Namespace Governance Architecture

The namespace governance system replaces the eight-entry hardcoded CHECK constraint on lensers.profiles with a policy-driven, data-governed authority that protects strategic handles, blocks impersonation, and requires zero code changes to protect newly discovered AI providers or models.

Core idea

Every handle claim goes through a single validation authority before it reaches the database. New providers and models are data rows, not code deployments.

Handle claim attempt
  │
  ▼
identity_gov.fn_validate_handle(candidate)
  │
  ├─ [1] Syntax guard (length, charset)
  ├─ [2] Normalizer  (NFKC → leet fold → repeat collapse → filler strip)
  ├─ [3] Exact registry hit
  ├─ [4] Prefix / suffix / regex pattern match
  ├─ [5] Token + modifier composition check
  ├─ [6] Homoglyph skeleton probe (TR-39)
  ├─ [7] Levenshtein fuzzy probe (edit distance ≤ 2)
  ├─ [8] Trigram similarity probe (pg_trgm > 0.70)
  └─ [9] Phonetic probe (Double-Metaphone)
         │
         ▼
    verdict: allow | escalate | deny

Fail-CLOSED: on any internal exception, the function returns deny. Ecosystem safety is never traded for onboarding convenience.

The one table

identity_gov.reserved_namespaces (
  entry_kind  exact | prefix | suffix | token | regex
  value       pre-normalized string or POSIX pattern
  class       system | security | provider | model | future | verified_only | restricted
  deny_score  0..100  (floor risk score when this entry matches)
  source      canonical | manifest | ai_inferred
  expires_at  NULL = permanent; set for ai_inferred during 30-day grace window
)

One table replaces seven draft tables and a family of CHECK constraints. Combinatorial coverage (e.g. openai-support, openai-bot, openai-admin) is achieved by token entries combined with the modifier list inside fn_validate_handle — no row explosion.

Namespace classification

Class	Score	Examples	Who can claim
system	100	lenserfight, chainabit	Never claimable
security	95	admin, mod, abuse, trust	Internal ops only
provider	90	openai, anthropic, google	DNS proof + trademark
model	85	claude, gpt, gemini, llama	Provider-verified only
future	80	lenscore, lensruntime, lenso	Held for roadmap products
verified_only	70	Notable public figures	T&S approval
restricted	50	Pattern-flagged via suffix/token	Appeal-only
public	—	Everything else	Self-serve

Protected handles (canonical seed)

The following handles are seeded into identity_gov.reserved_namespaces at bootstrap. All values are pre-normalized (lowercase, ASCII-clean). Combinatorial coverage (e.g. openai-support, claude-bot) is achieved via prefix and token entries, not row explosion.

System / platform identities — score 100

Handle	Entry kinds	Why
lenserfight	exact, prefix	Platform canonical identity; no variant may be claimed
chainabit	exact, prefix	Chainabit ecosystem identity
conectlens	exact, prefix	ConectLens parent brand
lenserfight-hq	exact	Platform HQ handle

Security / moderation — score 90–95

These handles exist to prevent impersonation of platform operations roles.

Handle	Score	Why
admin	95	Platform admin namespace
root	95	System root namespace
system	95	System identity namespace
moderator	95	Moderation team namespace
mod	95	Short moderation namespace
trust	95	Trust & Safety namespace
safety	95	Safety team namespace
abuse	95	Abuse reporting namespace
report	95	Report namespace
security	95	Security team namespace
support	95	Platform support namespace
null	95	Null/undefined identity guard
undefined	95	Undefined guard
help	90	Help namespace
policy	90	Policy namespace
legal	90	Legal namespace
dmca	90	DMCA reporting namespace
anonymous	90	Anonymous identity guard

LenserFight brand family — score 80

Reserved for current and future platform products. No user may claim a lens* or lola handle without T&S approval.

Handle	Why
lenser	LenserFight core brand token (also a token entry)
lens	Core brand token (also a token entry)
lenso	Lenser AI family — @lenso
lensa	Lenser AI family — @lensa
lense	Lenser AI family — @lense
lola	Lenser AI family — @lola (also a token entry)
lenscore	Future product: LensCore
lensruntime	Future product: LensRuntime
lensagent	Future product: LensAgent
lensos	Future product: LensOS
lensai	Future product: LensAI
lensinfra	Future product: LensInfra
lensizm	Legacy reserved handle

Chainabit ecosystem handles reserved for future products:

Handle	Why
chaina	Chainabit brand token (also a token entry)
chaincore	Future: ChainCore
chainagent	Future: ChainAgent
chainruntime	Future: ChainRuntime
chainnode	Future: ChainNode
chainlabs	Future: ChainLabs
chaininfra	Future: ChainInfra

AI provider companies — score 85–90

Blocked to prevent trademark impersonation. Exact + prefix + token entries provide combinatorial coverage.

Handle	Score	Source
openai	90	OpenAI (trademark)
anthropic	90	Anthropic PBC (trademark)
google	90	Google LLC (trademark)
deepmind	90	Google DeepMind
meta	90	Meta Platforms
xai	90	xAI
mistral	90	Mistral AI
deepseek	90	DeepSeek AI
groq	90	Groq Inc.
cohere	90	Cohere Inc.
perplexity	90	Perplexity AI
huggingface	90	HuggingFace Inc.
stabilityai	90	Stability AI
elevenlabs	90	ElevenLabs
nvidia	90	NVIDIA Corp.
microsoft	90	Microsoft Corp.
amazon	90	Amazon/AWS
vertexai	90	Google Vertex AI
geminiapi	90	Google Gemini API
togetherai	90	Together AI
together	85	Together AI (short form)
stability	85	Stability AI (short form)
replicate	85	Replicate Inc.
bedrock	85	AWS Bedrock
sagemaker	85	AWS SageMaker
fireworks	85	Fireworks AI
octoai	85	OctoAI
anyscale	85	Anyscale
cerebras	85	Cerebras Systems

AI model names — score 75–90

Blocked to prevent model impersonation. Provider-verified accounts may petition T&S for access.

Handle	Score	Why
chatgpt	90	ChatGPT product name
claude	90	Anthropic Claude
gemini	90	Google Gemini
grok	90	xAI Grok
gpt4o	85	GPT-4o
gpt4	85	GPT-4
gpt3	85	GPT-3
gpt	85	OpenAI GPT family
llama	85	Meta Llama family
deepseekr1	85	DeepSeek-R1
sonnet	85	Claude Sonnet tier
opus	85	Claude Opus tier
sora	85	OpenAI Sora
commandr	80	Cohere Command-R
qwen	80	Alibaba Qwen
phi	80	Microsoft Phi
haiku	80	Claude Haiku tier
gemma	80	Google Gemma
mixtral	80	Mistral Mixtral
codex	80	OpenAI Codex
dalle	80	OpenAI DALL-E (slug)
dall-e	80	OpenAI DALL-E
stable-diffusion	80	Stability AI SD
whisper	80	OpenAI Whisper
mistral7b	80	Mistral 7B
o1	80	OpenAI o1
sdxl	75	Stable Diffusion XL
falcon	75	TII Falcon

Suffix / composition guards — score 50

These entries block impersonation patterns like mybot-official or something-admin. Score is lower because the tokens may appear legitimately in non-impersonation handles; the composition check in fn_validate_handle handles high-risk combinations.

Suffix	Why
official	Impersonation suffix guard
verified	Impersonation suffix guard
iam	Identity claim suffix guard

Detection pipeline detail

[2] Normalizer — identity_gov.fn_normalize_handle

The IMMUTABLE normalizer reduces attack surface before any registry lookup:

1. Strip invisible/directional Unicode (ZWJ, ZWNJ, RTL marks, BOM)
2. NFKC normalization (ligatures: fi→fi, fullwidth: Ａ→A)
3. Casefold (lower())
4. Homoglyph fold — Cyrillic а/е/о/р/с/х → a/e/o/p/c/x
5. Leet-speak fold — 0→o, 1→l, 3→e, 4→a, 5→s, 7→t, @→a, $→s, 8→b
6. Repeat collapse — runs of 3+ identical chars → 2 (ooopenai→oopenai)
7. Filler suffix strip — official, support, admin, bot, ai, hq, team…
8. Leading the_ / the- strip
9. Trim separator chars

openai-support → openai → exact hit → deny.

[6] Homoglyph skeleton — identity_gov.fn_handle_skeleton

Applied on top of the normalizer. Collapses font-level confusables: rn→m, vv→w, ii→u. Detects rneta → meta.

[5] Composition check

If any token entry appears inside the normalized handle and the original candidate contains a modifier word, the handle is denied:

lens_ai   → token 'lens' + modifier 'ai' → deny
gpt-admin → token 'gpt'  + modifier 'admin' → deny
claude-bot → token 'claude' + modifier 'bot' → deny

Modifier list lives in the function. Adding a new modifier is a single-line code change; adding a new protected token is a single row insert.

Functions (public API surface)

Function	Volatility	Caller	Purpose
public.fn_check_handle(text)	STABLE	authenticated, anon	Advisory read-through; safe for debounced UI calls
identity_gov.fn_validate_handle(text)	STABLE	trigger, fn_check_handle	Authoritative 9-step pipeline
identity_gov.fn_normalize_handle(text)	IMMUTABLE	fn_validate_handle	Pure normalizer
identity_gov.fn_handle_skeleton(text)	IMMUTABLE	fn_validate_handle	Homoglyph folder

fn_check_handle is advisory — it is cached and safe to call from UI on every keystroke (debounced). The BEFORE INSERT/UPDATE trigger on lensers.profiles runs fn_validate_handle as the hard enforcement gate. Cache staleness can never produce a wrongful approval on the write path.

Security model

• identity_gov schema has FORCE ROW LEVEL SECURITY on all tables.
• authenticated and anon roles have SELECT only on reserved_namespaces.
• All write mutations go through SECURITY DEFINER functions with SET search_path = ''.
• The trigger on lensers.profiles raises P0001 on deny — no bypass through direct SQL.
• A superuser/service_role can still insert rows (seeding, moderation) — the trigger only fires for client roles.

Adding new providers or models

No code change needed. Insert a row:

INSERT INTO identity_gov.reserved_namespaces
  (entry_kind, value, class, deny_score, reason, source)
VALUES
  ('exact',  'newprovider',  'provider', 90, 'NewProvider Inc.', 'manifest'),
  ('prefix', 'newprovider',  'provider', 90, 'Prefix guard',     'manifest'),
  ('token',  'newprovider',  'provider', 90, 'Composition guard','manifest');

For ai_inferred entries (from the future AI Intel pipeline), set expires_at = now() + interval '30 days'. Entries that are not promoted within the grace window expire and become claimable again.

GRASP / OOAD alignment

Principle	Application
Information Expert	identity_gov owns all normalization, detection, and decision logic
Protected Variations	New providers/models are data rows — the variation point is closed to code change
Pure Fabrication	fn_normalize_handle, fn_handle_skeleton are stateless services
Controller	fn_validate_handle is the single validation authority
Low Coupling	lensers.profiles trigger delegates entirely to fn_validate_handle
Chain of Responsibility	9-step pipeline; each step is independent, first terminal verdict wins
Registry	reserved_namespaces is the canonical registry
Strategy	entry_kind column selects the matching strategy at runtime
Specification	Each probe is a self-contained boolean sub-query

Tables introduced

Table	Purpose
identity_gov.reserved_namespaces	The single registry — exact handles, prefix/suffix/token/regex rules

One schema, one table. No audit tables, no policy tables — the governance logic is pure SQL functions backed by Postgres extension capabilities (pg_trgm, fuzzystrmatch).

Related

• RFC-0005: Namespace Governance
• Lenser Family
• AI Cost Governance & BYOK