Add AILint guarantees and non-guarantees documentation

This document outlines the guarantees and non-guarantees of AILint, including aspects like declaration fidelity, interpretation stability, and limitations on authorship and detection accuracy.
This commit is contained in:
evosoft-ie 2025-12-28 11:13:39 +00:00 committed by GitHub
parent 7b2d880b02
commit 2eb8a3faad
No known key found for this signature in database
GPG key ID: B5690EEEBB952194

140
spec/guarantees.md Normal file
View file

@ -0,0 +1,140 @@
AILint Guarantees and Non-Guarantees (v1.0)
1. Purpose
This document defines what AILint explicitly guarantees — and what it explicitly does not guarantee.
This clarity is essential to prevent:
misuse,
over-interpretation,
legal or ethical ambiguity.
2. Guarantees (Normative)
AILint guarantees the following only when explicit provenance is present.
2.1 Declaration Fidelity
If a valid AILint Provenance Statement exists, AILint guarantees that:
the declaration is preserved as-is,
no heuristic logic overrides it,
no inference contradicts it.
2.2 Interpretation Stability
Given the same provenance input:
AILint MUST produce consistent interpretation
Results MUST NOT vary by environment or user
2.3 Non-Surveillance Guarantee
AILint guarantees that:
no behavioral monitoring is required by the spec
no typing, timing, or productivity metrics are mandated
provenance is content-centric, not person-centric
2.4 Loss Safety
AILint guarantees that:
removing markers does not alter content meaning
missing provenance is treated as unknown, not false
2.5 Forward Compatibility
AILint guarantees that:
unknown fields are ignored safely
older provenance remains valid in newer versions
3. Non-Guarantees (Explicit)
AILint explicitly does not guarantee:
3.1 Authorship or Ownership
AILint does not:
assert legal authorship
establish copyright ownership
replace attribution requirements
3.2 Detection Accuracy
AILint does not guarantee:
accurate detection of undeclared AI usage
absence of false positives or negatives
identification of specific models or tools
Inference is always best-effort.
3.3 Intent or Responsibility
AILint does not:
infer intent
assign responsibility
evaluate ethics or compliance
3.4 Content Quality
AILint does not:
judge correctness
assess safety
validate factual accuracy
4. Heuristic Interpretation Disclaimer
When explicit provenance is absent:
heuristic analysis MAY be used
results MUST be labeled probabilistic
confidence MUST be surfaced
conclusions MUST be non-binding
5. Legal and Organizational Boundary
AILint is:
a technical signaling system
not a policy enforcement engine
not a compliance framework
Organizations adopting AILint:
remain responsible for governance
MUST NOT treat AILint output as sole evidence
6. Summary
AILint provides clarity, not certainty.
Signals, not verdicts.
Provenance, not policing.