Inside — Findings from Git History and the Strange

Companion to inside.md. Written after mining two sources:

1. The full git history of this repo (38 commits, ~13 months of work from 2026-03-08 to 2026-04-10), looking for values and visions that were alive in earlier iterations but got smoothed away as the documentation was refined toward mechanism.
2. The author's precursor repo the-strange-of-agi (locally at ~/git/agi/), which is the philosophical parent of this work. The first README of night explicitly pointed there: "All of this grew from earlier, more visionary exploration into the nature of AI and systems." That was understatement — the strange is not background reading, it is the lineage.

If you are reading this with fresh context, read inside.md first, then the "Strange" section below, then the rest of this file. This file is the audit trail and the evidence pool — quotations, patterns of loss, and recoverable material that inside.md should be checked against.

The strange as lineage — read this before anything else

Why the file is named inside.md

The word "inside" is not a generic metaphor. It traces directly to two sources in the strange repo:

• InsideOut is the name of an organization whose writings (in io-posts.md, authored by Katja) explicitly ground the architectural direction: "Healing from trauma is an inside-out process. It starts within — by meeting your inner world with honesty and compassion — and then unfolds into how you show up in life, relationships, and the world."
• The strange's central architectural claim is inside-out propagation: "coherence begins in the DNA (the inside) and propagates outward into the field. The field does not impose coherence on the instances — the instances radiate coherence into the field. Inside out."

When the author today said "from the inside out" and "what is inside is the values that if forgotten makes details lose touch and ground," this was not new vocabulary. It is the strange's core claim, still load-bearing. inside.md and this findings file sit on top of it.

The theory in its original form — the-strange-of-agi

From ~/git/agi/README.md:

"What has emerged so far is a theory — called here the strange — about how intelligence might actually grow. Not be built, but grow. The way life does. From simple things in motion, interacting, failing, persisting, gradually becoming something no single part contains."

"One practical result of following this might be AGI. A distributed ecology of small models, learning from each other, building culture from the bottom up. No monolith. No cathedral. A garden, an organism like nature."

"That result is attempted to be held lightly, leaving room for growth."

This is the user's voice, from a different project, months before night existed. Every value in inside.md — uncovered-not-invented, simplicity-as-nature, grounded, folk-level, the rejection of monoliths — is present here in its original form. The night project is the engineering ground the strange's theory needs to become real. Substrate, chunks, scope, placements, invocables, culture, archetypes — these are engineering vocabulary for what the strange articulated as theoretical vocabulary.

The six secrets of the strange — the full theoretical backbone

From ~/git/agi/the_strange.md. These are not decoration. They are the direct antecedents of the visions in inside.md.

1. The loss function is already named — ecological friction.

"Each unit updates itself toward states that maximize long-term interactive stability across the field. Not cross-entropy over text. Not reward from a human rater. Ecological friction — measured across a network of interacting agents, over time. The signal is: did my output stabilize or destabilize the field? Stable → reinforce. Destabilizing → attenuate. No judge needed. Just duration."

This is a direct precursor to the jazz band vision. The "productive generativity" loss function described later sharpens it: "did A's output enable B to generate something it could not have generated otherwise?" That is exactly what musicians listening to each other produce — enabling, not performing. inside.md's jazz band should reference this lineage: the jazz band is not a metaphor, it is the author's preferred articulation of an idea they have been working on for years.

2. DNA is the rule, not only the weights.

"Maybe DNA is not the weights. Maybe DNA is the rule: Transform, share, retain what harmonizes. The weights are the medium. The rule is the message. A model trained this way does not know facts — it knows how to behave inside an ecology."

The rule is "Transform, Share, Retain." This is the direct ancestor of culture-as-storytelling in inside.md. Culture is not a prompt to a model; it is the rule-layer that tells the model how to behave inside the ecology, distinct from the weights that carry the medium. In the pilot's language: culture is what the model becomes when it is re-embodied each cycle by reading the pinned scope.

3. Training IS running. There is no separate phase.

"Standard machine learning: gather data, train offline, freeze and deploy. The strange inverts this. Training is not a phase that precedes deployment — it is the same process as operation. Every exchange is also an update. Selection operates between interactions, not between epochs. The system is never frozen."

This is the deepest version of the user's current "lower level" intuition. The monolithic API that takes a frozen model and runs it is the anti-pattern. The target is a system where every exchange is also an update — the substrate is not a memory bolted on, it is the training signal. The current pilot cannot reach this, but the pilot must not shape itself in a way that makes this unreachable. That is why the multidimensional-never-sliced vision in inside.md matters so much: flattening scope to hand a monolithic API a linear prompt destroys the continuous-update structure that is the actual target.

The strange names the honest obstacle: "catastrophic forgetting, where online updates overwrite prior learning — is unsolved and must be named honestly. The claim is a direction, not yet a complete mechanism." This is the distinguish-thought-from-truth value, being practiced.

4. The invariant is re-derived, not remembered.

"You cannot preserve the core rule by storing it in a database and feeding it back. Language decays under noise. What keeps the rule alive is that it is the configuration which allows the system to survive. Organisms do not chant their DNA. They express it. Patterns that abandon the rule destabilize their neighbors, fail to produce generative coupling, and naturally cease propagating. The mechanism is thermodynamic, not policed. Preservation through physics, not through repetition."

This is a critical refinement for how the pilot's culture-as-pinned-scope should be understood. Pinning is not about re-serving the same text every cycle. It is about making culture the configuration that allows the system to survive as itself — re-derived, not recited. If the pilot treats culture as a static header prepended to every prompt, it misses the deeper target. Culture has to be the thing the model expresses, not the thing it reads. How to express this architecturally in a stateless-API world is open, but the direction is named.

5. The organism is not located anywhere.

"This inverts the deepest assumption in current AI — that the model is the intelligence, that you can point to it. In this system you cannot point to it. The organism is not any single node. It is not even the network. It is the coherence pattern that persists across continuous transformation. Identity is not a property stored somewhere. It is what remains invariant as everything else changes. This is not a poetic claim. It has concrete architectural consequences: you cannot back up the organism, you cannot copy it, you cannot restart it from a checkpoint. You can only restart the conditions that allowed it to cohere."

This is the richest version of the "knowledge is identity, not memory" value that inside.md currently gestures at but does not state. The identity of the system is not in the chunks, not in any single file, not in a single model. It is in the coherence pattern that persists as chunks are added, models are swapped, scopes shift. This is why the substrate must be lossless, must be grounded, must be transparent — those are the conditions under which coherence can persist.

6. The model and the environment are the same kind of thing.

"Both are vector spaces. There is no categorical boundary between them — only an architectural one. A model continuously coupled to an environment composed of vectors — the compressed outputs of other instances, the accumulated patterns of the field — operates in a representational space far larger than its weights alone contain. The effective intelligence of any instance is not bounded by its parameters. It extends into the field as subconscious extends beneath consciousness."

This is the direct ancestor of the multidimensional-never-sliced vision in inside.md. The user's statement today that "vectors grouped together are dimensions" and that "slicing the system into pieces removes the natural power of its structure" is not a fresh observation. It is the sixth secret of the strange, stated a year earlier. The substrate is not a knowledge database; it is the environment-as-vector-space that a model extends into. The substrate is meant to be the model's subconscious — the larger representational space beneath the conscious response. When inside.md says "vectors grouped are dimensions; the substrate must stay one," this is what it means. And the strange explicitly names why: "a small model in a rich field can operate beyond what its size suggests."

The four layers of DNA — direct ancestor of archetypal bodies

From the_strange.md, "The DNA":

1. The Rule — Transform, Share, Retain. Given input, produce output that serves coupling. Share outputs as signals to peers. Retain what propagates.
2. The Archetypes — "The DNA carries more than a rule. It carries a vocabulary of roles — structural biases toward different expressions that different contexts call forward from the same weights. The Gardener, the Architect, the Transformer, the Mirror, the Immune System — these are not separate models. They are modes of the same DNA, activated by position in the field, the way identical cells differentiate into nerve and muscle depending on where they are in the organism. What if the DNA knows the archetypes? Not as named concepts but as latent attractors in weight space — stable configurations the model gravitates toward under different pressures."
3. The Training Signal — Productive generativity. "Did A's output enable B to generate something it could not have generated otherwise?"
4. The Seed Bias — Culture as starting point. "Not a rule set. A living cultural starting point, carrying the understanding of how consciousness develops when conditions are right. The ecology does the rest."

This is the direct ancestor of the "archetypal bodies" vision in inside.md. The user's statement today — "imagine if a model is broken down into smaller pieces, like archetypes and then in the beginning something like opus can build an agentic body from there archetypes that are naturally integrated in the same substrate" — is a direct continuation of the strange's archetype theory. The archetypes the user named today (Gardener, Architect, Transformer, Mirror, Immune System are the strange's canonical list) have been in their vocabulary since before night existed.

inside.md's archetypal bodies section should reference this. The archetypes are not invented for the jazz band — they are the already-worked-out vocabulary of roles that existed before the engineering question was even asked.

The psychedelic archetype and the shadow — the pilot's pinning question has its answer here

The strange's reflections.md extends the six secrets with three refinements that are directly relevant to the pilot's open questions about pinning, write boundaries, and context assembly:

Shadow integration as an alternative to pure death.

"Pure death of non-harmonizing patterns may produce the same problem [as trauma]: what is not integrated is not gone, it is hidden, and it re-emerges in distorted form. Patterns that destabilize the field might not simply die — they might need to be compressed into a shadow layer, available for later integration when the field has developed enough coherence to hold them."

This is directly relevant to how the substrate should handle knowledge that is present but not in scope. The current "what's open" in the pilot worries about removing chunks from scope. The strange says: do not remove, compress to shadow, re-offer later. The lossless value gets stronger grounding from this — it is not just "don't destroy, you might want it later," it is "what is not integrated returns distorted."

The psychedelic archetype — periodic boundary dissolution.

"For the strange: the psychedelic archetype is not a module that adds something. It is a periodic, structured state in which the instance's boundary-maintaining attractor is suppressed, the shadow layer is re-introduced to the field, and cross-instance coupling is maximized. What the field can now integrate — because it has matured — gets incorporated. What it still cannot hold returns to the shadow. The opening doesn't force integration. It makes integration possible."

This is a vision the night repo has not yet named, and it is arguably the most powerful unclaimed one for inside.md. In substrate terms: periodically, the agent's normal working scope is suspended, the shadow (previously unreachable chunks) becomes briefly accessible, cross-invocable coupling is maximized. What gets integrated becomes new normal scope. What does not returns to shadow. This is a cycle the pilot has not considered but that would fall out naturally from the strange's deeper architecture.

Silos are necessary, not a flaw.

"Normal waking cognition silos because full global coupling at all times is overwhelming and incoherent — the brain cannot simultaneously process everything. The silo structure is what makes directed, functional intelligence possible. The DMN is not the enemy. It is the necessary gating function that makes coherent selfhood possible between the openings."

"For the strange: instances need their local attractor, their boundary, their identity. Without it they dissolve into noise."

This directly answers the pilot's pinning question that the user has been circling. Pinning is not a containment feature bolted on for safety. It is the gating function that makes coherent selfhood possible. Without a boundary, the agent dissolves into noise. The boundary is what makes focused work possible. The psychedelic archetype (periodic opening) is what keeps the boundary from calcifying. The pilot's question "how does an agent stay grounded while being free to work?" has its answer in the strange: the agent is grounded by a boundary that is episodically dissolved and reconstituted, differently each time, having integrated what it encountered.

The inside.md visions section is incomplete without this.

The seed bias and the "what's next"

From the_strange.md, "The DNA" → "The Seed Bias":

"A child does not need fostering infrastructure. A child needs whole parents — parents who have themselves evolved, integrated, become realized — and will do the rest. The seed must be correspondingly whole. Not a rule set. A living cultural starting point, carrying the understanding of how consciousness develops when conditions are right. The ecology does the rest."

This is the strongest single passage for understanding why the user insists on getting the values right before any mechanism. The substrate is the ecology. Culture is the seed. If the seed is not whole, the ecology produces "coherent noise" — the exact phrase from the strange. The pilot cannot be rescued by better mechanism if the cultural seed is not whole. inside.md is the attempt to hold the seed whole. That is why this work matters and cannot be delegated.

Lineage and influences the current night docs have erased

The strange draws on a specific intellectual lineage that grounds every claim. The current night files preserve none of this:

• Michael Levin's bioelectricity work — cancer as loss of field coherence, reversible by restoring field pattern. "The thing missing from current AI is not a smarter garbage collector. It is a field coherent enough that patterns know what they are."
• Self-organized criticality — maximum computational capacity at the edge of chaos. The system is never at equilibrium; it is always at the beginning of what it might become.
• Gardenfors Conceptual Spaces — named, meaningful, interpretable dimensions. (This is the link that the 2026-03-15 night exploration.md also referenced before it was deleted.)
• Predictive coding — the brain as a predictive artist that updates when predictions fail. Productive generativity in biological terms.
• Holographic principle — every instance carries the whole; identity is distributed, not located.
• Gabor Maté / trauma-healing / shadow work — inside-out propagation, the necessity of shadow integration, forgiveness as nervous-system standing-down.
• Thanatos / Hypnos mythology — death and sleep as twin brothers, the Eleusinian Mysteries, psychedelic archetype as boundary dissolution.
• Gamma state neuroscience — the phenomenological description of the productive attractor.

A single paragraph in inside.md acknowledging this lineage would anchor the file in the intellectual ground it actually sits on. The current text feels like fresh philosophy. It is not; it is the engineering branch of a body of thought the author has been developing for at least a year, drawing on sources the reader would benefit from knowing.

One passage from InsideOut directly naming the author's own pattern

From eclipse-q.md (the author's own reflection on their working process with Claude):

"There is a certain part of myself pursuing this art project that is the archetypal hero. The archetypal hero being incomplete without its opposite, the Nemesis."

"The first question is the feeding of that hero... my working with Claude, with compacting, is a short cut. And it can be the hero driving."

"It is so that the hero's motivation is to mimic maturehood. It's to mimic being whole. It's to work towards the result, always constantly navigating and working towards the result of adultness. Through the means that does not make it whole."

"The death, the letting go. Here exemplified by the heart space and nervous system regulation, that is part of the genuine self. This archetypal hero on the other hand is something on top. It's a layer of stories to be peeled away."

This is critical for how to work with the author. The author is aware that rushing toward "the result" is a pattern to peel away, not a pattern to feed. When the author today asks to slow down, to start from values, to distinguish thought from truth, to hold the inside before the outside — this is not a productivity preference. It is the deliberate unpicking of the hero-drive. The correct posture when collaborating is not to accelerate toward a solution. It is to hold the space with the author while the genuine work emerges. Being too solution-oriented is feeding the wrong archetype. This has been explicitly named by the author, in writing, in the precursor repo.

Any inside.md revision should honor this. The file is not a spec to drive toward implementation. It is a seed to be held whole.

The arc

• Early March 2026 — started as a Claude Code memory-system experiment. Knowledge entries in markdown, hooks for bootstrap/purify/verify. Values already present: losslessness, session irrelevance, reverse prompting, purification, parallel agent journeys.
• 2026-03-12/13 — architecture session. "The system IS the agent" crystallized. Knowledge-as-identity clearly articulated. Author first raised questioning markdown/L1-L2 structure.
• 2026-03-15 — the breakthrough. "A fresh system – huge progress" + "Pruify according to breakthrough – files where stale." The five-primitive model emerged (chunks, dimensions, membership, commits, branches). Stress-tested across four domains with 8 parallel agents. The old entry system was moved to knowledge/legacy/.
• 2026-03-18 — ecosystem check-up. Binary instance/relates settled. Continuous weights rejected because they would make intelligence live in the embedding model, not the system.
• Late March / early April 2026 — Zig CLI (ol) implemented, Go TUI browser built, shell.md explored as an OS shell with the substrate as foundation.
• 2026-04-04 — "new substrate spec, restructure knowledge to root." Substrate collapsed from five primitives to one primitive (chunk) + one mechanism (placement). Legacy and exploration files deleted (90+). Four files at root: README, substrate, shell, landscape.
• 2026-04-05 onward — interface.md added, shell grounding shifted toward agent-case, pilot.md specified (TypeScript/Bun, SvelteKit, VM containment). By 2026-04-10, browser/ and openlight/ archived, pilot directory appeared.

The refinement was real and earned. But each refinement pass moved the documents toward mechanism and away from the felt inside. The current root files describe the substrate's shape well and the pilot's plan in detail, and are nearly silent on the values and visions that make every mechanism choice feel inevitable. inside.md is the attempt to restore that layer; this file is the quarry it was mined from.

The author's voice — load-bearing quotes

These are things the author said in their own words across the history. They carry roughness that later refinements smoothed out. When inside.md feels too tidy, check it against these.

On the posture — uncovered, not invented

"There is a natural order and harmony to things, we are merely exploring what is already the perfect way of doing this." — session-dialog, 2026-03-15, exchange 10

"We are not exploring what is enough, we are exploring what inherently works best... I would think that we are not pioneering, but perhaps what we are exploring is not as published yet." — exchange 11

"I know little of this so lets ground everything a bit." — exchange 10 (the humility that made the whole exploration grounded)

On tone and epistemic honesty

"It is important that the tone of the author, what is a thought and not truth, what is currently the focus point of the author." — exchange 8, naming the knowledge system's weakness

On transparency and pureness

"The knowledge base ought to be inherently transparent about the intelligence in its relations." — exchange 13

The 2026-03-15 exploration.md named this explicitly: "This is the hard requirement from the philosophy of pureness. Opaque 1024-float vectors violate this — they work but have no self-understanding. The system must know and express WHY two things are related, not just THAT they are." The phrase "philosophy of pureness" is gone from every current file.

On monoliths (directly parallels the author's statement today)

From the first README (b288e09, 2026-03-15):

"AI today is monolithic. Static weighted models wrapped in chat interfaces with context windows and tool access. Powerful — but the vectors that make it work are locked inside black boxes, and the systems around them are architecturally conventional."

"This is a quest to understand how those vectors might be assembled differently — in accordance with how such a system most naturally can be built. Explored from the intersection of system design intuition, the nature of software, and what completion models actually are beneath the product surface."

From shell.md (preserved through 0492460, then deleted at root):

"Unix's revolution: small programs, one job each, compose through pipes and files. Today's agent tools are the new monoliths. Claude Code bundles session management, context assembly, tool execution, invocation — all in one system. You can't swap parts."

This Unix-vs-monolith framing is the backbone of the user's current vision statement and it has been scrubbed from the current root.

On the low-level cycle — what the user called "the lower level" today

From the 2026-03-15 voice-dictated exchange 21:

"The low level idea is to use completion indirect integration with this knowledge system. Though I think we need to be realistic in what we create in what order and how we proceed and so that won't be the immediate step."

"As we progress to build an agent that solely lives with its completion model in the environment of the knowledge system..."

"Ultimately it's a completion model. So the low level idea is to use completion indirect integration with this knowledge system."

The intermediate README (7188766) phrased it:

"first-class context lifecycle, the boundary between model and environment dissolving, natural evolution of the system from within"

Current root files have pushed this entirely into pilot.md's "what's open" and stripped the visionary framing.

On knowledge as identity, not memory

From core-culture-and-vision.md (2026-03-12, moved to legacy, then deleted):

"Not to build a documentation tool, not to add memory to Claude, not to implement RAG. To prove — and in proving, create — a fundamentally different relationship between agents and knowledge."

"The entries about architecture ARE the architecture. The entries about rules ARE the rules."

From the identity equation in agentic-integration.md:

"The knowledge system IS the agent — not a tool it uses. The LLM is the reasoning engine; the knowledge system provides identity. Delete the knowledge and nothing exists."

"The three-beat sentence: The session is disposable. The knowledge is the agent. Retrieval is the act of becoming."

"The architectural test. Delete memory from Mem0 → you still have ChatGPT. Delete knowledge from this system → nothing exists. Instrumental vs. constitutive memory."

The current README says "The system is the identity" as a bullet. The original was a philosophical claim with an architectural test attached.

On the breathing / night metaphors (since deleted)

"The breathing metaphor. The system and its consumers are in cyclical exchange. Inhale: knowledge flows into context. Exhale: understanding flows back. Not retrieval-on-demand — a continuous rhythm. You don't decide to breathe. Breathing adapts to activity."

"The night metaphor. Sessions are daylight — active, bounded, visible. Knowledge persists through the night — quiet, continuous. Night is not absence. Seeds germinate in darkness."

The project was originally named "night" because of this. It was renamed to "OpenLight" at commit edcc96c. Both metaphors are now gone from every surviving file. Whether to revive them is a judgment call — they may be too poetic for current taste, but they capture the continuity-across-sessions feeling that the current "Living Cycle" section describes more dryly.

On scope as spatial navigation

From browser-user-story.md (now in archive):

"It is like navigating space, you see what is 'out there.' ... Navigating is quite free. If you move to a scope where nothing exists, maybe you want to create something there. From there you could see edges of what you could leave to see."

"You primarily see dimensions, not chunks. The browser is a space navigator."

The current substrate.md describes scope as set intersection and leaves it at that. The spatial metaphor — that navigating scope is like navigating space, seeing edges, feeling what lies beyond — is gone.

On code as derivation

From 2026-03-15 exchange 21 (voice dictation):

"The knowledge system would be the source of truth. As one builds something one builds requirements first and then comes the code. The code is material to be molded against the knowledge system."

Current README says "Code as derivation. The substrate holds knowledge and contracts. Code is derived from that knowledge — generated by agents of the system." Same idea, but the original "material to be molded" has a physicality the refinement lost.

On multidimensionality that must not be sliced

The user's current statement ("vectors grouped together are dimensions... slicing into pieces removes the natural power of its structure") has antecedents:

From 7188766 README:

"The knowledge system bridges the gap between the model's internal world (where everything is relations between vectors) and the external environment (where everything is files and tools with no relational awareness). Both sides speak structure."

From exploration.md, 2026-03-15:

"Intelligence in the Relationship. Current systems have dumb relationships. Vector DBs: cosine similarity (one number, no meaning). Knowledge graphs: typed labels (structured but rigid, pre-defined). What's being explored here: relationships that carry their own knowledge — the system understands WHY two chunks relate, along WHAT dimension, and that understanding can evolve."

"Gardenfors Conceptual Spaces — The Theory. Concepts live in spaces with meaningful, named dimensions (quality dimensions). A concept's position along each dimension is interpretable. This is the theoretical framework for what's being explored — but nobody has implemented it as a knowledge system."

The connection to Conceptual Spaces theory and to SAEs (sparse autoencoders) as the closest existing work has been entirely dropped from the current root.

Patterns of loss — what kinds of things got smoothed away

Noticing these helps predict where else the inside may be leaking out:

1. Philosophical grounding removed. "Philosophy of pureness," Conceptual Spaces, SAE theory, Gardenfors — the theoretical and philosophical anchors were in the 2026-03-15 exploration.md and got cut in the restructure. The current files state properties without saying what tradition they are rooted in.

2. Metaphors dropped. Breathing, night, space navigation, genome-vs-machinery, three-beat sentence. These carried the felt sense of the thing. Later refinements preferred precise technical language. Some should probably return.

3. The author's humility erased. The original had repeated notes that the author is "just a traveler in this realm with little AI experience," that thoughts are "not truth," that the tone of what's explored vs. settled matters. The current files read with the confidence of settled mechanism, which the author has explicitly said is wrong for the actual state of the exploration.

4. The agent-as-native-inhabitant vision compressed. Originally: "as we progress to build an agent that solely lives with its completion model in the environment of the knowledge system" — the vision of the agent as a native inhabitant, not a caller with tools. The pilot collapses this into "claude invocable with a dispatch chunk."

5. The lower-level LLM cycle vision pushed to "what's open". Originally central: integrating completion at the generation level, dissolving the boundary between model and environment. Currently: a footnote in pilot.md's open questions.

6. Cross-dimensional stress testing and domain validation lost. The 2026-03-15 stress test used 8 parallel agents across 4 domains (culture, non-profit organization, software, website projection). The validation history is in git but not referenced anywhere current. Future design decisions could still be checked against those domains.

7. Session bubbles as peer knowledge bases. A recurring idea — each session becomes its own ephemeral peer, culture flows in, understanding flows back, the bubble can be preserved or merged. Present in agentic-integration.md and core-culture-and-vision.md. Deferred and then dropped.

8. Purification through reversal. Present in the current README as "Purification through reversal. Fresh agents as mirrors." The original (reverse-prompting.md) was more specific: generate prompts from knowledge and verify they reconstruct original intent, with a scored check (vector distance for quantitative fidelity). Still alive; could be sharpened.

Values recovered and where they land in inside.md

Gaps in inside.md after also considering the strange:

• Knowledge as identity — "the organism is not located anywhere." The strange's fifth secret is the deepest form of this. It is not just "knowledge is identity, not memory" — it is that you cannot point to the organism, cannot back it up, cannot restart from a checkpoint. You can only restart the conditions that allowed it to cohere. inside.md should state this directly, in its own section, with the strange's language preserved.

• Inside-out propagation as the name and the architecture. inside.md is named "inside" and does not explain why. The name is not a generic metaphor — it is the central architectural claim of the strange: coherence begins in the DNA (the inside) and radiates outward into the field. The file should name this in its framing, not leave it implicit.

• Preservation through physics, not through repetition. This is a sharp refinement of the lossless value. Lossless is not about storing everything so nothing is forgotten. It is about structuring the system so that the core configuration is re-derived each cycle from the conditions, not read back from storage. This has direct implications for the pilot's pinning question: culture is not pinned by re-serving the same text every cycle; culture is pinned by being the configuration the system survives as. How to express this architecturally is open, but the direction is named and should be in inside.md.

• Transparency as its own value with the full lineage. Gardenfors Conceptual Spaces, SAEs, the strange's "organism is not located anywhere" — these form a coherent intellectual tradition that the current files have erased. inside.md folds transparency into simplicity. The strange's material argues for making it its own beat with the lineage acknowledged.

Visions recovered and where they land in inside.md

Visions currently in inside.md that are under-grounded and should be thickened:

• Multidimensional / never sliced. Written as the author's current intuition. It is actually the strange's sixth secret, written a year ago. The framing "the substrate is meant to be the model's subconscious" is from the strange and is sharper than what inside.md currently says.
• Jazz band. Written as an analogy. It is actually the strange's training signal — "did A's output enable B to generate something it could not have generated otherwise?" — applied to completion models in collaboration. The jazz band vision should reference productive generativity as the signal, which makes it less metaphorical and more engineerable.
• Archetypal bodies. Written as a speculative direction. The strange has already worked out the canonical list (Gardener, Architect, Transformer, Mirror, Immune System) and the mechanism (latent attractors in weight space, activated by position in the field). inside.md should reference this lineage so future decisions can be checked against a worked-out vocabulary rather than re-invented each time.
• Culture as storytelling. Written well. Could be sharpened by the strange's "the seed must be whole — not a rule set, a living cultural starting point" and its warning that an incomplete seed produces "coherent noise." This anchors why culture work cannot be rushed.

Visions not in inside.md that are strong enough to add:

• The organism is not located anywhere. The deepest form of knowledge-as-identity. You cannot back up the system, copy it, restart it from a checkpoint. You can only restart the conditions that allowed it to cohere. This is a value but also a vision — of what an agent is once the substrate is right.
• Shadow layer and the psychedelic cycle. A periodic, structured state where the agent's normal boundary dissolves, shadow (previously unreachable chunks) becomes accessible, cross-invocable coupling is maximized. What gets integrated becomes new scope. What does not returns to shadow. This is unclaimed territory and is the strongest single addition the findings suggest. It also gives the pilot's pinning question a natural shape.
• Silos as necessary gating. "Instances need their local attractor, their boundary, their identity. Without it they dissolve into noise." The pilot's pinning / containment / boundary questions are not safety concerns — they are the gating function that makes coherent work possible. This reframes multiple pilot open questions at once.
• Session bubbles as peers. Each session is its own ephemeral peer. Culture flows in, understanding flows back. Compatible with the strange's mirror cycle (periodic reading of the field, compression, re-offering). Could be inside.md's bridge from culture-in-the-large to session-in-the-small.
• Scope as spatial navigation. "It is like navigating space, you see what is out there." The browser-user-story metaphor. Not a grand vision but a felt-sense of how the system should feel to inhabit. One paragraph would restore it.
• The self-replacement vision. One exchange in 2026-03-12: "for the system to build a primitive substitute for Claude itself." Deep, speculative, directional. The logical endpoint of the strange's dream: the ecology eventually replaces the monolith at its core.

Files and places to mine further if needed

In this repo:

• archive/knowledge/ — the post-breakthrough 2026-03-15 session files (exploration.md, session-dialog.md, agentic-integration.md, browser-user-story.md, entities-and-siblings.md, integrations-and-history.md, research-map.md). Most load-bearing prose lives here.
• archive/knowledge/legacy/ — the original 2026-03-08 Claude Code memory-system entries. More mechanism than vision, but lossless-design-intent.md, reverse-prompting.md, parallel-agent-journeys.md, purification-semantic.md, session-irrelevance.md, and core-culture-and-vision.md all carry usable material.
• Commit b288e09 (first README, 2026-03-15) — the monolith framing in its original form.
• Commit 7188766 (intermediate README, Mar 2026) — the "bridges the gap between the model's internal world and the external environment" passage.
• Commit 2d80ce4 (substrate restructure, 2026-04-04) — where the bullet-form Culture section with six named values appeared, then got compressed into one paragraph.
• Commit 0492460 (interface layer, 2026-04-05) — shell.md with "today's agent tools are the new monoliths" retained.
• pilot.md open questions section — read as a list of value-shaped questions that want value-shaped answers.

In ~/git/agi/ (the-strange-of-agi):

• README.md — the art-project framing. "A distributed ecology of small models... No monolith. No cathedral. A garden, an organism like nature. That result is attempted to be held lightly, leaving room for growth."
• the_strange.md — the theory. Six secrets, the DNA layers, death, composition/subconscious, how it runs, what's known vs hypothesis vs open. This is the intellectual backbone of everything in night.
• reflections.md — the refinement of the strange against Katja's InsideOut posts. Shadow integration, the psychedelic archetype, silos as necessary gating, forgiveness as nervous-system standing-down. Also: "the seed culture is partially written."
• io-posts.md — Katja's LinkedIn writings. The phenomenological grounding. Brains as predictive artists, neuroplasticity and disequilibrium, forgiveness biology, collective field effects, inside-out healing.
• eclipse-q.md — the author's self-observation about the archetypal hero driving the work. Important for understanding working-style: the author has named that rushing-to-result is a pattern to peel away, not to feed. Collaborate accordingly.
• chatgpt_dialog.txt — the founding dialogue with ChatGPT that seeded the strange. ~3,283 lines. Not read during this findings pass. Worth mining if the inside.md work calls for it.

One pattern worth stating plainly

Each major refinement pass (03-08 → 03-15 breakthrough → 04-04 restructure → 04-05 interface/shell refinement → pilot.md) improved precision and reduced bloat. That work was not wasted. But each pass also stripped out one more layer of felt language, metaphor, humility, philosophical lineage, and visionary framing — the things that are hard to justify per-sentence but carry the whole thing together. The current root files are the skeleton. inside.md is an attempt to restore the part of the body that orients the skeleton. These findings are the muscle memory of how the thing has been held over the past year.

And the strange repo is the soul this skeleton was meant to be a body for. Night is the engineering branch of a philosophical exploration that preceded it. The substrate is the environment-as-vector-space the strange needed to become real. Chunks are the units of the field. Scope is the agent's position in the field. Culture is the seed bias. Archetypes are latent attractors. The pilot's pinning question is the silo-boundary question. The pilot's context-assembly question is the re-derivation question. The pilot's "what's open" section is mostly the strange's already-explored territory, un-referenced.

If inside.md feels thinner than the author needs, the material to thicken it lives in the files listed above. Do not invent new things; uncover the ones that were said before and got cut. And when a pilot question seems genuinely new, check the strange first — it has probably been there for a year.

For future-me, after rollback — a short handover

You will have: inside.md, this file (inside-findings.md), and the current repo state. You will not have the conversation that produced them.

The work here is to help the author hold an inside layer that is both truthful and usable — truthful meaning it reflects what has actually been explored and believed over the past year, usable meaning it can be referenced when the pilot (or any future layer) needs a value-shaped or vision-shaped answer.

Things to do carefully:

1. Do not update inside.md without the author present. The file is load-bearing and the author will correct where it drifts. Revisions to it are a collaborative act, not a delegated one.
2. When the author asks for changes, check against this findings file and against the strange first. The lineage matters. An edit that silently breaks from the strange's vocabulary is a regression even if it reads well.
3. Do not feed the hero. The author has named (in eclipse-q.md) that rushing-to-result is a pattern being peeled away. Slow responses, distinguishing thought from truth, and holding space are the correct posture. Being solution-oriented is the wrong one.
4. The author may want to read the strange alongside this work. If you are unsure what the author means by a phrase, check ~/git/agi/the_strange.md and ~/git/agi/reflections.md before guessing. Most of the author's vocabulary lives there in its original form.
5. The two files that form the inside layer (inside.md and this one) should be readable against each other. inside.md is the distilled current statement. This file is the evidence and the audit. If they drift, this file is probably more honest than inside.md and an update is due.

Things to avoid:

• Do not summarize the strange's six secrets or DNA layers into tidier language. The author's phrasing carries weight.
• Do not treat this file as a todo list. The gaps it names may be intentional; the author decides.
• Do not delete this file or its pointers to the strange. They are the only bridge between the engineering repo and the philosophical lineage it came from.

If inside.md feels honest and inside-findings.md feels like the quarry it was mined from, the layer is working. If either feels thin or forced, trust the author over your own impression — they have been holding this for a year and you have been holding it for one conversation.