The first confidentiality guardrail I built for my AI agents had a blacklist at its center: company names, client names and sensitive terms that must never appear in a commit to a public repository. It also treated my own biography as a leak.
My public websites already named employers, clients I had advised and collaborators I had worked with. I had written those pages and published them deliberately. Blocking a restatement of my resume made the guard harder to trust.
When I audited how my guardrails classified my repositories, I found sensitive-pattern scanning switched off for the repositories that published my websites. The name blacklist had made protection impractical there: author bios would trigger it. The classification had relieved the false alarms by leaving some of my most public surfaces without that scanning.
I needed to distinguish a fact I had chosen to publish from new information about the same organization. A name can itself be confidential. Once I've published a relationship, however, its name alone doesn't tell the guard whether the next sentence is appropriate to share.
One relationship, three different disclosures
Email, chat, meeting transcripts and private repositories give an assistant the context to help with real work. The same context includes facts its user has never published. A truthful, flattering draft can disclose those facts without repeating any name on a blacklist.
I've written about an assistant's accumulated memory and who controls it. Disclosure adds another question: which of those remembered facts may the assistant repeat to this audience?
Consider a fictional author whose public biography says, “I advised example-client on accessibility in 2024.” These are three possible additions to a public article:
| Proposed sentence | What the biography establishes | What still needs review |
|---|---|---|
| “I advised example-client on accessibility in 2024.” | The same relationship, topic and year appear in the author's published biography. | Whether this destination and use are permitted, and whether the mention serves the article. |
| “At example-client, I helped resolve an internal launch dispute.” | The advisory relationship exists in the biography. | The internal event has no approval in that precedent. |
| “I helped a regional accessibility vendor resolve an internal launch dispute.” | Omitting the name adds no permission. | The description could identify the organization, and the event remains unapproved. |
The same name can occur in an acceptable biographical statement and an unapproved operational story. An unnamed description can disclose the story too.
For review, I divide facts into two categories. The first contains what I've deliberately published or explicitly approved for the relevant use: a role, relationship, era or piece of public work. The second contains everything without that support, especially information learned in private context. A compliment about a private engagement belongs in the second category just as a criticism does.
This classification applies to facts and uses. A ledger entry for an organization cannot make everything the assistant knows about it publishable.
Stories and names I have chosen to publish already carry my editorial decision. The guard should preserve that decision while checking new material the assistant introduces from private context.
Record the precedent precisely
A precedent ledger is a version-controlled record of what was approved or published, with evidence someone can inspect. I want an entry to answer these questions:
- What exact fact does it cover?
- Where did I publish or approve it, and on what date?
- What kind of use does it cover? A biographical statement about a role is different from an account of an internal decision.
- Which destinations may use it, and does a new mention require another decision?
The third question is sometimes called the register: the kind of statement the approval covers. Writing “biography” in that field should mean something concrete, such as relationship, role, era and already public work. It should not be an allowance for operational details.
For the fictional biography, the record covers advising example-client on accessibility in 2024. It cites the biography passage and records the permitted use. The launch dispute doesn't acquire approval by sitting next to that record. If the author later approves a description of the dispute, that decision needs its own scope and evidence.
A citation makes an entry checkable. Someone still has to read it. A correctly formatted record can point to a page that never says what the record claims, or to a source that doesn't establish the author's permission. The assistant's belief that a fact is public is insufficient when that belief may have come from private context.
Precedent also permits a mention without requiring one. I can leave an approved fact out because it adds nothing to a particular article. For relationships that need a fresh decision each time, the record should say so explicitly.
What the inspected hook checks
Later source check (September 30, 2026): This section examines the September 27 hook revision linked below.
The public hook implementation makes part of this review mechanical. In the ledger loader inspected for this revision, an allowance requires evidence strings, a register from the configured vocabulary and identity patterns that exactly match eligible configured patterns. The loader then removes those eligible patterns from the scan for a permitted publication class.
That last step is broader than understanding a sentence. Once a name pattern is removed, the scanner does not decide whether a sentence using the name is biographical or operational. It does not fetch the cited page, authenticate an approval or compare the meaning of the sentence with the source. The staged scanner acquires bounded added lines and applies configured patterns. Its result needs the contextual review below.
Configuration failures have distinct meanings. If a ledger path is configured but the ledger is missing or fails the loader's checks, the loader raises an error; it cannot supply an allowance. If no ledger is configured, the loader returns no allowances.
To establish protection in a working environment, observe the check being invoked at the relevant action, with the actual configuration and failure behavior. A commit scan says nothing by itself about an email send, social scheduler or another path that can publish text.
Follow the content to its destination
A repository's visibility doesn't tell me where its contents will end up. A private repository may build a public website, whether release happens automatically or manually. A public open-source repository may have no reason to carry a client's name in a code comment, even when that relationship appears in my biography.
My policy distinguishes three destination classes:
| Destination class | Intended mechanical treatment | Remaining judgment |
|---|---|---|
| Private personal notes | Credential scanning, without the public-disclosure name restrictions. | Keep the material private; reassess any later transfer. |
| Public surfaces I author | Sensitive-pattern scanning with eligible, recorded name allowances. | Check the particular fact, register and context. |
| Other public-facing work and client work | Strict scanning, without those name allowances. | Obtain the authority appropriate to that destination. |
The inspected classifier checks configured remote patterns and falls back to the strict class when no class fits; strict matches take precedence. That policy depends on accurate configuration. A private note copied into a website draft has acquired a public destination even if the note's original repository remains private.
I apply the same destination question outside Git: where will this proposed text land, who will see it, and which approval covers that use? Merely keeping the draft in a private folder doesn't answer it.
Five checks before sharing
I use these five checks above the hook to examine what the proposed disclosure means:
- Precedent. Does the record cover this particular fact and use? In the example, the biography supports the advisory relationship and year. It supplies no precedent for the launch dispute.
- Provenance. Where did the information come from? If the assistant learned it in messages, transcripts, meetings or private repositories, find the independently published counterpart or the explicit approval before treating it as shareable. A public relationship is not a public source for an internal event.
- Negativity. Does the new text add criticism or a negative implication about an identifiable party? A neutral biographical approval doesn't cover that change. Put the actual wording before the person who can decide it.
- Identification. Could an outsider or insider work out who an unnamed passage describes? “A major metropolitan newspaper where I ran engineering” still points toward my published career history. Removing the name doesn't erase identification, and adding a private event would require separate review.
- Aggregation. Does the combination reveal something the individual facts don't? A public role and public dates may become identifying when joined to a new anecdote. Review the combination the audience will read.
Uncertainty should produce a specific question: “May this paragraph describe the internal launch dispute on the public site?” That is a decision the author can make. “Is example-client allowed?” is too broad to settle it.
The checks apply to this essay too. I can explain why my biography confused a name filter without adding details of an engagement. The newspaper description illustrates identification through an already published career category; it gives me no reason to append a private story. Describing how the ledger works also doesn't require publishing its entries. I don't need to name every relationship whose mention might be permitted.
Practice the review with a fictional workbook
The downloadable disclosure-review workbook contains the fictional biography, a source that supplies different information, a blank approval record, nine proposed disclosures and an answer key with reasons. Start with the source passages and make your own classifications before opening the key.
One case deliberately supplies a well-formed citation that doesn't support the proposed fact. Another moves a private-file draft onto a public website. The native-enforcement case has no observation of a running guard, so its answer remains unknown. The exact-biography case supplies a positive control: a review that rejects every precedent misses a valid match.
The included Python checker compares your recorded answers with the workbook's fixed answer key. It checks the fields and classifications you entered; it cannot read a new article and decide whether its disclosure is safe. All people, organizations, source passages and approval records in the workbook are fictional. A matching answer grants no real publication authority.
After extracting the ZIP, the README gives the commands for checking the completed example, trying the blank response and running the tests. The answer key distinguishes source support, precedent, destination, further review and native enforcement so one passing field cannot stand in for the others.
For the fictional draft, I'd retain the supported biographical sentence if it served the article. I'd keep the launch dispute out until the author had reviewed that specific disclosure and its destination. The unnamed version needs the same decision.
Enforcement is a velocity feature
I built this expecting to pay a speed tax for safety. The opposite happened.
Before the redesign, every outward-facing draft brought back a diffuse question: had the assistant said anything it shouldn't? That question re-litigated my entire disclosure posture at every familiar name. Recording precedents and applying five ordered tests changed the work. I could focus on the particular new fact or implication. The per-draft review cost collapsed, and what remains is the writing.
The policy model, ledger schema and hook engine are open source. The private entries belong to the person whose information they govern. Adopting the code still requires choosing the configuration, reviewing the precedents and checking the paths that actually carry information outward.