Conflicts — Signal
Per-category conflict detail. Summary + index: CONFLICTS.md.
Critical 3 — S9 H/S V7
Type: Hard vs Soft
S9 embeds principles in the prompt. The model applies them through language reasoning — probabilistic, can be overridden by adversarial prompting, cannot be audited with certainty. V7 externalises rules in a policy engine independent of the LLM — deterministic for defined violations, survives prompt manipulation, produces an audit record.
They are not alternatives. They are layered enforcement:
S9 (Constitutional Framing) — soft, broad, in-prompt
"I should not reveal confidential data"
$\to$ model usually follows; can be manipulated by injection
V7 (AgentSpec / Declarative Governance) — hard, specific, external
PROHIBIT: tool_call.name == "send_email" AND context.contains(classified_data)
$\to$ enforced at runtime regardless of what model "thinks"
Resolution rule:
- S9 for: values, style, judgment calls, broad ethical principles — anything requiring contextual interpretation
- V7 for: specific, enumerable prohibitions and obligations — anything requiring deterministic enforcement
- Always use both in safety-critical systems; S9 catches the cases V7 didn't anticipate; V7 catches the cases S9 was manipulated into allowing
Critical error: Using S9 alone and claiming the system is "aligned." S9 is probabilistic; call it what it is.
Connection B — S2 $\sim$ prefix cache
Type: Composability Tension ($\sim$)
Dynamic S2 (Retrieval-Augmented Few-Shot variant) changes the token sequence of the few-shot block on every call. This does not only forfeit S2's own cache entry — it invalidates the cache for the entire prefix that precedes it: S3 Persona, S5 Constraint Framing, S6 Output Template, S9 Constitutional Framing. Any stable content placed before the dynamic S2 block cannot be cached if S2 changes.
The economic cost is larger than it appears: if the stable prefix (S3+S5+S6+S9) is 2,000 tokens and dynamic S2 is inserted in the middle of it, all 2,000 tokens of stable content re-prefill at full cost on every call.
Resolution: If dynamic S2 is required, place it at the END of the prompt — after all stable content. This preserves the stable prefix cache for the S3/S5/S6/S9 block while still allowing the examples to vary.
Signal vs Signal
| Pattern A | Conflict Type | Pattern B | Resolution |
|---|---|---|---|
| S1 (Zero-Shot) | $\uparrow$ | S2 (Few-Shot) | S1 is the default; add S2 when output format is inconsistent. S2 costs 3-5× more tokens. |
| S3 (Persona) | $\sim$ | S5 (Constraint Framing) | Persona may imply latitude that constraints prohibit. Add explicit "constraints override persona." |
| S3 (Persona) | $\sim$ | S9 (Constitutional Framing) | Persona implies identity; constitution implies values. Conflict when persona's implied expertise contradicts constitutional safety constraints. Constitution wins. |
| S4 (Instruction Decomposition) | $\uparrow$ | O2 (Prompt Chaining) | S4 puts all steps in one prompt; O2 distributes across calls. S4 is cheaper but loses inter-step inspection. |
| S6 (Output Template) | $\uparrow$ | Structured Output API | Structured output API (JSON mode) is strictly better when available. S6 free-text templates only when API not available. |
| R17 (Self-Consistency) | $\oplus$ | H3 (Entropy Curiosity) | See CRITICAL 4. Never apply simultaneously. |
| S8 (Meta-Prompt) | $\to$ | R17 or V15 | S8 requires an evaluation signal to select between generated prompts. Without R17 or V15, S8 cannot function. |
| S9 (Constitutional Framing) | H/S | V7 (AgentSpec) | See CRITICAL 3. Complementary; S9 soft/broad, V7 hard/specific. |
Signal vs Reasoning
| Pattern A | Conflict Type | Pattern B | Resolution |
|---|---|---|---|
| S2 (Few-Shot) | $\sim$ | R17 (Self-Consistency) | S2 shapes what the model produces; R17 samples multiple versions and votes. They compose: S2 sets format, R17 improves reliability. Ensure S2 examples don't bias R17 toward a single answer style. |
| S4 (Instruction Decomposition) | $\uparrow$ | R3 (Plan-and-Solve) | S4 is a prompt-level step list; R3 is an agent-level planning cycle with separate plan and execution calls. R3 is more powerful but costs more. |
| S9 (Constitutional Framing) | $\sim$ | R7 (Reflexion) | Reflexion critiques outputs; constitution critiques against principles. If both are active, ensure they don't generate contradictory critique: R7 might say "be more detailed" while S9 says "be more concise." Make priorities explicit. |