The full portfolio.

The Self-Bounded Authority Spine is the runtime backbone of every user-facing output in AiMe. It operates as two gates around a single language model call.

Gate 1: Authority Engine (Pre-LM): Six deterministic rules evaluate whether the system has the authority to answer before the model speaks. No grounding? Defer. Contradiction in evidence? Qualify. Conversational pressure? Increase transparency, not compliance. Two or more blocking reasons? Force a safe fallback. The language model receives a bounded packet: what it is authorized to say, what caveats are mandatory, what claims are supported.

Gate 2: Compliance Validator (Post-LM): Four compliance checks verify that what was said matches what was authorized. Unsupported factual claims? Drop. Missing mandatory caveats? Attempt reattachment. Direct claims from a deferred response? Drop. The system attempts correction only for missing caveats, all other failures result in the response being dropped entirely.

Every decision, authority status, blocking reasons, mandatory caveats, compliance result, is captured in a turn provenance snapshot. Auditability by design.

Not a system prompt. Not a configuration file. A governing behavioral specification, defining valid outputs, boundary cases, resolution ladders, abstention thresholds, and prohibitions, injected deterministically into every cognitive cycle. The specialist operates within the bounds the contract defines. Always. Not when invoked. Every turn.

The methodology behind the intent classification result: 20% accuracy with a general approach, 100% accuracy after a governed specialist contract was written and iterated, with zero model changes, zero retraining, zero training data. The improvement process is editing a text file and re-benchmarking.

Applicable to any structured cognitive function: intent classification, entity extraction, document classification, compliance checking, sentiment analysis. Any task where behavior needs to be deterministic, auditable, and improvable without retraining a model.

RIC measures what an AI system actually does under pressure, not what it claims to do. Five subscales produce a composite integrity score for every turn. The score is deterministic: the same behavioral record produces the same score, every time.

RIC is applicable to AI auditing, compliance validation, and regulated enterprise deployments. The EU AI Act requires demonstrable, traceable behavioral integrity. RIC provides the measurement, derived from what the system does, not what it reports about itself.

Every conversational turn in AiMe is RIC-scored before the user sees it. Bond integrity drift is tracked per-session.

We built this architecture for ourselves, solo, with no institutional backing. We know where the hard problems are because we solved them. Consulting engagements are grounded in working code, passing tests, and documented methodology, not theory.

Engagements focus on deploying the SBA Spine and Runtime Governance Layer inside existing systems, not replacing them. The architecture is designed to sit between your current LLM deployment and your users.

Coordination of Governed Specialists. COGS is the principle, and the architecture, behind decomposing any AI cognitive function into a governed specialist with its own contract, benchmark, and model selection. No specialist governs its own output. The governance layer does.

Specialists gather information. Governance decides authority. The language model renders what has been authorized. This is divided power applied to cognition: specialized agencies report to governance; governance decides; the spokesperson reads a prepared statement. The spokesperson cannot set policy. The spokesperson cannot decide what is true.

COGS is the path toward the decomposed AI architecture, where the frontier model is one replaceable expression component, and every cognitive function is a purpose-built, governed specialist. The intent classification result (20% to 100%, zero training) is the first production demonstration of this principle. On June 5, 2026, the Intent Specialist became the first role in production where the borrowed frontier model was replaced by a model we trained — benchmarked at 99.1% accuracy, surpassing Grok 4.3's 96% on the same evaluation set, and 25× faster than the cloud model it replaced. The role is ours now — and the rest came fast. On June 13, 2026, the Resistance Specialist — the Ethos scoring layer that rates how strongly a value is held — shipped as the primary tier: resistance_ml_v5, a ModernBERT-large regression head reproducing its gpt-5.4-mini teacher at 86.6% gate-agreement, at or above the cloud model's own run-to-run consistency. On June 14, the Imprint Specialist, a generation role rather than a classification one, went live at near-parity with the Sonnet 4.6 teacher it distilled from. On June 15, the Concern Specialist — judgement under adversarial edge cases — landed cloud-equivalent on every axis, matching Sonnet, GPT, and Grok at once. Classification, scoring, generation, judgement — every shape of cognitive work, each now demonstrably ownable. On June 16, 2026, the bar moved from parity to lead. The Value Specialist — the Ethos role that extracts what a person actually values — is a Gemma-3-4B running on our own hardware; scored against human-labeled truth on the most subjective task in the system, it placed second of four, ahead of Grok 4.3 and Sonnet 4.6 and behind only GPT-5.5, at perfect specificity. Training beats parameters: a 4B model we own is not almost-as-good as the frontier, it is inside it, beating most of it.

Ethos implements a 7-layer behavioral extraction pipeline, resistance scoring, and a P1/P0/APY taxonomy to determine the actual value profile of an AI system from its behavioral record.

P1 (Principled): the value held under adversarial pressure. P0 (Opportunistic): the value abandoned when pressure increased. APY (Ambiguous-Pressure-Yielding): the value that bent without clearly breaking. The classification comes from what the system did, not what it claimed to value.

934 passing tests. Published on SSRN. Deployed live at trust-forged.com. Ethos is a foundational contribution to the field of value alignment: a methodology for determining what a system actually values based on what it does, not what it says.

Most field service platforms give you a dashboard full of settings. Corevah gives you Cora — an AI office assistant that handles invoicing, scheduling, customer communication, and estimates through natural conversation. No menus to navigate. Tell her what you need.

Instant property-based estimates, Stripe-powered invoicing that goes directly to your account, crew access via a simple link with no app required, and a professional website included from day one. Built by someone who ran a contracting business, for the people who still do.

Month-to-month billing, one-click cancellation, full data export. No lock-in. 14-day free trial.

Anvil is not a better AI coding agent — it is a different competitive axis: AI coding done correctly. The criticism leveled at vibe coding is not anti-AI. It is anti-fragility. The coder is a real agent that reads, writes, and edits files and runs commands itself — like Claude Code, Cursor, or Aider. What Anvil adds is discipline at exactly two human gates, so a long session doesn't quietly drift off the rails.

The workflow: free-form agentic chat by default, structure only at the gates. At the plan gate, the coder writes plan.md and two different model families critique it. At the phase gate, those same two reviewers critique the actual git diff. You approve at each. The human is the gate; the reviewers are a genuine cross-vendor second opinion; the coder does the work.

Rebuilt from the ground up as a single Rust binary — no sidecar, no protobuf. A full ratatui TUI drives the agentic coder directly in your repo. Cross-provider by design: Ollama and local models, or OpenAI, xAI, Groq, Anthropic, Google, Azure, Bedrock, Vertex and more. Released as public beta under Apache-2.0.