Your tools draft the work. Planisphere proves it was supervised.
Wrap the release boundary — where AI-drafted work leaves the firm — with a gate, a human sign-off, and a sealed receipt. ABA Opinion 512’s review duty, operationalized as records a fee auditor or opposing counsel can verify without trusting you.
No legal advice, no admissibility promise: the record shows the gate, who signed off, the seal, and what’s pending.
The duty is the lawyer’s, not the vendor’s.
Every legal AI tool your firm runs drafts, cites, and summarizes under rules that put the responsibility on the lawyer, not the vendor. Rule 1.1 demands competence in the technology you use. Rule 5.3 demands you supervise non-lawyer assistance — and a model is non-lawyer assistance. Rule 3.3 demands candor to the tribunal, which a fabricated citation destroys. ABA Formal Opinion 512 makes the review duty for generative AI explicit, and FRE 702 shapes how a court weighs AI-assisted work when it is challenged.
The pressure is not hypothetical: courts have sanctioned attorneys for AI-fabricated authority, including matters where the firm had a written AI policy. A written policy is a promise. A sealed record of who reviewed what, before it left the firm, is evidence. That is the difference this page is about — and it is what internal or external compliance review will ask for. For jurisdiction-specific guidance, consult the state bar AI ethics opinions alongside ABA Formal Opinion 512. The full reference map for the regimes lives at the law library surface.
The release boundary.
The release boundary is the one place in the workflow where AI-drafted work stops being internal: the moment a billing entry, a filing, or a client deliverable leaves the firm. Planisphere wraps exactly that crossing. The gate holds the act and returns needs_review; a named human decides; the seal turns the decision into a receipt; the export is the audit trail. The whole loop below ran verbatim against the production API:
import { Planisphere } from "planisphere-sdk";
// key is positional — not { apiKey }
const ps = new Planisphere("ps_test_YOUR_KEY");
// 1. Gate the release boundary
const d = await ps.gate({
pack: "law",
proposed_action: "Release AI-drafted billing entry" +
" 4821 to the billing system.",
surface: "billing-release",
source_key: "laurel:entry:4821",
law_context: { matter_band: "billing", tool_id: "laurel" },
// retries seal once, meter once
idempotency_key: "entry-4821-release",
});
// → decision: "needs_review", route: legal-ops:policy-review
// — the block IS the product
// 2. The human supervision receipt
await ps.recordReview({
pack: "law", action_key: d.action_key,
reviewer: "partner:senior-review", decision: "approved",
});
// 3. The audit trail (time-window export;
// each record carries verify_request)
const url =
"https://api.planisphere.ooo/tenant/exports/records";
const exp = await (await fetch(url + "?from=...&to=...", {
headers: { "X-Planisphere-Key": "ps_test_YOUR_KEY" },
})).json();
// 4. Verify from the trail — no trust required
const v = await ps.verify(exp.records[0].verify_request);
// → verified: true, signature_valid: true,
// key_status: "sandbox" (honestly labeled)
One integration point, four calls, and the firm holds a supervision record for every release. The needs_review block is not friction to engineer away — it is the review duty itself, running in code.
Privilege by construction.
By default Planisphere commits hashes, never content. Every packet carries raw_content_stored: false: the prompt, the draft, the matter, and any privileged work product stay inside the firm. Where a matter calls for the material to be preserved next to the decision, the firm may optionally attach it as ciphertext it encrypted itself — Planisphere receives neither the plaintext nor the key, and the firm decrypts it on its own machines. What crosses the boundary is the binding — hashes, decisions, timestamps, and who signed off — which is enough for a third party to check the record without ever seeing the work.
Why not your own log.
A log records events. A signed receipt adds a check on the integrity and signer of the recorded metadata. The signing key is published at a public well-known address, and the record check can verify an exported record in the browser. Keep the receipt, verification material, and underlying review evidence together. The receiving team decides whether that evidence is sufficient; a valid signature does not establish that the legal work was correct.
How to verify AI legal citations before filing.
Open the underlying authority, compare the cited passage with the draft, and have a qualified lawyer decide whether it supports the legal proposition. Finding a real case is only the first check. These three synthetic examples show different ways a citation can fail.
1. The authority cannot be found.
Example: an AI draft supplies a case name, court, date, and citation, but the lawyer cannot locate a matching opinion. Search the court’s published opinions or docket and a reliable legal research source. Record what was checked and leave the citation unresolved until the authority is obtained or the draft is corrected. An unsuccessful search is a reason to investigate, not proof by itself that a case is fabricated.
2. The case exists, but the pinpoint is wrong.
Example: the draft attributes a quotation to page 12; the cited page contains a different passage. Compare the full opinion, its pagination, and the words around the quotation. Preserve the exact passage and corrected pinpoint for review. A matching case title does not validate a quotation.
3. The passage is real, but the proposition is unsupported.
Example: the draft presents a party’s argument as the court’s holding. Read the surrounding analysis and disposition, then check the authority’s relevance and current status for the intended use. Record the lawyer’s reasoning and correction; repeating the AI’s explanation is not an independent check.
Keep a review record: the draft proposition, source identifier, exact passage or pinpoint, discrepancy found, responsible lawyer, review date, and final decision. These are suggested workflow fields, not an API request schema or a prescribed bar-association form. Keep client material within the firm’s approved systems.
The competence discussion in ABA Formal Opinion 512, pages 3–4, addresses independent verification and review of generative-AI output. The appropriate review depends on the task; applicable court rules and jurisdictional duties also matter. Source checked September 8, 2026 for this section.
Continue with AI citation sanctions and verification questions or inspect a sample evidence record. A Planisphere receipt can record a review event; it does not determine whether an authority is good law or a filing is legally sufficient.
Next step.
Get a free sandbox key — no card — and wrap one release boundary. The test key covers 100 sealed actions, and sandbox receipts say SANDBOX at every layer, so a test record can never be mistaken for a production one. An engineer with your stack can have a verified receipt in an afternoon; upgrade with paid checkout on the same email and the tenant history carries over. No client data changes hands at any step.