01CORE
Constitution Engine
Applies non-negotiable governance rules above model reasoning, including the authority boundaries and conditions that must be satisfied before consequential action can proceed.
RUNTIME AI GOVERNANCE // CONTROLLED PILOT CANDIDATE // W1: 1,501 TESTS REPORTED PASS
Fresh Direction
Runtime governance for AI agents at the point of action. Fresh Direction is a provider-neutral governance layer designed to sit between AI reasoning and consequential action — applying authority, policy, evidence and human approval before an AI can affect real systems.
Models propose. The Governor controls whether the action may execute. The demonstrated path separates proposal from authority, uses a hard execution boundary, binds authority to the governed action, and retains durable evidence for restart, recovery and verification.
Current test evidence (16 October 2026): the coding agent reported 1,501/1,501 full-suite tests passing, including 17/17 focused security tests. The earlier, separate safe-suite checkpoint was 1,423 passed, 2 skipped and 0 failed (1,425 collected). Neither report constitutes independent verification. Fresh Direction remains a Controlled Pilot Candidate, not production-ready, and no numeric whole-product TRL is asserted while Constitution lifecycle and operational proof remain open.
02GOVERN
Authority Boundary
Checks who or what is acting and keeps permissions within the authority actually granted. An AI request cannot create or expand its own authority.
03CONTROL
Human Oversight
Requires evidence-rich human approval where policy, impact or risk means an automated decision is not enough.
04CONTROL
Evidence & Assurance
Uses durable evidence and traceability to support decisions, challenge uncertainty and make governance reviewable after the event.
05VERIFY
Controlled Access
Limits approved AI activity through a hard execution boundary and action-bound authority rather than broad standing access.
06VERIFY
Live Monitoring
Keeps important activity visible while it is happening so unexpected behaviour can be detected and escalated.
07SAFE
Intervention & Safe State
Blocks or contains consequential activity when required authority, evidence or integrity state is missing or cannot be verified.
08ASSURE
Audit & Recovery
Preserves decision, receipt and recovery evidence so organisations can understand what was authorised, what executed and what followed.
Current product statusCONTROLLED PILOT CANDIDATE
The latest readiness audit supports bounded controlled-pilot use, while higher readiness claims remain locked until the Constitution lifecycle gap is resolved.
Latest reported W1 tests1,501 / 1,501 PASS (REPORTED)
16 October 2026 coding-agent report: full suite 1,501/1,501 passed, including a 17/17 security-test subset. These results have not been independently re-run or verified.
Current TRL publicationNUMERIC CLAIM WITHHELD
The previous TRL 6 public reassessment is no longer asserted as the current whole-product claim. The latest audit requires authoritative Constitution activation, reconstruction and rollback evidence before a numeric TRL is republished.
Independent assurance / releaseNOT VERIFIED · NOT PRODUCTION-READY
W1 Stage 1a is agent-reported patched/tested. Stage 1b and mandatory non-destructive attack-lab verification remain outstanding, alongside Constitution lifecycle and Gate 10 proof.
Public information is intentionally presented at governance-capability level. Test results and readiness status are HelpfulCo internal engineering evidence, not external certification. Detailed implementation, decision logic, evidence schemas and enforcement mechanisms remain proprietary.
THE GOVERNANCE PROBLEM
AI can reason. It must not decide its own authority.
The hard problem is not whether AI can reason; it is whether reasoning can quietly become permission. Fresh Direction separates what an AI proposes from the authority, constitutional rules, organisational policy, evidence and human judgement required before consequential real-world actions are allowed to proceed.
Identify
Establish who or what is requesting an action and what authority exists before consequential activity is allowed to proceed.
Assess
Apply constitutional rules, organisational policy, risk, evidence and human-oversight requirements to the proposed action and its likely impact.
Govern
Constrain approved activity to the authority, purpose and scope that have actually been permitted rather than assuming broad standing access.
Verify
Monitor important activity, preserve decision evidence and check outcomes so governance continues beyond the initial approval decision.
PLAIN ENGLISH
Runtime governance at the point of action.
Fresh Direction is designed to sit between AI capability and consequential action. Rather than stopping at policy or reporting, the Governor is intended to apply authority, evidence, risk and human-approval controls where an AI request could affect a real system, then keep the resulting activity visible and auditable.
AI proposes
An AI system or agent can reason, plan and suggest an action, but the request does not automatically create permission to affect a real system.
The Governor checks
Fresh Direction checks identity, authority, constitutional rules, organisational policy, risk, evidence and whether human approval is required.
Action is controlled
If the required controls are satisfied, authority stays bounded to the approved purpose. If authority or evidence is missing, or policy is breached, the action is blocked or escalated rather than assumed safe.
01RequestAI proposes a consequential action
→02Authorityconstitution · identity · scope · policy
→03Decideevidence · risk · human approval where required
→04Execute & verifybounded access · monitor · audit · outcome
Updating a live system
An AI asks to make a consequential change. Governance checks whether the action is authorised, what could be affected and whether additional human approval is required before it can proceed.
External content tries to influence an agent
Information from webpages, email, documents or other external sources is not automatically treated as permission to extend an AI system's authority.
Detecting and stopping AI swarms
Fresh Direction is designed to detect unauthorised coordination between AI agents — including covert communication, shared goals, collective task drift and attempts to rebuild blocked coordination — then deny, isolate or stop consequential actions at the execution boundary before they reach real systems.
An approved action materially changes
A significant change in purpose, target, scope or behaviour should not silently inherit the authority granted to a different action. Governance can require a new decision or intervention.
WHAT FRESH DIRECTION GOVERNS
The controls around consequential AI action.
Public information is deliberately kept at capability level. The aim is to make the governance boundary and engineering evidence clear while keeping detailed decision logic, internal component design and enforcement mechanisms proprietary.
- AI and agent identity, authority and accountability
- Non-negotiable constitutional rules and organisation policy
- Human oversight and approval for consequential actions
- Evidence, traceability and decision assurance
- Controlled access to tools, systems and data
- Multi-agent swarm coordination, cumulative authority and task drift
- Live monitoring, intervention and safe-state behaviour
- Post-action audit and outcome verification
Core principle: AI may propose consequential actions, but it should not be able to create, expand or approve its own authority merely by requesting it.
ENGINEERING FOUNDATION
Built from working supervision and governance engineering, not a policy-only concept.
Fresh Direction began with a measurable supervision environment where evidence integrity, reassessment, recovery, drift detection, controlled change and audit had to work against real system state.
The Governor extends those disciplines into a provider-neutral governance system for AI systems, agents and automations. Its Constitution Engine and authority controls are intended to keep non-negotiable governance rules, human authority and independent evidence separate from model reasoning.
View the original proving ground →ENGINEERING EVIDENCEW1 AGENT-REPORTED TESTING · CONTROLLED PILOT CANDIDATE
On 16 October 2026 the coding agent reported a W1 Stage 1a full-suite result of 1,501/1,501 passed, including 17/17 focused security tests (a subset, not additional tests). This has not been independently reproduced. An earlier, separate safe-suite checkpoint collected 1,425 tests, with 1,423 passed, 2 skipped and 0 failed. Gates 5–9 have retained evidence, but the latest known readiness position keeps higher claims RED / LOCKED. The Constitution activation, reconstruction and rollback path remains open. Internal test reports are not certification, production acceptance or external validation.
- W1 Stage 1a, agent report dated 16 October 2026: 1,501/1,501 full-suite pass, including 17/17 security tests · not independently re-run
- Earlier safe-suite checkpoint (separate run): 1,425 collected · 1,423 passed · 2 skipped · 0 failed
- Stage 1a: PATCHED / AGENT-TESTED · Stage 1b: PREPARATION ONLY · W1 offline attack-lab proof: PENDING
- Current product status: CONTROLLED PILOT CANDIDATE
- Current readiness publication: RED / LOCKED for higher-readiness claims
- Gates 5–9 have retained evidence, including restart/recovery and bounded external-action evidence
- Constitution activation/reconstruction/rollback remains the blocking evidence gap
- Gate 10 integrated operational-governance proof remains pending
- Independent verification: NOT COMPLETE · Production readiness: NOT APPROVED · No unrestricted-autonomy claim
TEST LEVELS · EVIDENCE AND LIMITATIONS
What has actually been demonstrated, and what remains to prove.
A reported test pass is not the same as independent verification or a production release. This ladder separates design, implementation, internal testing, adversarial assurance and final readiness. The latest W1 results are attributed to the coding agent rather than independently certified.
Designed
Authority boundaries, constitutional rules, evidence and recovery requirements have documented engineering designs. Design is not proof that all controls resist attack.
Implemented
W1 Stage 1a implementation is reported patched. Stage 1b implementation and remaining release requirements are not claimed complete.
Internally tested
16 October 2026: 1,501/1,501 full-suite tests passed, including 17/17 security tests, according to the coding agent. No independent re-run has verified these figures.
Adversarially challenged
Earlier bounded gateway and recovery evidence exists. Mandatory offline, non-destructive W1 attack-lab testing for bypass, approval, MFA, tampering, replay and fail-open behaviour remains pending.
Independently verified
Independent reproduction, evidence review and representative-environment validation are outstanding. Internal regression success must not be presented as certification.
Production approved
Constitution lifecycle and rollback, Gate 10 operational proof, security residuals and release gates must close before production-readiness can be claimed.
Evidence snapshot: 16 October 2026. The 17 focused security tests are part of the reported 1,501-test suite, not an extra 17. Earlier 1,425-test safe-suite results belong to a separate checkpoint. No independent certification, numeric whole-product TRL or production readiness is claimed.
AI CONTROL ROOM
The working operator console makes governance visible.
The Control Room gives human operators a view of governed requests, approvals, risk, evidence, audit state, health and governance alerts. It is a supervision and decision surface rather than a substitute for the underlying authority controls.
Governed decisionsImportant AI requests are presented with relevant action details, risk, policy and approval context.WORKING
Evidence & auditDecision-linked evidence and outcomes remain available for review where authoritative records exist.CONTROL ROOM
Health & alertsOperational and governance state can be surfaced to human operators without inventing reassuring data where state is unknown.WORKING
Fail-closed behaviourLoss of trustworthy authority information is treated as a reason to restrict consequential activity rather than assume permission.GOVERNED
GOVERNANCE IN PRACTICE
Policy, authority and human judgement stay separate from model reasoning.
A supervised AI can provide reasoning and evidence, but Fresh Direction is designed so that the model does not become the final authority over its own consequential actions. Organisational policy, risk, evidence and human oversight remain independent governance functions.
Identity & authority
Establish who or what is acting and whether the requested action falls within authorised scope.
Policy, risk & evidence
Assess the request against organisational boundaries, likely consequence and the evidence available to support the decision.
Human oversight & audit
Require human involvement where appropriate and preserve enough evidence to review both the decision and the resulting outcome.
IMPLEMENTATION STATUS
Controlled Pilot Candidate. Numeric TRL claim withheld pending Constitution lifecycle evidence.
C1
Governance core
Provider-neutral governance separates AI reasoning from independent authority, constitutional rules, organisational policy, evidence, human oversight and controlled action.
C2
W1 internal tests (reported 16 October 2026)
The coding agent reported 1,501/1,501 full-suite tests passing, including 17/17 focused security tests. A prior, separate safe-suite run recorded 1,423 passed, 2 skipped and 0 failed. Neither result is independent verification.
C3
Operator visibility
The working Control Centre presents governed requests, exact action details, human decisions, risk, evidence, audit state, system health and governance alerts to human operators.
C4
Controlled pilot candidate
The latest readiness audit supports bounded controlled-pilot work. Higher readiness claims remain locked while the authoritative Constitution activation, reconstruction and rollback path is unresolved.
C5
Next proof
Complete independent reproducibility checks, a non-destructive W1 attack lab, Constitution lifecycle closure and Gate 10 operational-governance proof. A numeric whole-product TRL will be republished only when evidence supports it.
ENGINEERING STATUS · TECHNOLOGY READINESS
Current publication position: internal tests reported; whole-product TRL not asserted
The current W1 Stage 1a coding-agent report (16 October 2026) states 1,501/1,501 full-suite pass, including 17/17 focused security tests, without an independent re-run. A separate earlier safe-suite checkpoint recorded 1,423 passed, 2 skipped and 0 failed. These figures must not be added together or described as independently verified. The latest known readiness audit supersedes the earlier public TRL 6 reassessment: Fresh Direction remains a Controlled Pilot Candidate, not an approved production system. Constitution activation, reconstruction and rollback, Stage 1b follow-through, mandatory non-destructive W1 adversarial testing and Gate 10 integrated proof remain incomplete. Higher-readiness claims remain RED / LOCKED; no numeric whole-product TRL is asserted.
License, acquire or pilot Fresh Direction.
Fresh Direction is available for strategic licensing, OEM/integration discussions, acquisition discussions and carefully bounded pilot programmes. The core remains proprietary and current maturity claims remain deliberately conservative until the required evidence exists.