Research as Website
Working research drafts, specifications, and case studies from HUMMBL, with links to earlier deposited snapshots where recorded.
Research question: How can agent actions be governed through explicit authority, observable controls, and inspectable evidence? Each paper states its assumptions and evidence limits. Draft status does not imply peer review.
Multi-agent delegation requires connecting task scope, authority, execution evidence, and verification across multiple records. We present the Intelligent Delegation Profile (IDP), a six-type governance representation comprising delegation context, contract, capability token, evidence, attestation, and system event records. Cross-links support reconstruction, while evidence-before-verify, capability attenuation, and bounded delegation depth define validation obligations. We describe a briefing workflow and a peer-review mapping as design case studies. Historical operational summaries are retained as author-reported observations whose raw traces and benchmark protocol are not available in this artifact. A new exact-arithmetic companion reconstructs the stated no-decay trust-update model: recovery after one failure grows linearly with prior successful rounds, and does not establish economic resistance to attack. The profile contributes a composition and audit vocabulary; complete mediation, authentic evidence, and implementation conformance require separate assurance.
Multi-agent delegation creates an audit problem: task intent, capability grants, lifecycle transitions, execution receipts, and verification decisions are often recorded in different systems. We propose Typed Governance Tuples v0.3.0, a candidate interchange profile connecting these records through a signed envelope and explicit identifiers. Six core tuple classes represent contracts, delegated capabilities, context, system decisions, evidence, and attestations. The candidate also considers multi-parent task graphs and policy-defined co-approval.
The contribution is a specification of trust boundaries and validation obligations. An independently controlled monitor must mediate the declared side-effect boundary; signatures bind records to keys; verifiers evaluate observed evidence against contracts. These mechanisms do not establish the truth of arbitrary tool responses, the completeness of unobserved activity, or semantic goal containment for unrestricted natural language. The public reference implementation demonstrates an earlier HMAC-based design and does not substantiate all v0.3.0 requirements. This manuscript identifies the remaining protocol definitions and conformance artifacts needed before an interoperable implementation or deployment-security claim is warranted.
Agentic AI systems need interruption mechanisms that are more expressive than binary on/off switches but still simple enough to audit. We present a four-mode safety state machine for agentic runtime interruption: `DISENGAGED`, `HALT_NONCRITICAL`, `HALT_ALL`, and `EMERGENCY`. The machine distinguishes normal operation, graceful degradation, full operational halt with safety tasks preserved, and human-reset-only emergency halt.
We specify the transition function and task-admission policy, verify three safety properties over a complete table of 20 state/input pairs, and compute four reachable modes with maximum shortest-path depth one from normal operation. Eight task-class/state admission decisions distinguish the four required control regimes. We then compose the state machine with adapter-level circuit breakers and compare the resulting control stack to production mechanisms including Kubernetes probes, Envoy circuit breakers, SageMaker deployment guardrails, and Vertex AI rolling deployments.
Our claim is scoped. The four-mode design is sufficient for the stated safety properties under a binary critical/non-critical task taxonomy, and four states are minimal for the four specified admission regimes when admission depends only on state and task type. P1–P3 alone do not require four states. We do not claim that state-machine structure alone proves safe governance. Emergency irreversibility and auditability require supplementary identity, access-control, persistence, and append-only evidence mechanisms.
Base120 organizes 120 named reasoning prompts into six families: Perspective, Inversion, Composition, Decomposition, Recursion, and Systems. It combines a versioned catalog with a Python lookup and prompt-generation SDK, a command-line interface, and a JSONL record format. We describe the taxonomy as a design framework for selecting and documenting reasoning strategies, and audit its correspondence to a pinned software release. That release contains 20 operators per family and no third-party runtime dependencies. Its engine constructs prompts and records externally supplied recommendations; it does not execute or validate the reasoning itself. We distinguish catalog coverage from cognitive completeness, prompt composition from a typed program calculus, and append-mode storage from tamper-evident audit infrastructure. The contribution is a reproducible vocabulary and interface design. Improvements in decision quality, completeness of the taxonomy, and production safety remain empirical questions.
File-producing subagents need a shared specification of where outputs belong and how completion claims will be checked. We propose six prompt sections: output path, naming convention, sibling context, verification step, claim discipline, and return format. The proposal is motivated by a first-party report of three document-producing dispatches whose outputs occupied two directories and used two naming conventions. A reported 979-line completion claim was compared with a PowerShell result of 535, but the incident report also lists a heading at line 869. This inconsistency prevents treating the difference as a verified 83% overcount. A synthetic counterexample confirms that line-counting procedures can disagree when empty lines are handled differently. The resulting contribution is a measurement-aware output contract: receipts should bind a metric definition and measurement procedure to the same artifact snapshot, and a parent should verify that snapshot independently. We provide a portable measurement artifact and a controlled evaluation plan. The case does not establish that the contract reduces error rates, that the subagent fabricated a count, or that a receipt guarantees substantive correctness.
Experiment receipts can preserve the connection between a research proposal, its execution conditions, its measurements, and a keep or discard decision. This paper examines that design in one first-party autoresearch corpus and audits the numerical claims made from it. At the pinned source revision, the results table contains 243 run rows: 55 labelled keep, 186 discard, and two crash. Those counts are not counts of accepted patches. Two initial control receipts each contain five seeds; their observed validation bits-per-byte ranges are 0.180419 and 0.004988 under different configurations. The program records a decision gate based on those controls, but subsequently identifies the proposed QK-Norm patch as already present in the baseline. Five later candidate settings were discarded against a distinct BATCH 2x baseline. The audit also exposes changes in the definition of "noise floor," baseline mismatches, and counting errors in the earlier manuscript. We propose a receipt contract that makes units, measurement definitions, baseline lineage, and decision timing explicit. The case supports the inspectability of recorded decisions, not a causal claim that receipts improve model performance, eliminate missing results, or establish a scale threshold for architectural features.
DRAFT
Agent coordination
•
2026-10-10
•
Reuben Bowlby
An agent coordination bus can preserve work assignments, blockers, review claims, and decisions, but recording a message must not be confused with authorizing an action. This paper specifies a bus as evidence infrastructure, with execution controlled by a separate authorization boundary. We analyze two historical HUMMBL design documents, distinguish their proposed formats from deployment evidence, and identify the additional semantics needed for reliable queue reconstruction: stable work identifiers, ordered event admission, correction links, artifact revisions, and separate activity and progress clocks. An executable offline synthetic example illustrates stale evidence, contradictory claims, and blocker recovery without exposing operational records. We propose a controlled reviewer study and operational measurements with explicit denominators and censoring rules. No operational incident dataset or empirical throughput result is presented. The contribution is an inspectable design and a falsifiable evaluation plan, not a demonstrated improvement in reliability or a new append-only logging mechanism.
DRAFT
Foundations and cybernetics
•
2026-10-10
•
Reuben Bowlby
HUMMBL provides a setting for studying governance expressed as versioned norms, action checks, and decision records. We propose the term *normative cybernetic institution* for an arrangement in which feedback about agent behavior informs revisable institutional rules, some of which are enforced at an action boundary. The contribution is a design argument and a testable separation of normative authority, executable policy, runtime enforcement, and evidence of outcomes. We describe a minimal action-gate model, its assumptions, and an evaluation protocol. Base120 supplies a candidate vocabulary for describing reasoning activities. An exploratory analysis of 120 pinned definitions compares all 15 pairings of its six families: the proposed pairing ranks 13th by mean paired lexical distance and 5th by smallest maximum absolute contrast cosine. Near-orthogonality occurs across alternatives; these text measurements do not establish bipolar cognitive axes. A finite vocabulary can support arbitrarily long expressions, but neither grammatical validity nor an audit log establishes termination, safety, or legitimate authority. The paper makes no claim that the current HUMMBL deployment enforces every norm or improves institutional outcomes. Those claims require implementation-bound tests and independently assessable evidence.
DRAFT
Normative governance
•
2026-10-10
•
Reuben Bowlby
Human Flourishing Accelerationism (HFA) is considered here as a normative research position: improvements in agent capability should be evaluated together with the conditions under which affected people can authorize, understand, contest, and revise agent actions. We propose an explicit translation record between a flourishing-related concern and a narrowly testable runtime constraint. The record distinguishes the concern, local rule, observable predicate, enforcement boundary, failure modes, and unresolved normative judgments. A model of action admission illustrates what such a constraint can establish and why it cannot guarantee flourishing. We offer counterexamples and a prospective comparison against instructions and ordinary permission checks. The contribution is a specification and evaluation proposal, not a claim that HUMMBL is the first system to combine flourishing and runtime governance. No human outcome study, operational effectiveness estimate, or proof of flourishing preservation is reported.
