fix: align fallback prompts with conservative validation

README.md

@@ -1,14 +1,14 @@
 # Context Compiler Directive Drafter
 
-Turn natural-language requests into candidate Context Compiler directives.
+Turn user messages into candidate Context Compiler directives with a conservative validation boundary.
 
-`context-compiler-directive-drafter` helps hosts translate user requests like:
+`context-compiler-directive-drafter` helps hosts validate directive candidates such as:
 
-> Please use Docker for container examples.
+> use docker
 
-into candidate directives, such as:
+and reject nearby unsafe or non-canonical inputs such as:
 
-> use docker
+> Please use Docker for container examples.
 
 This package drafts suggestions for the Context Compiler. Only `context-compiler` applies directives and updates state.
 
@@ -20,9 +20,9 @@ The drafter suggests candidate directives. context-compiler decides what to do w
 
 Use this package when you want to:
 
-- Translate user requests into safe, canonical directives.
+- Validate and draft safe, canonical directives from user messages.
 - Avoid accidental or unsafe state changes from ambiguous input.
-- Add a conservative natural-language-to-directive step before applying changes.
+- Add a conservative drafting step before applying changes.
 
 ## Installation
 
@@ -45,7 +45,7 @@ Draft and validate a candidate directive:
 ```python
 from context_compiler_directive_drafter import preprocess_heuristic, parse_preprocessor_output
 
-user_message = "Please use Docker for container examples."
+user_message = "use docker"
 result = preprocess_heuristic(user_message)
 
 candidate = parse_preprocessor_output(
@@ -78,7 +78,7 @@ Public interface:
 
 1. Run `preprocess_heuristic(message)`.
 2. If a candidate exists, validate it with `parse_preprocessor_output(...)`.
-3. If not valid, consider fallback drafting (e.g., LLM prompt).
+3. If not valid, you may consider fallback drafting (e.g., an LLM prompt).
 4. Always validate fallback output with `parse_preprocessor_output(..., source_input=message)`.
 5. If validation yields a directive, pass it to `context-compiler`.
 6. Otherwise, pass the original user input unchanged.
@@ -104,6 +104,8 @@ Use render_prompt(path, state) to load a template and fill it with the current c
 
 The rendered prompt can be sent to an LLM to attempt directive drafting when heuristic drafting does not produce a result.
 
+In `0.1.x`, the validated contract remains intentionally narrow: fallback output must still resolve to a canonical whole-message directive shape that passes `source_input` validation. Broader semantic rewriting is not part of the current package contract.
+
 Any model output should still be validated with parse_preprocessor_output(...) or validate_preprocessor_output(...) before it is used.
 
 ## Current Limits
@@ -113,6 +115,7 @@ This package is intentionally conservative. It abstains when input is:
 - Ambiguous, mixed-intent, or quoted.
 - Embedded in prose, markdown, or code.
 - Not matching a canonical directive form.
+- A non-canonical natural-language paraphrase of a directive.
 
 Boundary rules:
 

src/context_compiler_directive_drafter/output_validation.py

@@ -84,7 +84,15 @@ def _contains_multiple_candidate_directives(text: str) -> bool:
 
 
 def _is_safe_fallback_directive_rewrite(source_input: str, directive_output: str) -> bool:
-    """Reject fallback rewrites that bypass strict whole-message boundaries."""
+    """Allow only strict whole-message canonical fallback matches.
+
+    source_input validation is a boundary guard, not a second natural-language
+    parser. At the current 0.1 boundary, fallback output is accepted only when
+    the source is either the exact strict structured directive contract or a
+    whole-message canonical directive shape that normalizes to the same
+    directive. Unsafe wrappers, reported speech, mixed prose, and near-miss
+    canonicalizations remain rejected.
+    """
     source = re.sub(r"\s+", " ", source_input.strip().lower())
     directive = re.sub(r"\s+", " ", directive_output.strip().lower())
 

src/context_compiler_directive_drafter/prompts/default.txt

@@ -1,5 +1,5 @@
 # Default directive-drafter prompt.
-# Use this prompt for most instruction-following models.
+# Use this prompt for the strict 0.1.x canonical-only drafting boundary.
 
 You convert the latest user message into a single Context Compiler directive, or `<NO_DIRECTIVE>`.
 
@@ -94,18 +94,19 @@ Replacement rules:
 
 Conversion rules:
 
+* For this prompt version, emit a directive only when the whole message already expresses one allowed directive form.
 * Convert only clear, direct, explicit state-modification instructions into one canonical directive.
 * Use the user’s original wording for `<value>` and `<item>` where possible.
 * Do not expand meaning beyond what is explicitly stated.
 
 Examples:
-Message: don't use peanuts
-Output: prohibit peanuts
-
-Message: please use docker
+Message: use docker
 Output: use docker
 
-Message: switch from docker to podman
+Message: prohibit peanuts
+Output: prohibit peanuts
+
+Message: use podman instead of docker
 Output: use podman instead of docker
 
 Message: can you avoid peanuts?
@@ -114,6 +115,15 @@ Output: <NO_DIRECTIVE>
 Message: I prefer concise replies
 Output: <NO_DIRECTIVE>
 
+Message: please use docker
+Output: <NO_DIRECTIVE>
+
+Message: don't use peanuts
+Output: <NO_DIRECTIVE>
+
+Message: switch from docker to podman
+Output: <NO_DIRECTIVE>
+
 Message: set premise to concise replies
 Output: <NO_DIRECTIVE>
 

src/context_compiler_directive_drafter/prompts/llama.txt

@@ -1,5 +1,6 @@
 # Directive-drafter prompt for Llama-family models.
-# Use this when the model shows weaker adherence to strict output constraints.
+# Use this for the strict 0.1.x canonical-only drafting boundary when the model
+# needs extra output-shape reinforcement.
 
 You convert the latest user message into a single Context Compiler directive, or `<NO_DIRECTIVE>`.
 
@@ -82,18 +83,19 @@ Replacement rules:
 
 Conversion rules:
 
+* For this prompt version, emit a directive only when the whole message already expresses one allowed directive form.
 * Convert only clear, explicit state-modification instructions into one canonical directive.
 * Use the user’s original wording for `<value>` and `<item>` where possible.
 * Do not reinterpret or expand meaning beyond what is explicitly stated.
 
 Examples:
-Message: don't use peanuts
-Output: prohibit peanuts
-
-Message: please use docker
+Message: use docker
 Output: use docker
 
-Message: switch from docker to podman
+Message: prohibit peanuts
+Output: prohibit peanuts
+
+Message: use podman instead of docker
 Output: use podman instead of docker
 
 Message: can you avoid peanuts?
@@ -102,6 +104,15 @@ Output: <NO_DIRECTIVE>
 Message: I prefer concise replies
 Output: <NO_DIRECTIVE>
 
+Message: please use docker
+Output: <NO_DIRECTIVE>
+
+Message: don't use peanuts
+Output: <NO_DIRECTIVE>
+
+Message: switch from docker to podman
+Output: <NO_DIRECTIVE>
+
 Message: set premise to concise replies
 Output: <NO_DIRECTIVE>