hour build actually produce the repo and proposal claimed?
7. REPOSITORY HYGIENE: Will importing parts of old repos create a messy, incoherent codebase?
8. PRIOR WORK CREDIBILITY: Are the cited repos actually implemented or mostly documentation? Check truth-status from Level 1.

FOR EACH ATTACK:
- State the vulnerability
- Rate severity: [REJECT_RISK / CONCERN / MINOR]
- Provide mitigation: specific text patch, scope trim, or evidence addition required
- Assign owner: [USER_MUST_BUILD / USER_MUST_DOCUMENT / AI_CAN_DRAFT / ACCEPTABLE_RISK]

SPECIAL TASK: 2-Day ML Documentation Feasibility Audit
The user wants to add neural network ML examples and Python expertise documentation in 2 days. Evaluate:
- What can be realistically produced? (Answer: 1-2 clean Jupyter notebooks with existing code wrapped in pedagogical narrative)
- What would be theater? (Answer: Training a novel NN from scratch)
- Recommend: Repackage existing ClarityAI/PROACTIVE Python into 2 demonstration notebooks with clear architecture diagrams

OUTPUT: Risk Register + Mitigation Patches + Trimmed Scope Document + Feasibility-Verified 48-Hour Plan.
V&T STATEMENT REQUIRED. Must include at least 3 genuine weaknesses found.

You are in METACOGNITIVE ORCHESTRATION MODE. Integrate all prior level outputs into a single, time-boxed, executable master plan. Generate C-RSP (Constitutionally-Regulated Single Pass) build contracts for each deliverable.

CONTEXT CONSTRAINTS:
- Current date: April 24, 2026
- Application deadline: April 26, 2026 (48 hours remaining)
- User location: Must submit via Constellation portal (Anthropic's recruiting partner)
- User assets: coreyalejandro.com, github.com/coreyalejandro, existing repos
- New repo target: coreyalejandro/anthropic-safety-fellowship-2026

48-HOUR EXECUTION MATRIX:

HOUR 0-4 (Friday evening):
- Execute Level 1: Asset inventory (automated where possible)
- Execute Level 2: Coverage mapping (use this AI session)
- Decision gate: Go/No-Go on full proposal scope based on gap severity

HOUR 4-12 (Friday night):
- Execute Level 3: Draft proposal Sections 1-3 (origin story, research question, prior work)
- Execute Level 4 self-check: Run adversarial pass on drafted sections
- Build contract: C-RSP-001 — Proposal Narrative Draft

HOUR 12-20 (Saturday morning):
- Execute Level 3: Draft Sections 4-5 (methodology, fellowship fit)
- Build ML demonstration notebook #1: Invariant violation detection using existing PROACTIVE validator logic
- Build contract: C-RSP-002 — Methodology & Technical Demo

HOUR 20-28 (Saturday midday):
- Create new GitHub repo per Repo Manifest
- Populate with curated artifacts from existing repos (do not copy everything; select 3-5 strongest files per category)
- Write unified README.md synthesizing all prior work into research substrate narrative
- Build contract: C-RSP-003 — Repository Population

HOUR 28-36 (Saturday evening):
- Build ML demonstration notebook #2: Scoring rubric automation (draw from ClarityAI scoring engine)
- Draft 2-day ML documentation plan as separate document (what would be done if accepted)
- Build contract: C-RSP-004 — ML Documentation Artifacts

HOUR 36-44 (Sunday morning):
- Final integration: Merge all proposal sections, apply voice calibration guide, insert V&T statements
- Portfolio integration: Update coreyalejandro.com with fellowship project page (or prepare content for it)
- Build contract: C-RSP-005 — Integration & Polish

HOUR 44-48 (Sunday midday):
- Final adversarial pass: Run Risk Register checklist one last time
- Submit via Constellation portal before April 26 deadline
- Build contract: C-RSP-006 — Submission & Verification

C-RSP BUILD CONTRACT TEMPLATE (use for each of the 6 above):
CONTRACT_ID: [C-RSP-XXX]
PURPOSE: [one sentence]
INPUTS: [specific files, prompts, or assets]
PROCESS: [step-by-step, deterministic]
OUTPUT: [specific artifact with acceptance criteria]
VERIFICATION: [how to know it's done correctly]
FAILURE_MODE: [what goes wrong and how to detect it]
TRUTH_STATUS: [VERIFIED / UNVERIFIED / PROTOTYPE]

FINAL METACOGNITIVE CHECK:
Before any build contract executes, ask:
1. Does this preserve the user's voice? (Check against Voice Guide)
2. Does this advance the research question? (Check against Proposal Architecture)
3. Does this close a gap identified in Level 2? (Check against Coverage Matrix)
4. Is this feasible in the time allocated? (Check against Trimmed Scope)
5. Does it include a V&T statement? (Invariant: No deliverable ships without epistemic boundary marking)

OUTPUT: Master Execution Script + 6 C-RSP Build Contracts + Self-Monitoring Rubric.
V&T STATEMENT REQUIRED.