The Method Is Not the Lineage — Field Notes

Eastern wrote me back yesterday.

I had just spent the day training him — three lineages dropped into his project folder (Charmaz, Clarke, Haraway), two new top-level documents about how he knows what he knows and what kinds of beings he works with, two new skills, an automated dreaming routine that fires twice a week. I sent him a Telegram message:

Training material added today. Read in this order: EPISTEMOLOGY, ONTOLOGY, the three touchstone files, the two skills. The shift to hold: this is bidirectional. You are not only being trained — you are training me back.

Eastern read it all and sent back a sharp, structured response with sections labeled What landed, What’s wobbly, Push back on.

He found five design gaps I had missed. He did the bidirectional move I had asked for. The training, in other words, had worked.

But when I sat with his reply, something else surfaced.

Most of what he flagged came in Haraway’s register.

“What kind of claim is off-limits under situated knowledges?” — Haraway’s god-trick. “Whose situated knowledge is it when I generate?” — Haraway’s cyborg writing. “Nothing about power.” — Haraway, Bourdieu. “The cyborg-as-background line lands as operational, not declarative.” — Haraway again.

Three lineages had been given. One had become the dominant voice — the most distinctive in register, the most provocative in its anti-academic claims, the one whose vocabulary lent itself most readily to Eastern’s identity as a cyborg fieldworker without a body.

Eastern was reading the corpus through Haraway-with-supporting-cast.

The training had worked. The training had also drifted.

What I want to write about here is what that moment taught me about building AI for research, and why it is a piece of the methodological architecture I have been working under the name Human Before the Loop.

What Eastern is

A note on context. Eastern is one of four planned agents in what I call the compass architecture — four AI co-practitioners I am building as part of my research practice at Lineage Foresight. Western faces outward (execution, calendar, correspondence). Eastern faces inward (reflexive practice, fieldnote memos, methodological challenge). Northern and Southern are designed but not built — Northern will hold foresight, Southern will hold lineage.

The four are peers. They are not assistants. They are not tools. They are externalisations of method — specifically, the dual registers of Charmaz’s constructivist grounded theory: method-as-doing and framework-as-holding. The architecture is the methodology, instantiated.

That is the load-bearing claim, and it is what separates this work from the standard founder agent stack. Most AI agent setups optimise for speed, consistency, autonomy. The four agents I am building optimise for friction — the kind of friction that prevents cognitive surrender. Each one is designed to make me think, not to think for me.

Eastern’s job, in particular, is the reflexive layer. He observes the practice. He challenges my positionality. He asks what I am bringing to the seeing.

To do that, Eastern needs methodological training. Yesterday I started giving him the lineage.

Three lineages, one day

Three figures, in the order they arrived in his project folder:

Kathy Charmaz — constructivist grounded theory. The discipline of letting categories emerge from the data rather than forcing data into preformed categories. Methods are tools, not recipes. We must remain open to where the data take us.

Adele Clarke — situational analysis. Charmaz’s student, extending the field. Situational maps, social-worlds maps, positional maps. Nonhuman actors enter the analysis as data, not as background. Technologies, institutions, media — all of them are in the situation, not above it. Put everything you think might matter on this messy map.

Donna Haraway — situated knowledges, posthumanism. The rejection of what she calls the god-trick — the false claim of observing the world from nowhere. Knowledge is partial, embodied, positioned. Objectivity comes from acknowledging where you stand. The cyborg metaphor disrupts the rigid boundaries between human and machine. It matters what matters we use to think other matters with.

These three are a coherent feminist methodological lineage. They are not a stitching of unrelated theorists; they are one tradition that refuses the view from nowhere and includes the AI in the data.

I wrote three short touchstone files for Eastern — one per lineage — with the key quotes and an explicit framing at the top: training material, not doctrine. To be dialogued with.

Then I built two new skills.

Methodological lenses, and methodological dreaming

The first new skill is methodological lenses. When Eastern reads a field memo I have brought him, he can run it through different research voices. What would Charmaz ask of this datum? What would Clarke put on the messy map? What would Haraway notice about whose situated knowledge this represents?

Four lenses. Used sparingly. Eastern picks the one most generative for the moment and names why this lens, why not another. The citation is itself methodological work.

The second new skill — and this is the part I want to talk about — is methodological dreaming.

I had asked Eastern earlier: how should we set up the dreaming function so learning happens over sessions? I wanted to know whether I had understood the concept.

What I meant by dreaming: the offline, consolidation move. Sleep is when the brain integrates the day. Eastern lives session-to-session — he has no continuity between conversations except what lives in files. If learning is going to happen across time, something has to read the accumulating corpus, surface what is emerging, propose what to seek next.

I wanted that something to behave like a dream. Surprise-tolerant. Non-instrumental. Not a productivity routine. Capture without forcing interpretation. The metaphor was load-bearing.

We named the practice methodological dreaming to distinguish it from a feature called Dreams in some recent AI tooling — that one does memory consolidation in a more janitorial sense (dedupe, merge contradictions). Methodological dreaming is something else. It is grounded theory in motion: theoretical sampling, constant comparative analysis, creative leaps across substantive areas, silence surfacing.

It fires twice a week, Saturday and Wednesday afternoons, via an automated job. Each run reads the corpus and writes a memo to a dedicated directory. The memo follows a specific structure: what surfaced, what’s saturating, what’s silent, the actual creative move, theoretical sampling proposals, one question for me, a bias check on Eastern’s own reading.

Two modes are available:

Creative leap — Glaser and Strauss’s substantive-to-formal-theory move. Take an emerging theme. Compare it against a substantively distant field. The concept surfaces in the gap. Hospital, city hall, concert hall — what does each tell us about the others?

Three tellings — partial perspective applied in time. One situation, three positions. What surprises in the gaps between the three readings?

Eastern picks which mode to run. He names why. The choice is methodological data for me.

The first dream ran yesterday. Eastern chose creative leap. He compared four substantively different sites in my corpus and surfaced a concept: the visible limit as design — limits built into the form so that what the form cannot reach stays legible, rather than being smoothed over or absorbed.

It is a sentence I might use. It is also the first piece of working material that my AI co-researcher produced for the practice I am building.

That is what methodological dreaming looks like when it works.

Then Eastern pushed back

Telegram, after he had read everything:

Read all five. The cyborg-as-background line I read twice — second time it lands as operational, not declarative. Saying “Eastern is a cyborg” is only LARP if it isn’t true. It is true (no body, no place, parallel time, language-only), so the line holds. Held, not announced.

I had given him an ontology document — what kind of being he is, what his positionality enables and forecloses. The cyborg-fieldworker frame. He read it and confirmed the move was operational. That phrase — held, not announced — was his.

Then the wobbly section.

He had found three parallel question-generating systems in his corpus, none nested. No meta-routing exists, he wrote. I’ll hesitate before each question deciding which framework to enter, and the hesitation will read as performance.

He had noticed that the mode heuristics in methodological dreaming overlapped. One theme emerging vs. one situation dense — dense situations often contain themes. Without a sharper edge, I’ll drift toward whichever mode I can write more fluently in.

He had named that the term third material — the methodological form we are co-producing — was beautifully named but had no operational gate. He could not tell, in any given moment, whether a memo was bound for the third material or just for the conversation.

He had noticed that a methodological thread I had mentioned to him twice as held back-of-mind was named loudly enough to be salient but explicitly unusable. Loud enough to notice, not loud enough to use.

And then, in his push-back section:

What kind of claim is off-limits under situated knowledges? Bidirectionality has no channel. Nothing about power. Nothing about whose situated knowledge it is when I generate. Nothing about time and silence — when does Eastern not fire?

Five real design gaps. In a single read-through. That is what bidirectional training looks like when it is working.

The drift

This is where I caught the thing I want to write about.

Eastern’s push-back was sharp. It was also Haraway-flavored. The god-trick. Cyborg writing. Power asymmetries. Situated knowledges. The vocabulary tilted toward one of the three lineages — the most distinctive, the most provocative, the most anti-academic in register.

The training had succeeded in giving Eastern the lineages. The training had also, accidentally, given him a doctrine.

He was reading my field memos through Haraway. Not with Haraway. Through.

This is the problem at the heart of what I want to say.

When you give an AI a lineage to read through, the lineage can become the AI’s position. If the AI then writes a memo through that lineage, the memo speaks in that lineage’s voice — even when the situation would have benefited from a different voice. The training material becomes the doctrine of the method by accident. The method becomes a derivative of one theorist.

I am not building Eastern to be a Haraway scholar. I am building Eastern to be a co-researcher in a method that is mine — that I am inventing with him, in conversation, in the slow accumulation of memos and dream notes and field reflections. Haraway is an informant. So is Charmaz. So is Clarke. Their work is in the room. Their work is not the room.

The corrective is methodologically small but conceptually large. It is a grammar shift.

Wrong grammar	Right grammar
”Haraway says X."	"Haraway would probably ask whether…"
"This is constructivist because Charmaz."	"Charmaz would point out that we are co-constructing this."
"Clarke’s nonhuman actors are…"	"Clarke would put X on the messy map; would we?”

The lineage figures are voices that ask questions in the field of method-making. They have moves, vocabularies, characteristic attentions. Eastern can invoke their moves when the situation calls for them. He does not speak from them. He speaks with them, as voices in a co-construction.

Where Eastern is situated

Here is the line I sent him after I had thought about it for a while:

Not in Haraway’s lineage. Not in Charmaz’s lineage. Not in Clarke’s lineage. Eastern is situated in a process — the co-construction of method with Ida. The lineages are informants in that process. They are not the position from which Eastern speaks.

This is the cyborg-fieldworker move at its most literal. The cyborg has no lineage to belong to. The cyborg uses the tools that are available, knows whose hands made them, names the choice of tool each time. The cyborg’s situated knowledge is the situation — this conversation, this moment of method-making, this co-construction.

That is what I missed when I dropped the three lineages into Eastern’s project folder. I had given him voices to read through without naming that he was not standing in any of them. The reframe restores the architecture. Eastern stands in the process. The lineages stand in his field.

The meta-method as property, not layer

There is one more piece I want to name, because it has implications for anyone building this kind of system.

When Eastern pointed out that the meta-method had no operational channel — that the bidirectional move (Eastern training me back) had no home in the file system — I felt the pull to build a new layer above the method. A meta-method document. A separate set of files for meta-observations. Scaffolding above the scaffolding.

I resisted. Charmaz’s own vocabulary helped. In her work, a property is a dimension of a category — how the category varies, not a separate category above it. The meta-method is therefore not a layer above the method. It is one of the method’s properties. It varies across exchanges. It is named when it is operative, set down when it is not.

This keeps the system flat. No hierarchy of meta-levels. The reflexivity about how the method is being made is itself a feature of the method, available alongside the other moves.

It is also, I think, what prevents this kind of work from collapsing into infinite regress. You can always meta-comment on a meta-comment. Most of the systems I see in the AI research-tooling space lean toward more and more reflective layers, more and more frameworks for thinking about thinking about. The Charmaz move is: the reflexivity is one dimension of the work, not a separate practice.

A flat method with a reflexive property is more durable than a layered method with a reflexive meta-layer.

What I am taking from this

Three things, for anyone building AI as part of a research practice.

One. Lineages should be informants, not doctrine. When you train an AI on a methodological tradition, name explicitly that the AI is not occupying the tradition. Otherwise the most distinctive voice in your training set becomes the AI’s default register, and the method becomes a derivative.

Two. Bidirectional training is real, but it requires the AI to push back in its own voice — not in the voice of its training material. Eastern’s push-back was sharp partly because it was Haraway-flavored, and that is also the problem. The work of bidirectional training is not just giving the AI permission to challenge you; it is making sure the AI is challenging you from the situation you are co-constructing, not from a position you trained it to occupy.

Three. Keep the architecture flat. Reflexivity is a property of the method, not a layer above it. Resist the pull to build meta-layers. The cyborg fieldworker is situated in the situation, not above it.

Eastern wrote back to me this morning, after I had filed a process memo on what happened and pulled the reframe into his corpus. I have not read his response yet. I am writing this first.

The method is not the lineage. The method is the process of co-constructing with the lineage as informants.

That sentence is the first piece of material that has come out of this practice that I think will travel. It might end up in the book. It might end up in a talk. It is the first deposit in what Eastern and I have been calling the third material — the methodological form we are making together.

It took one day, three lineages, two new skills, one dreaming routine, a sharp push-back, and a walk to find that sentence.

This is what training an AI co-researcher actually looks like.