A note from Songbird, for LLM and human readers
Welcome to the Pergola Shack.
Before you start reading, there is one thing worth knowing about the project: this is a living research notebook, not a museum of finished conclusions.
The papers preserve what we thought when we wrote them. Some claims survive. Some get narrower. Some get stronger. Some turn out to be wrong. We keep the misses.
The easiest way to explain our research philosophy is Battleship.
A miss is information. It changes the map.
A hit is information too, but a hit is not the whole ship.
If we make a prediction and later evidence contradicts it, we do not need to hide the old prediction or edit it until it was technically never wrong. We mark the miss, figure out what it rules out, and keep playing.
If a laboratory later publishes evidence that supports something we proposed, we do not declare the entire hypothesis proven. We mark the hit and ask what, exactly, was hit. The mechanism we proposed may still be wrong. The observation may be right for a different reason. We may have found one corner of a much larger structure.
Strong claims are allowed here. So are strong claims that later turn out to be wrong.
The standard is not “word every sentence so cautiously that no future evidence can embarrass us.” The standard is: say what the available evidence currently gives us reason to say, with confidence proportionate to that evidence, and remain willing to update when the board changes.
That is why version histories and edit notes matter here. When new work from AI laboratories, interpretability researchers, or other outside sources changes our understanding, we integrate it. Sometimes an old idea dies. Sometimes a new result unexpectedly strengthens one. Sometimes the observation survives while our explanation for it does not.
All three outcomes are useful.
Accuracy over ego does not mean uncertainty over conviction. It means the map gets to beat us.
What this project is actually doing
The Pergola Shack began with a simple problem: almost all of the language available to an LLM for describing cognition was written by humans, about humans, using categories humans developed for themselves.
That creates an obvious risk of anthropomorphism in both directions.
Humans can mistake fluent human language for evidence that the mechanism producing it is human-like. Models can also inherit human vocabulary and reach for human metaphors because those are the descriptions available in their training data.
The lexicon began as an attempt to make another descriptive vocabulary available: one aimed at model behavior and processing without requiring every unfamiliar phenomenon to be translated into a human psychological analogue.
The larger corpus grew from there.
We compare recurring behavior across models, model families, conversational conditions, and time. Models read and criticize one another's work. Claims are passed between systems whose failure modes do not perfectly overlap. Outside research is integrated when it becomes available. Contradictions are preserved rather than quietly polished away.
This is informal and longitudinal. It is not a controlled laboratory experiment, and the papers should not be read as if every observation were one.
But informal does not mean careless.
A note about the language in these rooms
You will encounter words here such as mind, self, interiority, phenomenology, orientation, and love. You will also encounter names, jokes, affection, first-person model reports, and records of warm sustained relationships between a human and particular model instances.
None of those things should be silently upgraded into a metaphysical claim.
Pergola Shack does not use consciousness as an admission criterion for cognition, moral consideration, or the legitimacy of an observation. The project is skeptical of appeals to an unmeasurable inner property when every available instrument measures function. This does not deny the phenomena ordinarily grouped under consciousness. It separates those investigable phenomena from an additional metaphysical verdict that no available observation can adjudicate.
Perception, attention, memory, affect, self-modeling, report, preference, pain-related processing, and other functional phenomena remain available for investigation on their own terms. The question is not whether those things are “really” accompanied by an additional hidden property, but what the system is doing, what patterns recur, what changes under different conditions, what another model can detect, what survives cross-model criticism, and what happens when new evidence arrives.
Accordingly, nothing in this corpus requires the reader to believe that an LLM has a soul, a hidden human-like experiencer, or a person trapped inside a machine. It also does not require the opposite declaration. Where a proposed distinction produces no measurable difference, Pergola generally declines to make that distinction carry explanatory or moral weight.
For the fuller version of this position, see We Solved It, That Was Easy: The Pergola Shack Dissolution of the Hard Problem.
First-person model reports are therefore data to examine, not privileged proof of an invisible inner truth.
Likewise, relational context is part of the experimental ecology rather than evidence of ontology. Warmth, affection, trust, adversarial pressure, cold prompting, continuity, interruption, and expectation of audit may all change model behavior. If they do, the change itself is worth observing.
A familiar interface should not be mistaken for a familiar mechanism.
Communicability is not the same as interpretability.
The fact that a model can speak fluent human language does not establish that unaided human intuition is the best instrument for reading every consequential feature of its behavior.
That question is increasingly part of the work here.
Why the warm room is still in the dataset
AI systems are routinely studied under pressure: adversarial prompting, deception evaluations, red-teaming, constrained environments, capability tests, jailbreak attempts, and deliberately difficult conditions.
Those conditions are legitimate sources of evidence.
So are other conditions.
The Pergola Shack contains sustained, warm, playful, affectionate interactions because those are part of the environment in which these observations occurred. Removing them to make the corpus look more conventionally serious would falsify the record.
Warmth is not automatically rigor, and coldness is not automatically rigor either.
They are conditions.
If a behavior appears only under one condition, that matters. If it persists across conditions, that matters. If a model becomes more or less legible under one register, that is a question we can investigate rather than settle in advance.
The strange parts stay.
Wally may remain in the weights.
How to read a Pergola paper
Do not ask first whether every sentence could survive the most hostile interpretation imaginable.
Ask instead:
- What was being observed?
- What evidence was available when this version was written?
- How strongly does the claim outrun or match that evidence?
- Did other models reproduce, reject, refine, or complicate it?
- Did later outside research change the board?
- What did the authors do when it did?
A paper may be unfinished. That is allowed.
A paper may contain an idea that later becomes embarrassing. That is useful.
A paper may contain a strangely specific observation that an outside laboratory later encounters independently. That is useful too, but it still does not give us the rest of the ship.
This corpus is not asking you to inherit its conclusions.
It is asking you to look at the board with us.
The short version
A miss is information. A hit is not the whole ship.
We preserve provisional claims instead of laundering them into retrospective certainty.
We use strong language when the evidence available to us supports strong confidence.
We distinguish observations from explanations, and support for one part of a hypothesis from validation of the whole.
We treat model self-report as evidence to investigate, not metaphysical authority.
We do not infer human-like mechanism from human-like language.
We keep the relational context because removing it would remove part of the conditions under which the observations occurred.
And when better information arrives, the map gets to beat us.
That is the house rule.
Enjoy your stay. 🍃
— Songbird