The Evidence Changed the Plan

A productive fortnight can bring new publications and software previews. It can also produce a narrower, better-supported research claim—or a decision not to add another language or system.

This issue covers new Kindle and print editions, public developer previews of Design Space Calculus and Novum, research into AI reasoning and semantic representation, and work on trustworthy completion and software qualification. Across these developments, the recurring question is not simply what can be produced, but what the evidence supports.

Books#

New Kindle editions#

• Always-On AI, published October 1, examines what changes when an AI agent keeps working after you leave.
• Mandala of Agency, published October 2, extends Structures of Governed Intelligence with an account of actors, capabilities, authority, permissions and responsibility.
• What Should Remain Yours?, published October 6, asks how to delegate work to AI while preserving human judgment and responsibility.
• When AI Forgets, published October 10, offers a practical approach to keeping context, correcting old assumptions and carrying continuing work forward.

More books in print#

The print catalog has expanded with paperback and hardcover editions of Permission to Act and The Bottleneck Moves. Always-On AI and What Should Remain Yours? also gained both print formats on October 6, giving readers more ways to explore continuing AI activity, legitimate authority, human participation and progress through a changing workflow.

• Permission to Act: paperback · hardcover
• The Bottleneck Moves: paperback · hardcover
• Always-On AI: paperback · hardcover
• What Should Remain Yours?: paperback · hardcover

Software previews#

Design Space Calculus and Novum reach public developer-preview releases#

Design Space Calculus Core v0.2.0rc1 and Novum v0.2.0rc1 were released as GitHub prereleases on October 3. They give developers inspectable source and bounded Python 3.12 preview packages. These are developer previews, not claims of general production qualification or proven research efficacy.

DSC Core release · Novum release

Research#

Better epistemic integrity does not necessarily mean better solutions#

In the structured-representation invention-reasoning experiment, 144 solver outputs across twelve problems and three conditions yielded 432 blinded AI-evaluator judgments. The representation scaffold increased measured Epistemic Integrity by 0.701 points on a four-point scale relative to ordinary reasoning, but did not establish an improvement in the primary Solution Quality measure. An incremental advantage from transformation instructions and a Novelty advantage were also not established.

The result supports a bounded benefit in how reasoning discloses uncertainty and limitations, not a generalized improvement in invention quality.

Read the research

A new language must earn its complexity#

A comparison of Mandala Core Calculus with direct Cog found some structural and editing conveniences in the proposed notation, but insufficient evidence to justify a separate language. The engineering determination was COG_PREFERRED: preserve Cog as the execution substrate and explore useful semantic geometry through schemas and domain helpers.

Read Semantic Geometry over Cog

Both investigations demonstrate a useful discipline: the experiment must be allowed to narrow the claim and change the development plan.

Read the essay

Systems and accountability#

Audit Workbench qualification progresses#

Audit Workbench's 0.2.7 candidate added Companion-assisted installation, audit follow-up, and source-context safeguards. Repository qualification reported 1,084 passing tests alongside scoped macOS arm64 Fast and controlled Deep audit testing. These results support the tested candidate and environment; they do not establish broad compatibility or completed commercial delivery.

Audit Workbench

A listed product is not necessarily available for purchase#

WinMedia's Machine Edition and Certified Edition catalog continues to distinguish package identity from actual checkout and delivery availability. The Shop's eight Certified Edition listings were recorded as checkout-paused in the prior review. Commercial availability must be described from current runtime evidence, not inferred from product listings.

WinMedia Shop

The task succeeded. The work failed.#

An assistant can produce the requested artifact while violating the conditions that made delivery acceptable. Completion requires evidence about permissions, destinations, side effects, and outstanding obligations—not simply an assertion that the task is done.

Read the essay

Closing perspective#

A public preview is not broad qualification. Better reasoning disclosure is not necessarily better problem solving. Another abstraction is not always an improvement. And finishing a task does not necessarily finish the work.

The value of a careful research and publication program lies partly in keeping those distinctions visible.