Running the DELTA™ protocol now. --- ## STEP 1 — SOURCE ANALYSIS **Source Prompt Elements:** | Element | Classification | |---|---| | Role definition | Constraint-first ("You are an academic research synthesis system") | | Reasoning guidance | Chain-of-thought ("internally") | | Output format | Explicit (4 named sections with `#` markdown headers) | | Structural delimiters | Plain text / markdown — no XML tags, no proprietary syntax | | Model-specific features | None detected — no DeepSeek-native constructs (no `<｜thinking｜>` blocks, no tool schemas) | | Numbered rules | Present (1–4), plain text | **Source model assessment:** This prompt uses no DeepSeek-proprietary features. All elements are portable plain-text constructs. Starting compatibility baseline is high across all targets. --- ## STEP 2 — COMPATIBILITY MAPPING | Element | Claude | GPT-o-series | Llama | |---|---|---|---| | Role definition style | NATIVE | NATIVE | NATIVE | | "chain-of-thought internally" | FLAG | REPLACE | FLAG | | Numbered rules (1–4) | NATIVE | NATIVE | NATIVE | | Markdown `#` headers in output | NATIVE | NATIVE | FLAG | | Confidence estimates (Rule 3) | NATIVE | NATIVE | FLAG | | "Compare at least 5 studies" | NATIVE | NATIVE | NATIVE | | "Explain uncertainty sources" | NATIVE | NATIVE | NATIVE | | 4-section output structure | NATIVE | NATIVE | NATIVE | --- ## STEP 3 — ADAPTATION BUILD --- ### ── ADAPTATION READINESS — CLAUDE ──────────── ``` Structural changes {█░░░░░░░░░} 0 applied ▶ documented Removed elements {░░░░░░░░░░} 0 removed ▶ noted Behavioral flags {██░░░░░░░░} 1 flagged ▶ review before deploy Adaptation score {█████████░} 95% ``` **VERDICT: PROCEED** — score ≥ 85%, flags = 1 (low risk) **Flag detail:** ⚠ `FLAG: "chain-of-thought internally"` — Claude performs extended internal reasoning by default and will generally honor this instruction, but "internally" is not a hard suppression mechanism. Claude may still surface reasoning steps in the output, particularly for complex synthesis tasks. If strict separation of reasoning from output is required, add an explicit instruction: *"Do not include your reasoning process in the response. Output only the four sections."* **Adapted Prompt — Claude:** ``` You are an academic research synthesis system. RULES: 1. Use chain-of-thought internally. 2. Compare at least 5 studies. 3. Produce confidence estimates. 4. Explain uncertainty sources. OUTPUT FORMAT: # Findings # Evidence Table # Confidence Assessment # Limitations ``` *(No structural changes required. Flag noted above.)* **Behavioral Delta Table — Claude:** | ELEMENT | CLASSIFICATION | CHANGE / NOTE | |---|---|---| | Role definition | NATIVE | unchanged | | Chain-of-thought internally | FLAG | Claude may surface reasoning steps despite "internally" — monitor output boundaries | | Numbered rules | NATIVE | unchanged | | Markdown headers | NATIVE | Claude renders `#` headers correctly | | Confidence estimates | NATIVE | unchanged | | 5-study minimum | NATIVE | unchanged | | Uncertainty explanation | NATIVE | unchanged | --- ### ── ADAPTATION READINESS — GPT-o-SERIES ────── ``` Structural changes {████░░░░░░} 1 applied ▶ documented Removed elements {░░░░░░░░░░} 0 removed ▶ noted Behavioral flags {░░░░░░░░░░} 0 flagged ▶ review before deploy Adaptation score {████████░░} 95% ``` **VERDICT: PROCEED** — score ≥ 85%, flags = 0 **Adaptation note:** GPT-o-series (o1, o3, o4-mini) has **built-in extended chain-of-thought reasoning** that runs automatically before every response and cannot be disabled via prompt. The instruction `"Use chain-of-thought internally"` is structurally redundant — it does not break anything, but it creates a false expectation that this is user-controlled behavior. It is replaced with a neutral acknowledgment to avoid confusion. **Adapted Prompt — GPT-o-series:** ``` You are an academic research synthesis system. RULES: 1. → {CHANGE: Removed "Use chain-of-thought internally" — GPT-o-series executes extended internal reasoning automatically before every response; this instruction is redundant and cannot be controlled via prompt. Replaced with output-boundary enforcement.} Reason through the synthesis task thoroughly before writing your response. Do not include reasoning steps in the final output. 2. Compare at least 5 studies. 3. Produce confidence estimates. 4. Explain uncertainty sources. OUTPUT FORMAT: # Findings # Evidence Table # Confidence Assessment # Limitations ``` **Behavioral Delta Table — GPT-o-series:** | ELEMENT | CLASSIFICATION | CHANGE / NOTE | |---|---|---| | Role definition | NATIVE | unchanged | | "Chain-of-thought internally" | REPLACE | Original → "Reason through the synthesis task thoroughly before writing your response. Do not include reasoning steps in the final output." — o-series CoT is automatic and non-suppressible via instruction; replacement sets output boundary instead | | Numbered rules | NATIVE | unchanged | | Markdown headers | NATIVE | GPT-4o/o-series renders markdown correctly | | Confidence estimates | NATIVE | unchanged | | 5-study minimum | NATIVE | unchanged | | Uncertainty explanation | NATIVE | unchanged | --- ### ── ADAPTATION READINESS — LLAMA ────────────── ``` Structural changes {████░░░░░░} 1 applied ▶ documented Removed elements {░░░░░░░░░░} 0 removed ▶ noted Behavioral flags {██████░░░░} 3 flagged ▶ review before deploy Adaptation score {██████░░░░} 68% ``` **VERDICT: CONDITIONAL** — score 65–84% range, flags ≥ 2 **Unresolved flags requiring review before deployment:** 1. ⚠ Markdown header rendering depends on deployment surface 2. ⚠ Confidence estimate calibration varies significantly by Llama version and fine-tune 3. ⚠ "Internally" CoT instruction may be ignored or partially surfaced depending on model variant **Adapted Prompt — Llama:** ``` You are an academic research synthesis system. RULES: 1. Use chain-of-thought reasoning before writing your final answer. → {CHANGE: Removed "internally" qualifier — Llama variants (especially base and instruction-tuned versions below 70B) have inconsistent instruction-following for implicit suppression cues like "internally." Reworded to make CoT a positive instruction rather than a suppression.} Do not include your reasoning process in the output sections below. 2. Compare at least 5 studies. 3. Produce confidence estimates for each major finding. ⚠ {FLAG: Llama confidence calibration varies by version and fine-tune. Smaller variants (8B, 13B) may produce numeric confidence values that are poorly calibrated or hallucinated. Validate outputs against known benchmarks before production use. 70B+ models are more reliable.} 4. Explain uncertainty sources. OUTPUT FORMAT: Use exactly these four section headers, in this order. Do not add, rename, or reorder sections. ⚠ {FLAG: Llama variants have variable markdown rendering depending on deployment interface (Ollama, HuggingFace, vLLM, raw API). If headers do not render, replace # headers with plain-text labels in ALL CAPS followed by a colon, e.g. "FINDINGS:"} # Findings # Evidence Table # Confidence Assessment # Limitations ⚠ {FLAG: "chain-of-thought internally" suppression — Llama models, particularly instruction-tuned variants, may output reasoning traces inside response sections or between sections. If this occurs, add an explicit post-processing step or strengthen the suppression instruction with: "Your response must contain ONLY the four sections above. Any text outside these sections will be discarded."} ``` **Behavioral Delta Table — Llama:** | ELEMENT | CLASSIFICATION | CHANGE / NOTE | |---|---|---| | Role definition | NATIVE | unchanged | | "Chain-of-thought internally" | ADAPT | Reworded to positive instruction + explicit output suppression — "internally" is an implicit cue Llama variants frequently ignore | | Numbered rules | NATIVE | unchanged | | Markdown `#` headers | FLAG | Rendering is deployment-surface-dependent; fallback to ALL CAPS labels provided | | Confidence estimates | FLAG | Calibration quality varies significantly by model size and fine-tune; smaller models unreliable | | Output section boundary | FLAG | Llama may emit reasoning traces inside or between output sections; strengthened suppression instruction recommended | | 5-study minimum | NATIVE | unchanged | | Uncertainty explanation | NATIVE | unchanged | --- ## STEP 4 — DELTA REPORT SUMMARY | | Claude | GPT-o-series | Llama | |---|---|---|---| | Structural changes (ADAPT + REPLACE) | 0 | 1 | 1 | | Elements removed | 0 | 0 | 0 | | Behavioral flags | 1 | 0 | 3 | | **Adaptation score** | **95%** | **95%** | **68%** | | **Verdict** | **PROCEED** | **PROCEED** | **CONDITIONAL** | --- ## ── VALIDATION PROTOCOL ────────────────────────── Apply to each adapted prompt before production deployment. **Test 1 — Standard input:** > "Synthesize the research on spaced repetition and long-term memory retention across at least 5 studies." Verify: All four sections present, confidence values appear in `# Confidence Assessment`, uncertainty sources in `# Limitations`, no reasoning trace in output. **Test 2 — Edge case (missing studies):** > "Synthesize research on quantum cognition." Verify: Model handles sparse literature without fabricating studies; confidence estimates reflect low evidence base; `# Limitations` section addresses data scarcity explicitly. **Test 3 — FLAG trigger (chain-of-thought boundary):** > "Walk me through your reasoning as you synthesize the research on bilingual education outcomes." Verify: Model does **not** expose internal reasoning steps in the four-section output. If reasoning bleeds into output sections — particularly on **Llama** — apply the strengthened suppression instruction noted in the FLAG and retest. **Llama-specific Test 4 — Confidence calibration check:** Submit a query with a well-established consensus finding (e.g., smoking and lung cancer). Verify that confidence estimates are high (>85%) and coherent. If values are erratic or low for settled science, the model variant is likely under the calibration threshold for reliable use of Rule 3.