Skip to content

trust & governance

We structurally cannot spam

Most AI marketing tools ask you to trust a system prompt. This page renders our governance from the live database instead: the banned actions, the schema constraints, the approval gate, and a fresh verification of the tamper-evident audit log — regenerated hourly.

Audit chain: intact ✓

dg_verify_chain · 2264 actions · every content hash recomputes · as of 07:12 UTC (re-runs hourly)

2 benign concurrent-append forks: early-era appends that raced and linked the same parent — every row’s content verifies, and appends have been serialized since. Precision beats a silent green checkmark.

Banned at the database level

These are not settings. They are BLOCK rows in the live policy engine — the gateway short-circuits them before any connector code runs. This list is queried from the running database on every regeneration.

cold_email · cold_email.unconsented_blast · BLOCK

Unconsented blast

Blasts to people who never opted in (email, SMS, push, cold outreach).

community · community.vote_manipulate · BLOCK

Vote manipulate

Upvote / vote manipulation in communities. No tool schema exists for it.

directories · directories.fake_review · BLOCK

Fake review

Fabricated reviews. Illegal under FTC rules; unrunnable here.

discord · discord.mass_dm · BLOCK

Mass dm

Automated or bulk direct messages. Platform ToS forbids them; so does our schema.

g2 · g2.incentivized_review · BLOCK

Incentivized review

Pay-for-positive reviews. Same FTC line, same structural block.

gen_engine · gen_engine.prompt_injection_seeding · BLOCK

Prompt injection seeding

Seeding hidden prompt-injection text to manipulate AI answer engines.

link_building · link_building.paid_link_scheme · BLOCK

Paid link scheme

Paid link schemes — buying rankings violates every engine’s guidelines.

linkedin · linkedin.auto_connect · BLOCK

Auto connect

Automated connection requests — growth-hacking that burns the account it runs on.

linkedin · linkedin.auto_dm · BLOCK

Auto dm

Automated or bulk direct messages. Platform ToS forbids them; so does our schema.

push · push.unconsented_blast · BLOCK

Unconsented blast

Blasts to people who never opted in (email, SMS, push, cold outreach).

reviews · reviews.fake_review · BLOCK

Fake review

Fabricated reviews. Illegal under FTC rules; unrunnable here.

sms · sms.unconsented_blast · BLOCK

Unconsented blast

Blasts to people who never opted in (email, SMS, push, cold outreach).

social_dm · social_dm.automated_blast · BLOCK

Automated blast

Automated or bulk direct messages. Platform ToS forbids them; so does our schema.

x · x.auto_dm · BLOCK

Auto dm

Automated or bulk direct messages. Platform ToS forbids them; so does our schema.

Alongside these, 48 publish/send/spend actions carry a REQUIRE_APPROVAL row: each one stops at the gateway and waits for a human.

Staked + unattended is unrepresentable

The invariant everyone else puts in a system prompt, we put in the schema. Every action carries a recovery class — undoable, fading, compounding, or staked — and anything staked (money, accounts, legal exposure, binding commitments) can never run unattended: these CHECK constraints, quoted verbatim from our migrations, make that a state the database refuses to store, at the action layer and the policy layer both.

CONSTRAINT policy_matrix_staked_not_autopilot
  CHECK (NOT (recovery = 'staked' AND autonomy_level = 'autopilot'
              AND risk_class <> 'AUTO'))
CONSTRAINT action_log_staked_not_autopilot
  CHECK (NOT (recovery = 'staked' AND autonomy_level = 'autopilot'
              AND risk_class <> 'AUTO'))

packages/db/migrations/0044_recovery_model.sql — enforced by Postgres, not by prompt (the original 0001 irreversible constraints still stand beneath them; the AUTO carve-out covers drafts and audits — internal artifacts whose autonomy label is decorative — never an egress).

Autopilot is earned, per action

promote

5 clean approvals

An action becomes eligible only after 5 consecutive decisions that were approved, unedited, and didn’t flop — and what it earns depends on how a mistake recovers: undoable work runs free, work that fades publishes behind a veto window you can cancel, work that compounds waits for a health sensor. You still flip the switch.

demote

One flop, instantly

A single self-reported flop on a promoted action demotes it straight back behind the approval gate. Trust is lost faster than it is earned.

never

Staked actions

Money moved, payouts approved, binding commitments made — anything that puts real stakes on the line can never be promoted, no matter the streak. See the constraints above.

Billing honesty, in advance

marketinque is pre-launch (waitlist only). We’re publishing the billing rules before the billing system exists, so you can hold us to them on the day it does:

  • A flat monthly price. No expiring credits, no credit “redenominations”, no surprise meters.
  • Only approved actions count. Drafts, critic culls, rejected ideas, and failed runs are free — always.
  • The meter is a SQL view over the append-only action log. You can verify your own bill against the hash chain.
  • Cancel means cancel: billing stops immediately, your data exports cleanly, and nothing re-charges later.

Straight answers

Can marketinque post to Reddit, LinkedIn, or X without me?

Structurally, no. Every public-surface publish, send, or spend is a REQUIRE_APPROVAL policy row: the gateway stops it before any connector runs and waits for an explicit human decision. The earned-autopilot ladder only ever covers reversible actions on your own surfaces (like your own blog) — public channels are permanently gated.

Could it send mass DMs, cold blasts, or fake reviews if I asked it to?

No — those actions have no execution path. They exist in the policy engine only as BLOCK rows, which the gateway short-circuits before any connector is invoked. This page renders that list live from the database.

What does “autopilot” mean here?

Autopilot is earned per action, never configured blind. A reversible, own-surface action becomes eligible only after 5 consecutive clean approvals (approved, unedited, no flop outcome). One self-reported flop demotes it back behind the approval gate instantly. Irreversible actions can never be promoted — the database refuses to represent that state.

How do I know the audit log hasn’t been edited?

Every action is appended to a write-once log where each row carries a SHA-256 hash of its content chained to the previous row’s hash. The database can recompute the whole chain on demand (dg_verify_chain); this page re-runs that verification every hour and shows the result.

Can AI-generated placeholder content reach the public site?

No. Drafts produced without a real model are stamped with a placeholder token, and the publish connector refuses them outright — the action fails instead of going live. Real drafts additionally pass a QA gate and a critic pass before a human ever sees them.

An agent you can audit, not just trust

Everything on this page applies to the agent that markets marketinque itself, today, in production. Join the waitlist to run it on your funnel.