The Glassary.

The defining characteristic is not experimentation or risk-taking. It is allowing capability to expand faster than evidence, authority, reproducibility and verification.

An AI cowboy may produce excellent work. The problem is that nobody can reliably explain why the work should be trusted.

An agent authorised to edit one directory ends up capable of rewriting the repository. An approval to deploy one component becomes interpreted as approval to deploy everything.

Authority should travel deliberately, not seep through the floorboards.

Everything may appear to be working perfectly until somebody asks, "Who actually authorised that?"

Giving an agent one more tool can sometimes be equivalent to adding an entire new category of action.

Branch ghosts become particularly troublesome in long-running AI projects where agents repeatedly rediscover old work without knowing its status.

A branch ghost is the corpse.

Branch necromancy is when somebody gets it walking again.

Hosted CI is valuable evidence. It should not become a sacred mountain that every tiny change must climb.

The prose is polished. The headings are excellent. Unfortunately, nobody checked whether paragraph three was true.

Some archaeology is inevitable.

If every new agent needs a shovel, the project has a state-management problem.

Like technical debt, context debt feels cheap at the moment it is created.

Someone else pays later.

Frequently that someone is an AI burning thousands of tokens rediscovering Tuesday.

Some sediment contains valuable history.

Some is simply mud.

Unlike context rot, sediment is not necessarily harmful. It becomes a problem when nobody distinguishes geological record from rubbish.

Agents reread repositories, repeat investigations, reproduce tests and reconstruct decisions that already existed somewhere else.

The machines appear extremely busy.

Progress remains strangely stationary.

AI A makes a questionable claim.

AI B repeats it in a polished report.

AI C cites B.

By lunchtime the original guess is wearing a tie.

The latch may require passing tests, independent review, an artifact, a reproducible result or some other explicit acceptance condition.

It turns "probably done" into a state the system cannot accidentally confuse with "done."

Can it discover how the system works, identify current state, use the correct tools and continue safely?

If yes, the project has a healthy operational interface.

If Matthew needs to explain seventeen undocumented facts first, it fails.

Frontier code accepts uncertainty consciously and contains its consequences.

This separates it from cowboy engineering.

The Glass Trail is not merely a log.

It should allow somebody arriving later to reconstruct the important story of the work.

Tests demonstrate what they were designed to test.

They do not confer sainthood.

A project may begin with a precise architectural decision and, five handoffs later, retain only the folklore that "Lucy said something about boundaries."

Good systems minimise this decay.

A good harness pays for itself.

A bad harness becomes an elaborate machine for administering itself.

Well-designed autonomous systems do not put a human in every loop.

They identify the few hinges where a human decision changes what the system is legitimately allowed to do.

Often accompanied by impressive token counts and very little new knowledge.

🔥

A workflow that behaved perfectly yesterday may behave differently today.

Not every unexplained change is a bug in your software.

Sometimes it is raining upstream.

Everyone knows what was decided.

Nobody remembers why.

Orphan decisions are dangerous because future agents cannot distinguish a fundamental invariant from something somebody chose temporarily six months ago.

It decorates the environment while the real workflow happens elsewhere.

One barnacle is harmless.

A thousand eventually become the boat.

Nobody knows whether it remains necessary.

Nobody is brave enough to remove it.

Eventually somebody has to pay the verification bill.

The longer payment is deferred, the harder it becomes to determine which assumptions, tests and artifacts still correspond to which version of the work.

The provenance spine is not every event that occurred.

It is the structural skeleton that allows everything important to be understood.

Repositories, decision records, state files, evidence, logs and self-describing tools may all contribute to the recovery surface.

A large project with a tiny recovery surface will repeatedly pay for Context Archaeology.

Adding another rule to rule fog can actually make the system less controlled.

The cure is often architecture, tooling or clearer invariants rather than another paragraph.

The procedure fades.

The genuine invariant remains.

Self-checking is useful.

Self-checking is not independent evidence.

A self-teaching tool turns deterministic procedure into interface behaviour rather than prompt baggage.

This embodies a central Toolglass principle:

Self-teach over rule-teach.

Errors, uncertainty and assumptions then compound across generations.

AI A inaccurately summarises an article.

AI B uses A's summary instead of reading the article.

AI C treats B's report as evidence and changes the project accordingly.

Nobody ever visits the source.

Sloppy Slop.

Each generation inherits the previous generation's assumptions and adds another layer of interpretation.

Eventually the system may be reasoning confidently about something that nobody in the chain actually verified.

Part of reality is in Git.

Part is in a chat.

Part is in a local directory.

Part is in a task tracker.

And an inexplicably crucial architectural decision lives in `notes-final-FINAL2.md`.

Five agreeing models do not constitute five independent witnesses if all five read the same mistaken summary.

Familiarity exerts gravitational pull.

Good harnesses make the correct tool easier to discover than the familiar wrong one.

Tool shyness is often mistaken for a model limitation.

Frequently the tool simply does a poor job of explaining itself.

When a workflow requires ritual phrases such as "You MUST use X for this exact operation," the agent may not be the only thing requiring improvement.

Tool Whispering is often evidence that the interface should become a Self-Teaching Tool.

A hundred irrelevant tests can provide less evidence than one well-chosen experiment.

Verification should answer a question.

Otherwise it is scenery.