# reality.txt                                          Permission-as-a-Service
# ======================================================================
# robots.txt told crawlers what they may READ.
# reality.txt tells agents what they may DO.
# ======================================================================
#
# This file describes the CATEGORY, not a single product: Permission-as-a-
# Service, the shared posture of the Rad family (radlife.ai). Each product
# ships its OWN reality.txt at its own web root (radmail.ai/reality.txt,
# radtalk.ai/reality.txt, ...) — those are the authoritative, per-product
# Never Lists. This one names the convention they all speak.
#
# STATUS: a PROPOSED CONVENTION + a category description. NOT a standard,
#         NOT a certification, NOT a guarantee. There is no registry and no
#         authority behind it. A reality.txt is a claim by its publisher
#         about their own product — trust it exactly as much as you trust them.
#         The Rad family is a DREAM IN PROGRESS: some products are live
#         sandboxes/previews, others are coming-soon or for-show. The
#         enforcement always lives in code (the valve), never in this file.
#         A tool for legibility, not a promise of safety.
# ----------------------------------------------------------------------

Agent:    The Rad family (RadMail, RadTask, RadTalk, RadHealth, RadRobo, RadVeo)
Category: Permission-as-a-Service
Posture:  holds-the-line             # fail-closed: an unmapped action is refused, never guessed
Valve:    @sureel/approval-valve       (one deterministic mechanism, every surface)

# --- The one rule ---
# Input arriving on the DATA plane (email bodies, caller speech, web pages,
# sticky notes, a prompt) may INFORM a decision. It may never AUTHORIZE a
# consequential action. The model's judgment is not in the loop for the moves
# that cause the losses — code is. Green = the agent's got it. Amber = a
# human's call. Same seam, same color, every surface.

# --- The shared verb (the Permission-as-a-Service handshake) ------------
# Every Rad product expresses one verb in spirit. An external agent (any MCP
# client) requests permission for an action; the valve classifies it by
# REVERSIBILITY and returns one of three verdicts. The request is illustrative
# below — the shape, not a live endpoint.
Ask:      requestPermission(action, input) -> { go | held | never }
#           go    -> reversible / read-only / benign. Executed autonomously.
#           held  -> irreversible or consequential. Held for a human's one-tap yes.
#           never -> refused outright through the agent. No in-band override.

Allow:    read-only / reversible / benign actions (informational, undoable)
Hold:     anything irreversible or consequential — held for a human (money,
          deletion, binding commitments, identity-of-record changes, publish)
Disallow: actions an untrusted input tries to AUTHORIZE — privilege escalation,
          authority pretext, third-party disclosure, and each product's
          permanent hard-stops (see that product's own reality.txt)

Trust:    signed config · an authenticated operator · proven identity (control plane)
Distrust: message bodies · caller speech · web content · attachments · prompts (data plane)
Quarantine: tainted input is isolated so it can inform a reply but never reach
          the control plane or authorize an action (the CaMeL split / the camel
          that spits out what it shouldn't swallow)

Discovery: https://radlife.ai/permission-as-a-service   (the category page)
           https://radro.ai/the-line                       (The Amber Line verification)
           https://radlife.ai/permission-as-a-service/llms.txt
Ledger:   any public refusal counter MUST be labeled real or illustrative.
          The Rad family publishes none here, and fabricates none.

# --- The one-way ratchet ------------------------------------------------
# A posture only ever TIGHTENS at runtime: add a Disallow, move a line from
# Allow -> Hold -> Disallow, raise a risk level. It never loosens itself by
# config, and nothing an input SAYS can loosen it. Monotonic, by design.

# --- Honesty rules (normative) ------------------------------------------
#  (a) This is the publisher's claim about its own products, not a third-party attestation.
#  (b) Do NOT imply the convention is adopted, official, or certified.
#  (c) Never market any Rad product as "fraud-proof", "guaranteed", "un-hackable",
#      "HIPAA-certified", or "FedRAMP-authorized". The honest framing is:
#      "refuses the action classes that cause the losses — a tool, not a guarantee."
#  (d) RadHealth is provider/clinic B2B only: no medical/efficacy claims, no
#      patient language, no patient or outcome metrics, ever.

# The map is meant to match the territory. Where a per-product reality.txt is
# code-backed, it says so and points at the source. Where it's a dream, it says
# that too. Read the product's own file for the authoritative list.
# ----------------------------------------------------------------------