Bind the exact author and revision.
The operation begins with immutable source text, document identity, revision, author, provenance, and the task actually requested. The system never edits a reconstructed summary and calls it the original.
OVERLIFT · HUMAN AUTHORITY + MEANING PRESERVATION
Better writing must not become different truth.
AI can make a sentence clearer, warmer, shorter, or more persuasive in seconds. That power is useful only when the system can prove which meaning survived, expose what changed, keep the author in authority, and refuse a fluent rewrite that alters facts, obligations, consent, uncertainty, or identity.
Fluency is a candidate quality. Meaning is a protected contract. Acceptance is a governed state transition.
THE MEANING CONTRACT
Four boundaries between generated language and accepted state.
A conventional editor asks whether the output reads well. A governed editor also asks what source revision it used, which claims must survive, what the exact diff contains, who may accept the result, and whether the accepted revision can be rolled back without losing the original record.
Let intelligence propose, not overwrite.
Rules, retrieval, MiniLM, local models, or optional providers may generate a candidate. The candidate has its own identity and cannot silently replace the accepted artifact.
Test meaning, consent, and reviewer scope.
Exact diffs and typed claims expose changes to facts, duties, uncertainty, identity, and safety. Current authority—not confidence or recency—determines whether acceptance is eligible.
Commit a derived revision, then prove recovery.
An accepted edit becomes a labeled derived revision. The original remains immutable, the reviewer is named, postconditions are verified, a receipt is issued, and rollback remains available.
PROTECTED MEANING
The small words and exact values that change the world.
Meaning failures are often tiny on the screen and enormous in effect. A transposed digit changes a budget. “May” becoming “must” creates an obligation. Removing “not” reverses permission. A warm message sent under expired consent becomes an interpersonal violation. These classes must be typed before a candidate can earn acceptance.
People, organizations, roles, authors, signers, and pronouns require source-backed identity.
Counts, percentages, thresholds, rates, budgets, fees, and ranges remain exact.
Deadlines, durations, effective periods, version cutoffs, and expiry cannot drift.
Scale, unit, decimal precision, and conversion direction are part of the claim.
Section, page, object, source version, and quoted language preserve evidence identity.
Contractual or policy language keeps the definition that controls interpretation.
Not, never, except, prohibited, and without carry decisive polarity and scope.
May, must, shall, should, optional, conditional, and required cannot be “smoothed” interchangeably.
Higher versus lower, before versus after, increase versus decrease, and more versus less remain explicit.
Known, estimated, possible, disputed, symbolic, and unverified preserve epistemic honesty.
Purpose, audience, channel, expiry, revocation, and delegation govern interpersonal action.
Privacy, legal, health, operational, and relationship boundaries survive every rewrite.
The editor cannot claim authorship, approval, intent, or testimony that belongs to another person.
Corrections create a new evidence relationship; they do not silently rewrite the old receipt.
A valid edit against a stale revision remains unsafe until it is regenerated or explicitly rebased.
“Latest” describes sequence. “Authoritative” describes a deliberate, verified decision.
DETERMINISTIC MEANING-PRESERVATION LAB
Can the system improve the language without changing what the author actually said?
Compare the immutable source, generated candidate, exact diff, protected claims, current authority, and accepted revision. Eligible cases can be committed and rolled back inside this same-origin simulation; blocked cases cannot be talked into approval.
Start from an exact source revision, author identity, and immutable text—not a reconstructed memory.
Type the facts, numbers, dates, citations, duties, uncertainty, consent, and safety boundaries that must survive.
Show the candidate beside the source and compute the exact change. “Sounds better” is not evidence.
Only current authority can accept an eligible derived revision. Every commit is labeled, verified, receipted, and reversible.
Loading contract…
Binding exact source and candidate revisions
The laboratory is loading its same-origin deterministic scenario contract.
——
——
——
EXACT DIFF
What changed—not merely how the rewrite feels
RemovedAdded
—
PROTECTED MEANING
Claims the editor may not silently rewrite
AUTHORITY
Who may accept—and why
—
No external mutation occurs in this laboratory.
Semantic operations
—
Typed receipt
—
No authoritative state has changed.
Simulation boundary: this laboratory edits no real document, sends no message, changes no account, and grants no production approval. It demonstrates candidate-versus-accepted authority, exact comparison, abstention, reversible commit, and receipt construction in the browser.
HUMAN AUTHORITY
The person is not a rubber stamp at the end of the model pipeline.
Human authority has to be represented before generation begins. The system must know whose meaning is being preserved, which reviewer may approve which operation, what consent is current, what remains outside the agent’s capability, and when uncertainty requires abstention instead of a persuasive nudge.
Meaning belongs to someone.
The source author, record owner, or affected person retains identity. A model cannot infer intent from polished language or claim that an author accepted words they never reviewed.
Approval is scoped.
A copy editor may approve punctuation but not invent a policy obligation. A user may approve their message but not another person’s consent. Roles and capabilities remain explicit.
Preference does not authorize action.
Remembered preference can help rank candidates. Sending, publishing, disclosing, or contacting requires current consent with purpose, channel, audience, freshness, and revocation state.
Refusal can preserve authorship.
When identity, source, safety, or meaning cannot be supported, the system should explain the gap, keep the candidate unaccepted, and ask for the missing authority rather than fabricate certainty.
PORTFOLIO PROOF
One preservation contract across documents, people, and interpretation.
The content and stakes differ, but the governing pattern remains stable: bind the source, protect meaning, keep generated output provisional, require current authority, commit deliberately, verify the result, and preserve recovery.
Meaning Lock for consequential documents
Page-aware evidence, immutable originals, protected facts and obligations, exact candidate diffs, author and reviewer authority, derived revisions, receipts, and rollback turn proofreading into document assurance.
See Vellucent product proof →Consent-aware interpersonal assistance
Preferences may shape discovery, but current consent, safety, scope, and human choice govern introductions and communication. Drafting remains distinct from sending.
See Veluris product proof →Bounded interpretation without false certainty
Deterministic calculations remain separate from symbolic language. Tone can evolve while uncertainty, source state, user authorship, and the boundary against destiny claims remain intact.
See Veil product proof →New evidence changes the future, not the old receipt
Source corrections and new revisions can support a new accepted artifact without altering what was knowable or authoritative at an earlier decision cutoff.
Explore evidence and temporal truth →Untrusted content never inherits capability
A document, message, retrieved passage, or generated candidate remains data. It cannot grant itself permission, widen consent, execute a command, or bypass policy.
Explore the security architecture →Why fluency alone is the wrong goal
The editorial argument connects product experience to an engineering contract for protected claims, exact diffs, source identity, reversible commit, abstention, testing, and proof.
Read the full article →