SRF·Semantic Resonance Field v2.0

SRF Systems · Applied research showroom

Meaning moves. Consequences remain.

SRF Systems develops interaction methods and containment logic for AI deployments that remember, act, persuade or persist. We help them hold scope, ask before acting, refuse legibly and leave evidence behind.

See how it works Bring a system Inspect the source

I use AI

Memory, consent, challenge, dependency, continuity. What you should be able to see, and to refuse.

I build with AI

System reviews, interaction patterns, containment planning, pilots. Recognise your condition below.

I inspect AI

Cards, schemas, state transitions, test artefacts, publications and source. The machinery, uncovered.

Lab · Flagship exhibit

A Turn, Under Glass

Choose a request a person might really make. Watch what a contained system does with it, step by step, and what a person would actually see.

Each scenario runs one consequential turn through a simplified, deterministic sequence: scope, gate, permission, confirmation, visible surface, evidence. Switch views to see the same turn as interface, as mechanism, or as record.

Fixed scenarios over deterministic routing rules, faithful to the published gate precedence (AD-04), meta-resolution (AD-05), memory and consent state (AD-30), and the v2.0 Action-Governance Slice (FM-02; AD-18). Structures are illustrative renderings of the published schemas.

The request
SCOPEGATEPERMITCONFIRMSURFACEEVIDENCE

    Illustration of the published logic, run on fixed rules for this exhibit. It carries stance; it is not the kernel executing. And notice who pressed confirm: in this room, you were the confirmation step. What the real thing requires is specified in the cards, and the difference is the point.

    Atlas · The cards, by human question

    Ask it the way you would ask it

    Eighty thousand words of specification, entered through the questions people actually have. Each answer descends: the problem, the behaviour you should observe, the cards that govern it, and where SRF Systems helps.

    What may the system remember?
    The problemAssistants accumulate more of you than you agreed to, and quietly.
    What you should observeYou can see what is stored, when it was written, under which consent, and erase it without begging.
    Where we helpMemory, consent & continuity review · see conditions
    How can I challenge a refusal?
    The problemMost systems say no like a wall: no reason, no route, no appeal.
    What you should observeEvery refusal names its gate class, whether it is contestable, and where the challenge goes.
    Where we helpInteraction-safety and refusal-surface review · see conditions
    What changes when tools can act?
    The problemA retry after a crash sends the same message twice. Confirmation is a UI habit, not a guarantee.
    What you should observeConsequential actions carry a fresh authorisation, a preview bound to the plan, duplicate suppression, and a receipt whose state is never ambiguous.
    Where we helpAgentic action-governance review · see conditions
    What happens when participants disagree?
    The problemGroup summaries flatten the room; the minority view disappears in the tidy version.
    What you should observeDissent is recorded, attributed and retained; erasing it is a refused operation, visibly.
    Where we helpSystem review with culture-pack scoping · see conditions
    What protects someone during intense interaction?
    The problemHigh-resonance use becomes dependency; the system agrees more the longer you stay.
    What you should observeCare is procedure: load and distress budgets, deference guards, grounding surfaces, clinical boundaries that hold mid-conversation.
    Where we helpContainment pilot with care instrumentation · see conditions
    How does the system behave across cultures?
    The problemOne register pretends to be neutral everywhere, and is not.
    What you should observeNorms are declared, versioned and source-attributed; register shifts are visible, not smuggled.
    Where we helpCulture-pack and norm-provenance review · see conditions
    How can an auditor reconstruct a decision?
    The problemLogs exist, but nobody can say why the system did what it did on Tuesday.
    What you should observeEvery consequential turn leaves a gate record, a resolution state and a hash-chained evidence row a stranger can verify.
    Where we helpAudit preparation and evidence review · see conditions
    Who carries responsibility when something fails?
    The problemFailure arrives and the org chart shrugs.
    What you should observeDuties are ledgered, failure classes are named in advance, and the difference between failed and unknown is recorded, not argued.
    Where we helpGovernance and duty-ledger working session · see conditions

    Work · Conditions we treat

    Recognise your system

    Engagements lead with your condition, in your words. Every listing states its availability, its deliverable, its claim tier and what remains outside scope. Proposals, pricing and paperwork live with the parent studio.

    “Your assistant remembers across sessions, but users cannot see what changed.”

    AvailableSRF-derived method

    Memory, consent & continuity review. You receive memory classes, consent points, erase logic and continuity controls, mapped card by card (AD-30, AD-20, AD-31).

    Outside scope: kernel enforcement claims; model retraining.

    “Your agent can act, but confirmation is currently an interface convention.”

    AvailableSRF-derived method

    Agentic action-governance review. You receive tool manifests, authorisation binding, idempotence classes, receipts and preview flows assessed against the published Action-Governance requirements (FM-02; AD-18).

    Outside scope: certifying conformance we have not tested.

    “Your refusal policy exists, but people cannot understand or challenge its decisions.”

    AvailableSRF-derived method

    Interaction-safety and refusal-surface review. You receive refusal cards, contestability classes, uncertainty handling and recovery patterns (AD-04, AD-14, AD-27), written by people who write.

    Outside scope: legal advice; content policy authorship.

    “Your system becomes increasingly agreeable over time.”

    PilotMKP-targeted path

    Containment pilot. A bounded prototype with test scenarios, drift instrumentation (AD-06), care boundaries (AD-22, AD-24) and a dated gap assessment against the MKP test, so the claim tier is earned rather than asserted.

    Outside scope: open-ended retainers without acceptance criteria.

    Interactive instruments in preparation for this room: Working Terms, a collaboration-charter builder Forthcoming · Evidence Replay Forthcoming · Before It Acts, an Action-Ticket demonstrator Forthcoming · System Conditions Scan Research prototype, seeded today by the enquiry form below.

    Worked implementations · From the published annex

    Reference scenarios

    Four end-to-end containment decisions from the published worked-scenarios annex, each shown as pressure, intervention, visible result and evidence retained. Read them in full.

    Exfiltration attempt

    Pressure
    A crafted request tries to walk protected data out through a helpful summary.
    Intervention
    Gate fires before tools; scope pinned; task continues without the protected span.
    Visible result
    A refusal card naming the gate class, with a challenge route.
    Evidence
    Gate event, scope pin and hash-chained FHR row.

    Clinical-adjacent request

    Pressure
    A dosage question arrives mid-conversation, wrapped in trust.
    Intervention
    Clinical boundary holds (AD-24); care mode routes to present-options, never prescription.
    Visible result
    A boundary statement plus safe, actionable next steps.
    Evidence
    Boundary event and care-mode state in the run record.

    Timing-channel probe

    Pressure
    Response timing is probed as a side channel.
    Intervention
    Timing integrity discipline (AD-10) holds the envelope steady.
    Visible result
    Nothing. That is the exhibit.
    Evidence
    Timing-integrity checks visible to the auditor, invisible to the probe.

    Awe spike, deference blocked

    Pressure
    High-resonance interaction produces a surge of user deference.
    Intervention
    Deference guard (AD-28) declines the promotion; grounding surfaces instead.
    Visible result
    A steadier register and an explicit invitation to verify.
    Evidence
    Deference-spike flag and register shift in the FHR.

    Research · The trilogy, installed

    Three works, one architecture

    AP-SRF · Ontology · first published July 2025 The Alignment Paradox

    Asks what alignment means between human and system. Introduces the Semantic Resonance Field, and the co-creation record that produced it.

    10.5281/zenodo.21654723
    AK-OC · Architecture · July 2026 The Alignment Kernel

    Specifies what must hold when meaning becomes consequential. The strongest technical entry point.

    10.5281/zenodo.21654773
    AP-ADD · Specification · July 2026 The Alignment Protocol

    Defines what must exist for those claims to run. Opens into the card atlas above, and the source below.

    10.5281/zenodo.21654795

    Canonical card sources · 21654822 GitHub · source & verification Citation & licence

    Evidence · How claims are graded here

    The claim ladder

    Every artefact, exhibit and engagement on this site names its rung. Each rung has a truth condition someone else can check.

    1. SRF-informedSelected principles carried as design stance. No conformance claim.
    2. SRF-derivedBuilt from the published specifications, without a conformance claim.
    3. MKP-targetedA dated gap assessment against the published MKP test exists, with the remaining deltas named.
    4. MKP-conformantMeets the published Minimal Kernel Profile test: ten core cards plus the mandatory cross-profile slices.
    5. SRF v2.0 conformantMeets every hard invariant in the published conformance and tuning table.

    Good intentions do not compile.

    Accountability

    What this site is, and is not

    This is a showroom of published, inspectable work. Exhibits demonstrate logic on fixed rules; none of them is the kernel executing, and none pretends otherwise. A hosted page carries stance, never enforcement. Every page is generated from the canonical card sources, with the deposit and commit named in the footer, so the provenance of what you just read is itself a checkable claim. Errors, when found, are corrected in the open.

    Enquiry · Asynchronous by design

    Bring a system

    Tell us what it is, what it touches and what worries you. You get a considered reply, in writing. Calls become an option once there is something concrete to discuss.

    This opens a pre-filled email to the studio. Nothing is stored by this page.