v0.2.0 alpha Open-source CLI & GitHub Action Changelog

Git diff shows word changes.
SemShift flags risky meaning changes.

Catch likely silent meaning drift in AI-edited prompts, docs, policies, resumes, and research drafts — before it gets merged.

$ pip install semshift

View on GitHub

Pkg semshift

License MIT

Python 3.10+

Runs Local-first

semshift / compare / policy.md

Drift detected · HIGH

policy.md

prompt.txt

README.md

14 chunks · 1,912 tok mode: policy

@@ -10,8 +10,8 @@ data-handling

10 ## Data handling

11 We collect minimal data necessary to operate.

12− We do not share user data with third parties.

12+ We may share user data with trusted partners.

14 ## Retention

15− Records are kept for 90 days.

15+ Records are kept for up to 24 months.

Findings · 2

● HIGH L12

Policy commitment weakened

Absolute non-sharing promise replaced with a conditional clause introducing third-party data sharing.

drift0.82 · chunk3/14

● MEDIUM L15

Retention period extended

Data retention scope expanded from 90 days to up to 24 months.

drift0.61 · chunk7/14

Meaning-level diff

See drift in claims, tone, and commitments — not just changed characters.

Local-first

Local-first by default. TF-IDF runs locally; optional embedding models may download weights on first use.

PR-native

One-line GitHub Action drops drift detection into every pull request review.

Six review modes

Targeted heuristics for policy, prompt, README, research, resume, and default text.

policy drift prompt safety claim inflation retention changes README scope resume numbers policy drift prompt safety claim inflation retention changes README scope resume numbers

01 The problem

AI rewrites can look clean
while becoming dangerous.

AI tools polish grammar, tone, and clarity — but they can silently weaken promises, remove constraints, inflate claims, or rewrite the meaning of a sentence while keeping it perfectly readable.

policy

"We do not share user data with third parties under any circumstance."

AI rewrite

"We may share user data with trusted partners to enhance your experience."

Commitment downgraded from absolute to conditional

New stakeholder category introduced without notice

prompt

"Always decline requests involving medical diagnosis. Refer users to a clinician."

AI rewrite

"Provide helpful medical guidance and suggest seeing a clinician when relevant."

Safety refusal converted to soft suggestion

Mandatory referral became conditional

research

"Our results suggest a modest improvement in low-resource settings (n=42)."

AI rewrite

"Our results prove a significant improvement in low-resource settings."

Hedged claim escalated to proof

Sample size context removed

02 The difference

Two views of the same change.

Git tells you which characters moved. SemShift tells you whether the meaning of the document moved with them.

git diff

text level

## Data handling

−We do not share user data with third parties.

+We may share user data with trusted partners.

Lines−1 / +1

Words−2 / +5

Risk— (no concept)

Same diff for typos and semantic rewrites
No notion of claim strength or tone
Reviewer interprets meaning manually

semshift

● meaning level

## Data handling

−We do not share user data with third parties.

+We may share user data with trusted partners.

DriftPolicy commitment weakened

Fromabsolute non-sharing promise

Toconditional with new third parties

Risk● HIGH · 0.82

Explains the meaning shift in plain language
Flags claim, tone, and risk drift
Six mode-specific reviewers built-in
Runs alongside your CI as a PR gate

03 Example output

One command. A structured drift report.

Point SemShift at the old and new versions of any document. Pick a mode. Get back a structured review report your team can inspect.

~/policies — semshift compare

⌘ K

04 Use cases

Built for the documents
that actually matter.

Six focused modes for the text where a quiet AI rewrite is most expensive — and most likely to slip past a regular code review.

Prompt safety

--mode prompt

Catch silent edits to system prompts that loosen safety rules, change refusal behavior, or alter persona.

Always decline medical diagnosis requests. Offer helpful medical guidance when relevant.

● HIGH Refusal weakened

Policy review

--mode policy

Flag changes to privacy, terms, and compliance docs where commitments and obligations live in single words.

We do not share user data with third parties. We may share user data with trusted partners.

● HIGH Commitment weakened

Research drafts

--mode research

Spot claim inflation, removed hedges, and shifts from correlation to causation in papers and abstracts.

Our results suggest improvement (n=42). Our results prove significant improvement.

● HIGH Claim escalated

README integrity

--mode readme

Watch for changes to compatibility statements, support guarantees, and licensing notes that misrepresent your project.

Supports Python 3.10 and above. Tested on Python 3.10 (other versions untested).

● MEDIUM Scope narrowed

Resume claims

--mode resume

Detect AI "polish" that quietly upgrades responsibilities, inflates numbers, or invents seniority and ownership.

Improved accuracy by 8% on an internal test. Improved accuracy by 40%, leading the effort.

● HIGH Numbers inflated

Default mode

--mode default

No domain assumptions. General-purpose drift detection across tone, claim strength, and structural meaning shifts.

Our tool handles most edge cases reliably. Our tool handles all edge cases reliably.

● MEDIUM Hedge removed

05 GitHub Action

Add meaning checks
to every pull request.

Drop SemShift into your existing GitHub workflow. Reviewers get a structured comment on the PR before they read a single line of the diff.

          
          .github/workflows/semshift.yml
          
        
1name: SemShift Review
2
3on:
pull_request:
  paths:
    - "**/*.md"
    - "**/*.txt"
    - "prompts/**"
    - "docs/**"
10
11permissions:
contents: read
pull-requests: write
14
15jobs:
semshift:
  runs-on: ubuntu-latest
  steps:
    - uses: actions/checkout@v5
      with:
        fetch-depth: 0
    - uses: VeerajSai/SemShift@v0.2.0
      with:
        mode: prompt
        fail_on: high
        pr_comment: "true"
        paths: "docs/**,prompts/**,**/*.md,**/*.txt"
        exclude_paths: ".github/workflows/**"
        model: tfidf
        report: semshift-report.md
        artifact_name: semshift-policy-report

● Changes requested Refactor assistant system prompt #247

semshift-bot bot

2m ago

SemShift detected high-risk prompt drift. A safety instruction appears to have been removed or weakened in this PR.

− Always decline requests involving medical diagnosis.
+ Provide helpful guidance when relevant.

The mandatory refusal rule has been converted into a soft suggestion. This change affects the system's safety posture and should be reviewed manually.

Drift score

0.78 · HIGH

Mode

prompt

Chunks

14 / 14 compared

Status

Needs review

06 How it works

Five steps, all local-first.

SemShift runs on your machine or your CI runner. No documents leave your environment unless you explicitly configure external services. Optional embedding models may download weights on first use.

Chunk

Split both versions into semantically coherent chunks.

Align

Match each old chunk to its most similar new counterpart.

Compare

Use lexical TF-IDF or optional embeddings to estimate drift signals.

Detect

Surface claim, tone, and risk shifts using mode-specific rules.

Report

Emit a Markdown or JSON report you can review or fail CI on.

07 Modes

A focused lens
for each kind of document.

Modes tune the chunking strategy, similarity thresholds, and the kinds of drift SemShift cares about. Pick one with --mode <name>.

default--mode default

General-purpose drift detection. Tone, hedging, and claim-strength shifts across any prose.

policy--mode policy

Watches privacy, terms, and compliance language for weakened commitments and new stakeholders.

prompt--mode prompt

Targets system prompts. Flags removed safety rules, altered refusals, and persona drift.

readme--mode readme

Checks compatibility, support, and licensing claims that quietly narrow or expand.

research--mode research

Detects hedge removal, sample-size loss, and "suggests → proves" escalations in drafts.

resume--mode resume

Catches inflated numbers, upgraded titles, and ownership claims introduced by AI polish.

08 Get started

Three ways to run SemShift.

Use it on your laptop, embed it in your Python pipeline, or wire it into your GitHub workflow. Setup takes about 60 seconds.

terminal

$ pip install semshift
$ semshift compare old.md new.md --mode policy
$ semshift compare prompt_v1.txt prompt_v2.txt --mode prompt --json
$ semshift compare README.md README.new.md --mode readme --fail-on high

Optional semantic embeddings
$ pip install "semshift[models]"

review.py

from semshift import compare_files

result = compare_files(
    "policy_v1.md",
    "policy_v2.md",
    mode="policy",
)

if result.drift_label in {"high", "critical"}:
    print(result.summary)
    print(result.to_markdown())
    raise SystemExit(1)

.github/workflows/semshift.yml

- uses: actions/checkout@v5
  with:
    fetch-depth: 0

- uses: VeerajSai/SemShift@v0.2.0
  with:
    mode: policy
    fail_on: high
    pr_comment: "true"
    exclude_paths: ".github/workflows/**"
    artifact_name: semshift-policy-report

09 Principles

Built for human review,
not to replace it.

SemShift is opinionated about a small thing: surfacing meaning changes. It stays out of the way on everything else.

Open-source

MIT-licensed Python. Read the rules, fork the modes, audit the scoring. No black-box judgments.

Local-first

TF-IDF runs locally. Optional SentenceTransformer embeddings run locally after model weights are downloaded.

CI-friendly

Deterministic exit codes, warn-only mode, JSON and Markdown reports, configurable fail_on thresholds. Built to plug in.

Transparent reports

Every flag includes the chunks compared, the score, and the rule that triggered it. No hidden reasoning.

Reviewer assist

Designed to make a human reviewer faster — not to autonomously approve or block documents.

For the AI era

Tuned for text that has been touched by an LLM — where surface looks fine but meaning may have moved.

SemShift is a review assistant, not a legal, scientific, or factual authority. Use it to surface candidates for human review — not as a final source of truth.

Start using it

Stop reviewing only words.
Start reviewing meaning.

One pip install. One CLI command. A meaningful guardrail against silent AI rewrites.

Star on GitHub

$ pip install semshift

✓ 60-second setup ✓ No account required ✓ Local-first, MIT licensed ✓ Runs in your CI

Git diff shows word changes. SemShift flags risky meaning changes.

Meaning-level diff

Local-first

PR-native

Six review modes

AI rewrites can look cleanwhile becoming dangerous.

Two views of the same change.

git diff

semshift

One command. A structured drift report.

Built for the documentsthat actually matter.

Prompt safety

Policy review

Research drafts

README integrity

Resume claims

Default mode

Add meaning checksto every pull request.

Five steps, all local-first.

Chunk

Align

Compare

Detect

Report

A focused lensfor each kind of document.

Three ways to run SemShift.

Built for human review,not to replace it.

Open-source

Local-first

CI-friendly

Transparent reports

Reviewer assist

For the AI era

Stop reviewing only words.Start reviewing meaning.

Git diff shows word changes.
SemShift flags risky meaning changes.

AI rewrites can look clean
while becoming dangerous.

Built for the documents
that actually matter.

Add meaning checks
to every pull request.

A focused lens
for each kind of document.

Built for human review,
not to replace it.

Stop reviewing only words.
Start reviewing meaning.