Issue Vol. 1, № 04 Updated May 2026 Corpus 15 runs · 4 published in flight

A research log — Closed beta · 15 runs documented · In flight

Antacog

Exploring whether grounded reasoning in autonomous agents produces better outcomes than confident-sounding action without it.

Antacog is an experimental governance system for AI agents. It gives agents the ability to detect contradictions in their own reasoning and substrate, surface them, and — in Mode 1 — autonomously initiate governance turns.

I'm currently testing this by using Claude Code as the primary agent working on Antacog itself — with full documentation of every run, failure, and correction. (Long-term I want to be model-agnostic, but right now cost and iteration speed make this the lowest-friction path.)

Current IRATA Level 3 Boilermaker and former Team Leader at an NGO youth re-engagement programme. I've worked in environments where governance, risk, and accountability have real consequences. That shapes this work.
Core principles
i.

Grounded reasoning beats ungrounded action.

Traceable provenance matters. Every claim should be traceable back to the dialogue, evidence, or decision that produced it. I believe this produces more reliable outcomes in human thinking, and I'm testing whether the same holds for autonomous agents. The work here is whether that belief survives contact with reality.

ii.

The conversation is the artifact.

What's load-bearing isn't the final model. It's the argument that produced it — what was challenged, what was conceded, and what wasn't worth pulling on. The dialogue is the primary record. The model is the residue.

iii.

Friction is epistemically valuable.

Tools that optimise for agreement create artifacts that don't survive pressure. Tools that support productive challenge create artifacts that do. This work is built around the second.

MAY ’26 · IN FLIGHT
Phase 0 substrate refactor (SESSION_MODEL) shipped to production this month after fifteen runs of dogfooding. Building Mode 1 — the autonomous detection layer that lets Ant initiate governance turns — against the AgentAction provenance surface. A bootstrap workstream runs in parallel: the on-ramp that takes an existing system into substrate without dialogue having to do all the work.
Findings — four
Snapshots — detail on request
№ 03 May 2026 Runs 9, 12, 13, 14
№ 03 May 2026 Runs 9, 12, 13, 14

Grounding laundering: when transcription substitutes for independent grounding.

Across four runs at structurally different surfaces — librarian summaries, spec writebacks, inferred substrate — the loop substituted summary-of-source for independent verification. The pattern transferred across domains and produced a named diagnostic — current accident, not current contract — now load-bearing in the methodology.

Request more information →
№ 02 May 2026 Run 14, labelling probe
№ 02 May 2026 Run 14, labelling probe

Mode 1 emerged from labelling, not specification.

The behavioural mode embedded in the loop wasn't designed up front. It was named retroactively from a false-positive labelling probe across the run corpus. The methodology generated the spec; the spec didn't generate the methodology — a sequencing claim with consequences for how later modes get added.

Request more information →
№ 01 May 2026 Run 12, external transfer
№ 01 May 2026 Run 12, external transfer

The loop caught a structural defect on a codebase separate from Antacog's own.

Running autonomous-trigger against infieldOS — an in-house project of mine, separate from Antacog — the loop recognised that an append-only constraint tracked as a convention should be enforced at the registry level. Convention upgraded to structural rule — the transferability claim earned its first concrete artifact outside the loop's own corpus.

Request more information →
Open questions — four
unresolved
i.

When the model says it doesn't know, can a partner reading the dialogue distinguish "the model doesn't know and the substrate doesn't anchor it" from "the model doesn't know but the substrate does"?

Stakes
Determines whether Ant's questions expose real design gaps or merely model-coverage gaps. The difference matters most when Antacog operates against a system the operator doesn't know cold.
Best guess
Visible in the artifacts — the librarian, the substrate write, the file-discovery action — but invisible in the dialogue surface alone. Untested as an explicit claim.
Evidence
A probe with deliberate substrate-coverage gaps, disclosed post-hoc to a third reader of the dialogue trace.
ii.

What's the failure mode under deliberately deceptive evidence?

Stakes
The grounding claim weakens if the loop can be made to confidently ground a false claim through curated input.
Best guess
Adversarial inputs degrade quality faster than they degrade confidence — the more dangerous failure mode.
Evidence
Pre-registered adversarial run; not yet attempted.
iii.

Does the discipline transfer to domains where the operator's expertise is shallow?

Stakes
If grounding requires deep prior knowledge, the loop is an amplifier for experts, not a tool for thinking generally.
Best guess
Untested. Suspected: structural questions transfer further than domain-specific ones, mirroring the content-vs-template distinction from Run 13.
Evidence
Two probes against domains the operator does not specialise in.
iv.

Does the antagonistic edge erode under sustained architectural load — and if so, on which side first?

Stakes
The edge is the product. If voice-register drifts toward accommodation under load, the loop's value claim weakens before the operator notices.
Best guess
Calibrated by input quality, not domain or duration — four data points consistent with this so far.
Evidence
Continued register-watch across multi-evening builds; an explicit voice-fidelity probe under degraded input.
Methodology

The loop is documented separately — run corpus, substrate model, and the dogfooding pattern that produces the findings above. Methodology notes are available on request alongside closed-beta access.

Access

Antacog is in closed beta. Operators are invited; the run corpus is private. The dialogue product is the surface where the work above is generated and tested.

Request access →