RED PRESS

Our standards

Red Press is written to a public sourcing standard. Every claim is traced, dated, and checkable. Holding ourselves to rules readers can inspect is what makes the section worth citing.

The sourcing standard

Three tiers, and what each is allowed to carry.

Tier 1 Primary

Regulator reports, court filings, company disclosures, official post-incident reports, standards bodies, peer-reviewed research, and - for our own claims - first-party records disclosed as such.

May carry a claim on its own. Always linked and dated. Required for any claim about a named company or product (or Tier 2 verified against Tier 1).

Tier 2 Verified secondary

Established press and research houses with named authors and editorial standards.

Does not alone carry a claim about a named company or product - pair with or verify against Tier 1.

Tier 3 Commentary

Posts, newsletters, blogs, podcasts, vendor research.

Never carries a claim alone. Quoted as opinion, attributed, linked to the original and to a third-party archived copy.

Hard rules

  • Every claim traces to its origin. If we cannot reach the origin, the claim does not ship.
  • Claims about named companies or products require Tier 1 evidence (or Tier 2 verified against Tier 1).
  • Day-of-citation archives are taken through third-party archive services and linked. We do not host copies of third-party articles.
  • Numbers carry dates. Estimates are marked as estimates and show their working.
  • No superlative ships without a benchmark behind it.
  • Corrections are public. Each piece carries a revision note; material corrections are logged on the corrections page. We aim to correct within 48 hours of confirmation.
  • Every piece is dated. Publish date and last material revision are visible on the page.
  • Every piece is versioned. Nothing is silently overwritten; changed pieces carry a revision note.
  • Every piece carries a named author with a real profile. No team bylines.

Our own records

Claims about Red First can be sourced to documents we control - disclosed as ours, dated, attested by a named person with a public author profile, and marked for whether the capability is shipping or still design. That is Tier 0. Publishing the quoted line here turns a private record into something a reader can hold us to. Later pieces cite these excerpts the ordinary way.

A published excerpt is a one-way door: it is quotable against us forever. Material corrections go through Corrections, not silent edits.

Published excerpts

Each entry is the load-bearing sentence from an internal record, with its ratification date, attestation, maturity, and how you could check it.

  1. One environment per client

    Shipping

    One client's data is invisible to every other client, enforced at the database itself, not just in application code.

    Document
    canon/TENANCY_AND_RLS.md
    Agreed
    2026-08-03
    Attested by
    Rob Warner
    How you could check it
    Automated cross-tenant read-fail harness (own-tenant read succeeds; cross-tenant denied). Policy-on-table demo on request.
  2. The work spine

    Agreed design

    Work is organised as a connected chain - North Star → Objective → Milestone → Task - and agents are given that chain as context before they start, rather than working from a prompt alone.

    Document
    canon/WORK_SPINE.md
    Agreed
    2026-07-03
    Attested by
    Rob Warner
    How you could check it
    Walkthrough of a live environment showing a task resolving upward to the objective above it. Not independently checkable from outside a tenant.
  3. Removing personal data before the model sees it

    Agreed design

    Identifying details are stripped out before a prompt is sent to an outside model, and put back before the answer is shown to the person who asked for it.

    Document
    canon/PERMISSION_SERVICE.md · DEC-21
    Agreed
    2026-06-17
    Attested by
    Rob Warner
    How you could check it
    Redacted prompt shown beside its restored display, on request. Decision record: DEC-21.
  4. Bringing an existing agent in

    Designed, not built

    Agents built in other tools are appraised and mapped into governed versions carrying an owner, a budget, an approval route and a scope - and the shape of that intake is defined as a schema rather than as a description of a process.

    Document
    canon/AGENT_IMPORT.md
    Agreed
    2026-07-17
    Attested by
    Rob Warner
    How you could check it
    Published buyer-safe schema (linked). Designed, not running - you can check the shape; you cannot yet watch it work.
  5. Where our agent layer came from

    Shipping

    RedOS's agent layer originally forked Nous Research's Hermes Agent - a different project from OpenClaw. It was phased out as RedOS moved to a multi-tenant governed architecture and no longer forms part of the codebase.

    Document
    red-agent repository history · PR #203 (final clear)
    Agreed
    2026-06-24
    Attested by
    Rob Warner
    How you could check it
    As of the attestation date: (1) search the current red-agent and red-os trees for product self-reference to Hermes / Hermes Code Red - retired-term CI guards must stay green; (2) confirm no clone carries a NousResearch/hermes-agent git remote (hermes_severance_guard; remote severed 2026-06-12, rb #309). Final content clear: red-agent PR #203 merged 2026-06-24. Scoped to the current codebase - the fork remains in git history by design.

When we get it wrong

Material errors are logged publicly. We aim to correct within 48 hours of confirmation. Corrections are never silently deleted. Until the corrections log ships fully, write to hello@redfirst.ai or see Corrections.

Back to the glossary