Governance

A genome holds a business’s preferences and, through references, traces of the people who gave them. Governance is everything that lets a workspace control it. Most of it is built and tested. Most of it is also switched off unless an operator turns it on.

Heredity layers

An administrator wants one rule for every agent: “invoices go to the finance head for approval.” A team wants its own tone. A member wants their own drafts shorter. Heredity gives each of them a layer that agents inherit. Each layer is a log of its own, kept in the same store as agent logs, so it has the same tenant isolation, rollback and erasure.

The inherited layers

Layer	Written by	Limits
Workspace policy (org)	Administrators	At most 20 genes.
Team policy (team)	Administrators	At most 10 genes per team. A run applies at most 3 of the teams of the person who started it, in rank order. A workspace has at most 50 teams.
Shared knowledge (shared)	Administrators	Styles, approvals, decisions, business statements and roles by default. Rules and traits only when the administrator chooses. Never tool preferences, fact pointers or identity.
Personal (user)	The member, in their own log only	Style and trait genes only. Applied only to that member’s runs.

The characters of the org layer plus the three largest team layers are capped at 1,500 when a gene is written. Because of that cap, the compiler can treat org and team genes as protected and never drop them for size. An agent’s own customer gene outranks a shared one, because the more specific setting wins. One agent can opt out of a shared gene without affecting the others.

An administrator can promote a learned gene to the shared layer. The platform writes the shared gene, then retires the learned one. If the second step fails, it removes the shared gene again, so a failed promotion leaves no shared gene.

A “why” view shows, for each gene, whether it was included or left out and why. There are nine decision codes, such as included, dropped_budget and shadowed.

Built, flag off by default G24 to G28 Code and tests exist. The layers run only when the hierarchy, teams or shared flags are on.

Immunity

A genome that learns from customer text can be taught the wrong thing on purpose. A comment, a saved edit, an API call or an imported file can carry text that tries to address the model. The defence is a set of fixed rules. None of them calls a model.

• The screen. It normalises text first (removing invisible characters and applying Unicode normalisation), then looks for six families: instructions to ignore or forget earlier ones, role changes, system or jailbreak phrases, fake genome tags, encoded payloads, and requests to send data to an address. It covers English, Hinglish and Hindi in Devanagari script. A hit records the rule id only, never the matched text.
• Facts are refused. A gene holds no email address, URL, amount, phone number or run of four or more digits.
• The evidence floor. One person repeating the same text in one run on one day cannot reach it. These floors are constants. They are never traded away for a higher acceptance rate.
• The policy check. A candidate conflicts when its words overlap at least half with a protected gene and exactly one of the two is negated. “Always attach the invoice” against “Never attach the invoice” conflicts.

With the immune check off, only the older, smaller screen applies: five marker phrases, plus the fact patterns.

Built and tested G29 The rules themselves have code and tests and no flag. Built, flag off by default G30 Using them in the run path and the nightly job has code and tests and is off unless the immune flag is on.

Quarantine and review

A proposal is a gene offered to a person instead of being applied. A quarantined item is held out of the genome with the id of the rule that stopped it. An administrator can accept or reject a proposal, release a quarantined item, or discard it.

• A proposal that nobody decides expires after 30 days.
• A proposal the customer rejected is not proposed again for 30 days.
• Learned genes in an imported AGF file arrive as proposals, never as live genes.
• The nightly job re-screens stored learned genes and pending observations and quarantines new hits, at most 50 per workspace per night. Items an administrator released are exempt.

Built, flag off by default G31 Review routes have code and tests. They need the main genome flag plus one of the immune, format, reproduction or families flags.

Erasure

What the platform learned from a person is personal data too. When a person asks for erasure, the platform removes everything derived from them, across every workspace, by following provenance references:

• events and observations whose provenance names that person or their runs, and everything consolidated from them;
• their personal layer in every workspace, and their personal tenant;
• competence snapshots and quality scorecards that included their runs, which are recomputed;
• their pseudonyms in AGF exports and in any workspace that imported such an export;
• the previews of their own transfers, their team memberships, their episodes and their lessons.

Erasure never checks a feature flag. What is deleted does not depend on one. The platform accepts over-erasure: a learned gene that used one of the person’s observations is dropped even if it also used a colleague’s, and the colleague’s observation becomes pending again.

An optional erasure certificate records that the erasure ran. It holds hashed ids and per-store counts, no gene ids and no gene text, and is kept for 1,095 days. External agents can erase by an opaque subject reference.

Built and tested G45, G47 Erasure and workspace purge have code and tests. Built, flag off by default G46 The signed certificate is off unless the compliance flag is on.

Lineage and audit

An auditor can ask which genes shaped a given run’s prompt and where their data came from. The platform rebuilds the genome at the versions the run recorded, recompiles the section, and compares its SHA-256 with the hash it stored. If they are equal, the answer is marked as matching. The answer lists each included gene with its stratum, layer, source and classified provenance. Provenance is grouped (runs, members, imports) and never echoed raw. Writes to layers add audit rows with ids and the gene kind, never gene text.

Verification can fail for innocent reasons. A feature flag turned off after the run changes the rebuilt section. Episode and thread blocks are not rebuilt. Erasure leaves gaps in the log, which the answer reports. A run that is not verified still lists what can be read, with the reason.

Built, flag off by default G44 Lineage export has code and tests. It needs the compliance flag.

Rollback

Rollback restores an earlier state of an agent’s genome. It adds a rollback event and restores the whole saved state, so every event before it stays in the log, and a later rollback can go to a point after the first one. Built and tested G02 The rollback routes in the web app are Built, flag off by default G11 .

Limits and what is not built

The status of the rest of the genome is on the Technology page, and the open work is on the Future page.