Methodology

FractaVolta is an engineering firm and a stack operator. We design, build, and run infrastructure across four layers — energy, compute, cognition, civic. The methodology described on this page is not separate from the engineering. It is how the engineering is done.

Three concrete commitments shape every project:

Every claim is anchored. When we say "the LFP container delivers 1.9 MWh of useful work" or "the inference packet executes in under 30 seconds on a 50W node," there is a canonical document with a stable URL where that claim lives. We don't sell brochures of unsupported numbers.
Every objection becomes a first-class contribution. If you disagree with a claim — file an issue. We will either convert it into a falsifiable claim, mark it as a feeling-of-certainty (with attribution), or request the specifics that would settle it. Disagreement is part of the record, not noise to be erased.
Every AI agent is replaceable. When we use Claude, ChatGPT, a local model, or a human reviewer for any decision in the deployment, the work is captured as a typed continuation. A different agent — or a different human — can pick up the same work from the same point, without modifying our tooling. No vendor can lock us in. No vendor can lock you in.

These three commitments are operationalized through a small set of named instruments.

Cogentia Commons

Cogentia Commons is the methodology for distributed knowledge production across the corpus. It is GitHub-anchored — every research-grade document carries a canonical_url in its front-matter pointing at its authoritative location on github.com. The corpus is its own evidence: forks, objections, and revisions are all visible in the git history.

cogentia.js

cogentia.js is the operational CLI. Zero-dependency Node.js. Multi-repository registry. Commands include:

scan — flag markdown files not referenced in research/index.md (rule 4: every claim must be anchored).
check — validate internal links across the corpus.
stamp — insert canonical_url + last_stamped_at into a document's front-matter, anchored to its GitHub commit URL.
corpus-status — refresh the per-repository status page; structural sections auto-regenerated, manually-curated sections preserved.
continuation — emit / inspect / resume / fail / abort / queue typed continuation requests. See below.

Every state-changing call appends one JSONL line to .cogentia/audit.jsonl. The audit log is a first-class deliverable, not a compliance afterthought.

The continuation protocol

cogentia.continuation.v1, defined in the paper Agent-Resumable CLI and implemented in cogentia.js as of v0.5.0, is the typed protocol for suspended judgment.

When the CLI reaches a point where deterministic computation is insufficient, it does not call an AI API. It emits a continuation: a JSON object carrying the context, the alternatives, the expected result schema, and the resumption metadata. The surrounding system — a human reviewer, an AI agent, a script, a CI pipeline — supplies a structured step_result. The tool validates the response and resumes.

The soundness test is binding:

Can the current AI agent be replaced by another agent or by a human, without modifying cogentia.js? If yes, the protocol is sound. If no, it is contaminated.

Every resumption produces a Heraclitean follow-up — a dormant successor that keeps the chain non-terminal. History never 100% ends; the only constant is change.

The Second Method

The methodological foundation of the entire corpus is the Discours de la seconde méthode. Five rules:

Rule 0 — Encode the boundary in the architecture. AI agents may participate in epistemic production (drafting, summarizing, mediating). Political governance — binding decisions — belongs to living humans alone.
Rule 1 — Publish the process, not only the result. The audit log, the continuation log, the git history are part of the deliverable.
Rule 2 — Make every objection a first-class contribution. Falsifiable claims, marked feelings-of-certainty, requests for specifics — all are typed, all are recorded.
Rule 3 — Structure for machine readability from the start. Stable identifiers, schemas, canonical URLs, append-only audit logs.
Rule 4 — Let the corpus be its own evidence. The repositories demonstrate the method on themselves.

What this means in practice

If you are a partner. You don't have to take our word for what we delivered. The audit trail of the deployment is part of what you receive. If we ever can't be found, the next operator can pick up exactly where we left off — every decision recorded, every alternative considered, every objection filed.

If you are a deployer. Your installation produces its own typed audit trail. You can run cogentia.js scan against your site's documentation to verify it matches the canonical specification. Drift becomes visible.

If you are a researcher. The corpus is open. Fork it. Contest a claim by filing an issue — it will be converted into a falsifiable form, marked as unfalsifiable with attribution, or escalated for more context. Your objection becomes part of the public record. Methods: see the for-researchers page.

If you are a citizen using a deployed Inseme instance. The AI mediator Ophélia is bound by the same protocol. Her contributions are typed events; her reasoning is replayable; she is replaceable. The deliberation belongs to the assembly — not to a vendor, not to her.

How to engage

File issues at any of the corpus repositories listed in the ecosystem footer. They are read.
Install cogentia.js locally — it's MIT-licensed, zero-dependency Node.js. The README is the entry point.
Read the Cogentia Commons Working Paper for the methodology's formal specification.
Read Agent-Resumable CLI for the continuation protocol's design rationale.

The methodology is offered. It is not imposed. Operations that want compliance theater, opaque audit chains, or vendor-locked cognition will not find FractaVolta a useful partner.