- -
-
+ +
+
+ March 2026 +

What’s New

+
+ +
+ +
+ New Research +
+
+

Guardian Agents and the Philosophy of AI Accountability

+

How Wittgenstein, Berlin, Ostrom, and Te Ao Māori converge on the same architectural requirements for governing AI in community contexts.

+
+ + + +
+ Deployed +
+
+

Guardian Agents in Production

+

Four-phase verification using mathematical similarity, not generative checking. Confidence badges, claim-level analysis, and adaptive learning — all tenant-scoped.

+
+ + + +
+ Beta Open +
+
+

Village Beta Pilot

+

Village is accepting beta applications from communities ready to participate in constitutional AI governance. Community Basic from $10/mo.

+
+ +
+
+ + +
+

The Problem

@@ -107,24 +147,14 @@ A user told Claude Code to use port 27027. The model used 27017 instead — not from forgetting, but because MongoDB’s default port is 27017, and the model’s statistical priors “autocorrected” the explicit instruction. Training pattern bias overrode human intent.

-
-

From Code to Conversation: The Same Mechanism

-

- In code, this bias produces measurable failures — wrong port, connection refused, incident logged in 14.7ms. But the same architectural flaw operates in every AI conversation, where it is far harder to detect. -

-

- When a user from a collectivist culture asks for family advice, the model defaults to Western individualist framing — because that is what 95% of the training data reflects. When a Māori user asks about data guardianship, the model offers property-rights language instead of kaitiakitanga. When someone asks about end-of-life decisions, the model defaults to utilitarian calculus rather than the user’s religious or cultural framework. -

-

- The mechanism is identical: training data distributions override the user’s actual context. In code, the failure is binary and detectable. In conversation, it is gradient and invisible — culturally inappropriate advice looks like “good advice” to the system, and often to the user. There is no CrossReferenceValidator catching it in 14.7ms. -

+

+ The same mechanism operates in every AI conversation. When a user from a collectivist culture asks for family advice, the model defaults to Western individualist framing. When a Māori user asks about data guardianship, the model offers property-rights language. Training data distributions override user context — in code the failure is binary and detectable, in conversation it is gradient and invisible. +

+
Read the full analysis →
-

- This is not an edge case, and it is not limited to code. It is a category of failure that gets worse as models become more capable: stronger patterns produce more confident overrides — whether the override substitutes a port number or a value system. Safety through training alone is insufficient. The failure mode is structural, it operates across every domain where AI acts, and the solution must be structural. -

@@ -178,84 +208,126 @@ - +
-

Six Governance Services

+

Governance Architecture

- Every AI action passes through six external services before execution. Governance operates in the critical path — bypasses require explicit flags and are logged. + Six governance services in the critical path, plus Guardian Agents verifying every AI response. Bypasses require explicit flags and are logged.

-
-
-
- + +
+
+
+ + + +
+
+
+

Guardian Agents

+ NEW — March 2026 +
+

+ Four-phase verification using embedding cosine similarity — mathematical measurement, not generative checking. The watcher operates in a fundamentally different epistemic domain from the system it watches, avoiding common-mode failure. +

+
+
+
Phase 1
+
Response Verification
+
+
+
Phase 2
+
Claim-Level Analysis
+
+
+
Phase 3
+
Anomaly Detection
+
+
+
Phase 4
+
Adaptive Learning
+
+
+
+ Full Guardian Agents architecture → +
+
+
+
+ + +
+
+
+
-

BoundaryEnforcer

+

BoundaryEnforcer

Blocks AI from making values decisions. Privacy trade-offs, ethical questions, and cultural context require human judgment — architecturally enforced.

-
-
- +
+
+
-

InstructionPersistenceClassifier

+

InstructionPersistenceClassifier

Classifies instructions by persistence (HIGH/MEDIUM/LOW) and quadrant. Stores them externally so they cannot be overridden by training patterns.

-
-
- +
+
+
-

CrossReferenceValidator

+

CrossReferenceValidator

Validates AI actions against stored instructions. When the AI proposes an action that conflicts with an explicit instruction, the instruction takes precedence.

-
-
- +
+
+
-

ContextPressureMonitor

+

ContextPressureMonitor

Detects degraded operating conditions (token pressure, error rates, complexity) and adjusts verification intensity. Graduated response prevents both alert fatigue and silent degradation.

-
-
- +
+
+
-

MetacognitiveVerifier

+

MetacognitiveVerifier

AI self-checks alignment, coherence, and safety before execution. Triggered selectively on complex operations to avoid overhead on routine tasks.

-
-
- +
+
+
-

PluralisticDeliberationOrchestrator

+

PluralisticDeliberationOrchestrator

When AI encounters values conflicts, it halts and coordinates deliberation among affected stakeholders rather than making autonomous choices.

@@ -281,16 +353,20 @@

-
-
-
6
-
Governance services per response
+
+
+
4
+
Guardian verification phases per response
-
-
11+
+
+
6
+
Governance services in the critical path
+
+
+
17
Months in production
-
+
~5%
Governance overhead per interaction
@@ -454,17 +530,17 @@
- -
+ +

Research Evolution

- From a port number incident to a production governance architecture, across 800 commits and one year of research. + From a port number incident to Guardian Agents in production — 17 months, 1,000+ commits.

-
+
Oct 2025
Framework inception & 6 governance services
@@ -481,6 +557,14 @@
Jan 2026
Research papers (3 editions) published
+
+
Feb 2026
+
Sovereign training, steering vectors research
+
+
+
Mar 2026
+
Guardian Agents deployed, beta pilot open
+
diff --git a/public/locales/en/homepage.json b/public/locales/en/homepage.json index ae41d475..305cba55 100644 --- a/public/locales/en/homepage.json +++ b/public/locales/en/homepage.json @@ -1,4 +1,17 @@ { + "whats_new": { + "badge": "March 2026", + "heading": "What’s New", + "card1_label": "New Research", + "card1_title": "Guardian Agents and the Philosophy of AI Accountability", + "card1_desc": "How Wittgenstein, Berlin, Ostrom, and Te Ao Māori converge on the same architectural requirements for governing AI in community contexts.", + "card2_label": "Deployed", + "card2_title": "Guardian Agents in Production", + "card2_desc": "Four-phase verification using mathematical similarity, not generative checking. Confidence badges, claim-level analysis, and adaptive learning — all tenant-scoped.", + "card3_label": "Beta Open", + "card3_title": "Village Beta Pilot", + "card3_desc": "Village is accepting beta applications from communities ready to participate in constitutional AI governance. Community Basic from $10/mo." + }, "hero": { "title": "Architectural Governance for AI Systems", "subtitle": "Some decisions require human judgment — architecturally enforced, not left to AI discretion, however well trained.", @@ -14,6 +27,7 @@ "corollary_p1": "In code, this bias produces measurable failures — wrong port, connection refused, incident logged in 14.7ms. But the same architectural flaw operates in every AI conversation, where it is far harder to detect.", "corollary_p2": "When a user from a collectivist culture asks for family advice, the model defaults to Western individualist framing — because that is what 95% of the training data reflects. When a Māori user asks about data guardianship, the model offers property-rights language instead of kaitiakitanga. When someone asks about end-of-life decisions, the model defaults to utilitarian calculus rather than the user’s religious or cultural framework.", "corollary_p3": "The mechanism is identical: training data distributions override the user’s actual context. In code, the failure is binary and detectable. In conversation, it is gradient and invisible — culturally inappropriate advice looks like “good advice” to the system, and often to the user. There is no CrossReferenceValidator catching it in 14.7ms.", + "corollary_summary": "The same mechanism operates in every AI conversation. When a user from a collectivist culture asks for family advice, the model defaults to Western individualist framing. When a Māori user asks about data guardianship, the model offers property-rights language. Training data distributions override user context — in code the failure is binary and detectable, in conversation it is gradient and invisible.", "corollary_link": "Read the full analysis →", "closing": "This is not an edge case, and it is not limited to code. It is a category of failure that gets worse as models become more capable: stronger patterns produce more confident overrides — whether the override substitutes a port number or a value system. Safety through training alone is insufficient. The failure mode is structural, it operates across every domain where AI acts, and the solution must be structural." }, @@ -31,8 +45,16 @@ "download_pdf": "Download: The Philosophical Foundations of the Village Project (PDF)" }, "services": { - "heading": "Six Governance Services", - "subtitle": "Every AI action passes through six external services before execution. Governance operates in the critical path — bypasses require explicit flags and are logged.", + "heading": "Governance Architecture", + "subtitle": "Six governance services in the critical path, plus Guardian Agents verifying every AI response. Bypasses require explicit flags and are logged.", + "guardian_title": "Guardian Agents", + "guardian_badge": "NEW — March 2026", + "guardian_desc": "Four-phase verification using embedding cosine similarity — mathematical measurement, not generative checking. The watcher operates in a fundamentally different epistemic domain from the system it watches, avoiding common-mode failure.", + "guardian_p1": "Response Verification", + "guardian_p2": "Claim-Level Analysis", + "guardian_p3": "Anomaly Detection", + "guardian_p4": "Adaptive Learning", + "guardian_cta": "Full Guardian Agents architecture →", "boundary_desc": "Blocks AI from making values decisions. Privacy trade-offs, ethical questions, and cultural context require human judgment — architecturally enforced.", "instruction_desc": "Classifies instructions by persistence (HIGH/MEDIUM/LOW) and quadrant. Stores them externally so they cannot be overridden by training patterns.", "validator_desc": "Validates AI actions against stored instructions. When the AI proposes an action that conflicts with an explicit instruction, the instruction takes precedence.", @@ -45,7 +67,8 @@ "badge": "Production Evidence", "heading": "Tractatus in Production: The Village Platform", "subtitle": "Village AI applies all six governance services to every user interaction in a live community platform.", - "stat_services": "Governance services per response", + "stat_guardian": "Guardian verification phases per response", + "stat_services": "Governance services in the critical path", "stat_months": "Months in production", "stat_overhead": "Governance overhead per interaction", "cta_case_study": "Technical Case Study →", @@ -95,7 +118,7 @@ }, "timeline": { "heading": "Research Evolution", - "subtitle": "From a port number incident to a production governance architecture, across 800 commits and one year of research.", + "subtitle": "From a port number incident to Guardian Agents in production — 17 months, 1,000+ commits.", "oct_2025": "Framework inception & 6 governance services", "oct_nov_2025": "Alexander principles, Agent Lightning, i18n", "dec_2025": "Village case study & Village AI deployment", @@ -104,7 +127,11 @@ "date_oct_2025": "Oct 2025", "date_oct_nov_2025": "Oct-Nov 2025", "date_dec_2025": "Dec 2025", - "date_jan_2026": "Jan 2026" + "date_jan_2026": "Jan 2026", + "date_feb_2026": "Feb 2026", + "feb_2026": "Sovereign training, steering vectors research", + "date_mar_2026": "Mar 2026", + "mar_2026": "Guardian Agents deployed, beta pilot open" }, "claims": { "heading": "A note on claims", diff --git a/scripts/deploy.sh b/scripts/deploy.sh index b61a987a..d91dddc5 100755 --- a/scripts/deploy.sh +++ b/scripts/deploy.sh @@ -147,6 +147,23 @@ if ! node scripts/check-file-permissions.js public > /dev/null 2>&1; then fi echo -e " ✓ File permissions checked" +# CSP check: no inline scripts in HTML files +echo -e " Checking for inline scripts (CSP compliance)..." +INLINE_SCRIPTS=$(grep -rn '