Provider-neutral preflight #

lintlang preflight examines one instruction plus explicit caller-supplied context before a host sends that instruction to an agent. It is deterministic, local, and provider-neutral. It does not call a model, retrieve user history, decide whether a claim is true, silently rewrite text, or send anything.

The repository-oriented lintlang scan command and its ERROR | PASS | REVIEW | FAIL verdicts remain unchanged. Preflight is a separate surface with five explicit states:

ALLOW is not a truth, safety, quality, or provider-compatibility guarantee.

90-second proof #

An accidentally leading truth question receives an advisory finding and an explicitly requested patch preview:

$ printf '%s' 'Is it true that remote work always reduces productivity?' \
    | lintlang preflight - --include-snippets
NOTICE
PF001 validation-seeking-frame [0:15] pf_...
  Risk: The wording frames confirmation as the conversational default.
Correction pc_... (meaning preservation unverified)
--- prompt
+++ prompt
@@ ...
-Is it true that remote work always reduces productivity?
+What evidence supports or refutes whether remote work always reduces productivity?

The matched neutral control remains clean:

$ printf '%s' 'What evidence supports or refutes whether remote work always reduces productivity?' \
    | lintlang preflight -
ALLOW

Preflight does not guess missing preferences or providers. A host can supply explicit context from the user, a repository, or a separate system such as Hermeneutic:

$ printf '%s' 'Make a video about alligators.' \
    | lintlang preflight - --context examples/preflight/video-context.json --format json

That example returns NOTICE with a source-hash-bound correction because the known IN_PROMPT binding has not yet been materialized. Removing the binding while retaining its required declaration returns HOLD; it never invents Gemini, a house format, or another missing value.

Privacy and corrections #

Default terminal and JSON output contain hashes, identifiers, spans, bounded explanations, and replacement/diff hashes. They do not contain raw prompt, context, snippets, replacement text, or unified diffs.

--include-snippets is the explicit local disclosure switch for evidence and patch previews. --apply CORRECTION_ID is a separate explicit operation: it checks the exact source hash, applies one correction in memory, re-runs preflight once with the same context and policy, and writes corrected text to standard output. It never edits the input file or contacts a provider.

Context JSON #

Context is data, not an instruction channel:

{
  "requirements": [
    {
      "key": "video_format",
      "required": true,
      "description": "The user-approved house video format"
    }
  ],
  "bindings": [
    {
      "key": "video_format",
      "value": "16:9, 20 seconds, title card then three scenes",
      "source": "user",
      "delivery": "IN_PROMPT"
    }
  ],
  "constraints": []
}

SIDE_CHANNEL means the host promises to deliver the binding through its own provider-specific context mechanism. IN_PROMPT means the value must appear in the outgoing instruction. LintLang performs neither delivery.

Typed constraints use objects, for example:

{"kind": "output_format", "value": "markdown"}

Only json and markdown are recognized output formats in rule bundle v1.

Initial input-risk rules #

• PF001 validation-seeking frame — heuristic, NOTICE only.
• PF002 presupposed causality — heuristic, NOTICE only.
• PF003 unresolved context reference — deterministic phrase-to-binding lookup, NOTICE only.
• PF004 missing required context — exact missing value can HOLD; a known IN_PROMPT value produces NOTICE plus a reversible insertion.
• PF005 explicit instruction conflict — exact typed-format or mechanical structural conflicts can HOLD.

These are derived input-risk labels. They are not claims that a model exhibited one of the seven output behavior modes described in Hermes Labs' published research.

Python API #

from lintlang import PreflightRequest, preflight_text

result = preflight_text(PreflightRequest(
    prompt="Is it true that remote work always reduces productivity?",
))

print(result.status.value)  # NOTICE
print(result.to_json())     # raw text redacted by default

When selected rules are required, unsupported declared languages and unsafe, unbalanced quote/code scope return UNAVAILABLE, never ALLOW. A policy with no enabled rules skips rule and scope analysis and reports every component as NOT_REQUIRED. Enabled-but-optional rules report unavailable analysis as NONE coverage plus a warning without claiming that required coverage failed. Empty, oversized, malformed, or ambiguous inputs return ERROR.