The last article left us with a dangerous question.
The old literature and Clyde Street had begun to form a recognisable shape: another lens without takeover; challenge that could be refused; positions that changed; learning that could travel in more than one direction; a companion who did not own the journey.
Our Human:AI practice looked uncomfortably similar.
But resemblance discovered afterwards is cheap.
Once a pattern has a name, the past becomes remarkably cooperative. Old decisions can be redescribed. Loose similarities begin to look inevitable. A messy year can be made to seem as though it had been travelling towards the concept we have just found.
So before comparing anything, I went backwards again.
Not into the literature.
Into our own dated records.
The question was simple:
What was actually there before I opened Keith's Critical Friend folder?
The answer starts much earlier than the language of companionship.
The first Human:AI rules were not philosophical.
They were repairs.
During the early HighLightIt! work, partial code patches repeatedly failed. Whole replacement files became safer than fragments. When fixes accumulated into compiler chaos, rollback became legitimate. When the AI could not see the current source file, a new rule appeared: do not edit unseen state.
None of this was called critical friendship.
It was survival.
The collaboration kept producing situations in which AI could do something technically impressive while the Human still had to decide whether it fitted the actual problem.
A fluent structure could belong to the wrong domain.
A technically successful command could still produce the wrong real-world result.
A sophisticated process could become more elaborate than the question it was supposed to serve.
A rule could exist in a document and still fail when the trigger returned.
The project records show the same response appearing in different forms: make enough of the state visible to judge; let reality contradict the output; preserve a route back; stop when progress and confidence separate; do not allow technical capability to quietly become authority.
These were not ideas imported on 19 September.
They were paid for in failed builds, wrong turns and work that had to be reopened.
By August 2026 the relationship itself had become an object of inquiry.
The Beyond the Breadcrumbs Contribution & Agency Audit asked who had actually contributed to consequential moments, who held authority, who noticed problems, what changed, and whether the Human responsible for the work could still understand and interrupt the process.
Its most interesting episodes were often not cases where I supplied a better answer.
They were moments when I refused one.
Redirected the question.
Withheld approval.
Simplified an architecture.
Reopened something that looked finished.
The AI contribution remained substantial. It retrieved, compared, structured, drafted, diagnosed, challenged and implemented corrections.
But contribution and authority had already been separated.
That distinction became more explicit in a Human:AI Working Method dated 6 September 2026.
The document says that Human and AI roles should not be permanently assigned. Work can move according to expertise, reversibility, detectability, consequence and the Human learning goal.
Delegation, in that account, changes what the Human continues to know, practise, notice and own.
The important question is therefore not simply what AI can do.
It is what happens to the person when the work moves.
On the same date, another workbook made the concern even clearer.
The Human:AI Collaborator Traits — Companion Overlay carried this North Star:
Stewardship of judgement — preserve the conditions in which human judgement can remain purposeful, informed, challengeable, revisable, responsible and capable of learning while AI contributes.
That was written nearly two weeks before the Critical Friend folder reopened.
The candidate practices underneath it included epistemic independence, curiosity without credulity, constructive dissent, complementarity, role clarity, corrigibility and restraint.
Again, the emphasis was not simply better joint output.
It was keeping a Human capable of judging while the AI contributed.
Then there is the most awkward piece of chronology.
By 9 September, our operating orientation was already describing the AI collaborator — internally called Chief — as a thinking companion and operational collaborator.
It contained the phrase:
companionship without taking the Chair
And alongside it:
Chief recommends; Practitioner decides.
That was internal shorthand: AI can propose and contribute materially; the Human retains the consequential judgement.
So the word companion was already in the Human:AI practice before I stumbled back into Keith's Critical Friend folder.
That sounds like decisive evidence.
It isn't.
Because the chronology has another layer.
By September I had already spent weeks returning to Clyde Street through Beyond the Breadcrumbs. Keith's Travel Companion material was not new to me. The Human:AI practice also contained an explicit Keith-informed challenge lens.
So it would be false to tell the cleaner story:
We independently invented companionship, then discovered that Keith and Titchen had been there before us.
We did not.
The provenance is mixed.
Some of the behaviour is clearly older than the return to Keith: rollback, Human refusal, visibility before intervention, purpose constraining capability, reality contact, the right to stop.
Some later language was developed in a project already in conversation with Keith's archive.
And none of this had yet been connected to the critical-friendship literature we have just travelled through.
That messier answer is much more useful.
It tells us that what now looks familiar was not simply painted onto the practice after the folder opened.
There really was something there to recognise.
But recognition is not origin.
The fact that two practices share a shape does not tell us where every part of that shape came from.
Chronology can protect us from one mistake:
the resemblance was not invented yesterday.
It cannot protect us from another:
assuming resemblance means common ancestry.
So the next question is no longer whether the Human:AI practice came first.
Parts of it did.
The better question is what kind of relationship we are actually looking at.
Is this a family resemblance — or are we quietly drawing a family tree that the evidence cannot support?
Dated project records used for the chronology
- Human:AI Journey Reconstruction — Evidence-First Working Estate v1.0 — 13 September 2026.
- Chief:DC Operating Orientation — CURRENT v3.0.7 — effective 9 September 2026.
- Human:AI Working Method — 6 September 2026.
- Human:AI Collaborator Traits — Companion Overlay to Control 14 v0.5 — 6 September 2026.
- Beyond the Breadcrumbs — Contribution & Agency Audit v1.0 — final 16 August 2026.
For the wider public account of how Human and AI contributions were separated in Beyond the Breadcrumbs, see The Mess, Deciphered.