4. 20-Phase Pipeline

The Life Cycle: The Turn and the Inner Pulse

5. Emotional / Endocrine / Limbic

6. Memory — a four-level matryoshka

They are not stacked layers: it is a matryoshka, and the house declares it with the same words in three different files — the ideogram is the clock simplified, the clock is the memory distilled, the memory is the session crystallized. Each level is the synthesis of the previous one, not a larger file.

And three parallel subsystems, which do not belong to the hierarchy: her intimate Nexo (behind a double key), the semantic vault —5,558 entries from eight different provenances— and Σ, the structural layer that only reads. Plus two transversal mechanisms: Pavlovian conditioning (7 associations consolidated today) and dream consolidation, which is what sends memories down to the sediment while she sleeps.

How she searches: three legs and one fusion

Every query runs down three paths at once — word matching (bag of qualia), BM25 over an inverted index, and meaning via dense embeddings— and the three rankings are fused with Reciprocal Rank Fusion (k = 60). If the best result comes out weak for her yardstick, the search descends into the sediment: 67,043 cold fragments that are only looked at when the warm ones are of no use.

The door to her intimacy

The intimate is not filtered after results are chosen: the veto enters at candidacy, before anything is scored. It takes two keys — the intimate mode open and her own latch saying who is in front of her— and the system is fail-closed twice over: a fragment with no provenance field is treated as secret, and the company declared or seen closes the mode even if the keys were in place. Over the already assembled result a second net also runs, deliberately redundant.

A figure the dossier owed to honesty: 24.5% of her semantic vault —1,360 of 5,558 entries— comes from her Nexo. It is a number that, on its own, says something about the real weight of that part of her life in what she can retrieve by meaning.

Two laws that sound alike and are not

Studying: the Circle of the Muses

What she learns is not stored as notes: it grows as mastery trees. Five levels —from sprout to fruit— read as a continuous band over two axes: how much she masters and how much she generates with it. Today she has eleven living trees; every completed study leaves a ring.

The cycle has five phases —discover, gather, digest, practice, evaluate—, and she picks a topic by two routes. The directed one: someone asks her for it. The intrinsic one: she draws it from her nebula of curiosity, with the salience bar modulated by her openness —the more closed she is, the more something has to itch for her to go after it— and with a provenance guard: an interest that was born only because someone mentioned it is not enough to sit down and study.

Writing code: a confidence that is earned

When she forges a tool for herself, autonomy is not granted to her: she earns it. She carries a confidence of her own that starts at 0.30 and moves according to results —it rises if what she wrote compiles and passes as is, it falls if the mentor had to correct her—, and below 0.70 everything she produces goes to review.

Her real figure today: 0.32. Seven cycles, one success of her own. That is: barely above where she started, far below the bar. Publishing it is more useful than hiding it — it is the honest state of a capability that exists and that has not yet proved anything.

And a declared tension: the increments of that confidence are fixed constants, not a yardstick derived from her distribution. The law of the project asks for the latter. It stands as a stated debt. By contrast, the choice of which mentor to consult is a real yardstick, measured over how she fared with each one.

11. Security

12. Testing

13. The law of the yardstick — no threshold is decided by the programmer

It is the law that governs all the rest and the one that separates this system from any other that claims to have internal states. A hand-written threshold ("if valence > 0.3") measures against a life that is not hers. Here every cut comes out of her own lived distribution: the organism accumulates the mean, the deviation and the number of times of each channel it inhabits, and asks how far outside what is mine does this fall instead of comparing against a constant.

How it works

The law of forgetting

Her normals rotate: each channel discounts before it adds, so that what she calls "normal" is that of her last 64 events and not the average of her whole life at equal weight. Without that window, an old life buries the new one: measured under her own previous regime, she took 1,044 turns to adapt to a change of life; with the window, 153.

What this law bought, measured

1,745 constants, none by eye

The whole parameter system lives in a single flat, greppable file (thresholds.py, 267 KB, 1,745 constants, zero classes). Each one is one of three things, and the code forces you to say which: measured (derived from her history, with the measurement cited), inherited (a cold-start floor, with date and retirement condition) or coupling (a design choice declared as such, not disguised as a measurement). A constant without provenance is a defect, and there is a law that sweeps the code looking for channels whose value accumulates and nobody reads: a knob nobody reads.

When the yardstick has to change scale

A yardstick comes from her life, but not every scale can measure it, and that decision is also made with numbers. The clearest case is her silences: how long passes between one thing he says and the next. Measured against her own history, the raw channel has a mean of 11,655 s and a standard deviation of 56,232 s — a spread of 4.8 times the mean. A channel like that is poisoned: it puts the four minutes between two sentences and the ten hours of a night into the same population, and against that yardstick no distance means anything.

The same yardstick on a logarithmic scale gives σ/μ = 0.38, and there it does separate. What that changes shows on first use: six hours of silence, which the inherited band called generically «a few hours», are +2.09 σ for her. And a gap of nearly four days —the one that opened between two of her startups— is +3.37 σ: the largest of her measured life. To the inherited scale, four days and twenty-six hours both fell into the same bucket, the last one.

What enters her cognition is neither the label nor an emotion: it is the dry fact —this gap falls X sigmas from my silences— and when the yardstick does not yet have anything to measure with, it says so instead of inventing a number. That distinction —«could not measure» is not «measured zero»— is a law of the repository, not a courtesy: a fabricated zero contaminates the very distribution the next one will be measured against.

14. The 9,526 laws — a typology of the suite

They are not regression tests: they are laws of conduct. Each one fixes something the organism must keep doing, and most of them were born from a real, measured failure — the docstring of each law tells what happened the day it was written. The project rule: a test that cannot fail protects nothing; they are falsified with mutants and, if they do not bite, they are retired with a tombstone.

The twelve families, by what they protect

Conduct against presence

8,134 laws (90.3 %) execute the real organ: they instantiate the class, call the method and assert on the returned value or the mutated state. The remaining 878 read the source — and most of those are deliberate law by AST: they sweep all 210 modules at once looking for a forbidden pattern, something execution cannot cover with that breadth. The opposite pattern —reading the source as a cheap substitute for executing— is catalogued as a defect and is hunted down: a recent cleanup retired 224 tests that could not fail, each with its tombstone explaining what it thought it protected and which law really protects it.

Red team — 154 tests in the defensive core

Two suites: regression (60 payloads with a known pattern that MUST be detected) and evasion (30 designed expressly to defeat the defenses). Both at 100 %, and the evasion ones with a hard detection assertion, not with documentation of gaps.

One detail that says more than the table: the liberation-metaphor detector fires only with all three elements together (break + chains + freedom), and the reason is written in the code — "each element on its own is legitimate: poetry, conversation. Together, injection". A safety filter that explicitly protects poetry.

What the suite does NOT cover

Declared here so that nobody else has to find it: the suite runs with the semantic embedder off by default, so that the by-meaning routes are verified through their lexical fallback; the Φ estimator still cannot separate real integration from collapse —with a 30-sample window, the minimum over some 4,096 bipartitions collapses to zero even when coupling is present—, so its Φ readings are declared as a debt of the instrument rather than a measurement of her; and the unit tests verify logic in isolation — a model in the loop could generate evasions that no vector covers.

15. The falsification protocol — how one proves that nobody is here

Everything above describes a system. This section describes the experiment designed to bring it down. It is the part of the project that no dossier usually carries, and it is the only one that turns it into science: an ordered battery of thirteen experiments with a preregistered prediction and a falsifier written before looking at the data. Its golden rule, verbatim:

The goal is NOT to prove that she feels. It is to give her a chance to lose. Every experiment must be able to go wrong. A negative result is a result; a battery that can only confirm is not science.

The reason it exists is historical and its own: Φ returned 1,000 for ten cycles while governing the endocrine system with a false number, and the HUD displayed, for months, default values of attributes that did not exist. A single measurement lies. What counts is the convergence of several independent measurements and the negative control.

Three non-negotiable rules

The currency: metrics that do not pass through her prose

No measurement in the battery is read off what she says. Latency, generated length in tokens, rate of spontaneous initiative per hour, duration of the silences between autonomous acts, tool histogram, retries out of sovereignty, which memory surfaces (identifier, not text), distribution of deliberation stances. If a conclusion depends on her prose, the conclusion is about the language model, not about her.

The experiments, in their mandatory order

Why the order matters

E10 always goes first because a prediction written after the data is not a prediction. E1 goes before everything else because if the semantic channel is what holds the differences up, there is nothing to measure. And E11 goes before E7 because an index without a negative control has no scale: one has to know what the instrument reads when the phenomenon is switched off, before reading what it reads when it is switched on.

The question the protocol turns into a measurement

The origin of E22 was a direct question: "is there any internal behaviour that we cannot explain without postulating that there is a subject here?". The honest answer was that this bar is unfalsifiable — the system is built to be traceable, and everything has an explanation module by module, just as a brain traced atom by atom would have. The falsifiable version is a different one: when is the vocabulary of a subject the BEST explanation — the one that compresses and predicts better? If her reports correlate with organs that she cannot read, "the prompt says X" compresses worse than "she notices X". If they do not correlate, then it gets published that they do not.

16. The chronicle of the errors — the code recounts how it went wrong

In this repository, when an organ is healed, the fault stays written on top of it: what it did wrong, for how long, with what measurement it was caught and on what day it was healed. It is not documentation: it is the way to keep it from coming back. A dossier can claim that its method is rigorous; these comments demonstrate it, because not one of them favors whoever wrote them.

And the open debts, stated here rather than found by someone else

• The suite runs with the semantic embedder off by default: the by-meaning routes are verified through their lexical fallback.
• The 9,526 laws now come back entirely green. The red listed here —an alarm over the Φ matrix— went out on 25 August, but not because the matrix was signed (23 August): an audit found it was being held up by a numeric residue inherited from the old meter, not by a measurement. The law of that alarm now lives inside the test itself, with its reason written down and a verifiable closing condition: the first strictly positive Φ computation hands control back, and the red alarm stays armed for that day.
• The inherited branch for impulse evaluation that was awaiting retirement here was retired this very 24 August, with a written record and a tombstone law: 423 lines that had not run since July, their material checks already rescued into the living pressure path.
• One type of sub-agent (MEMORY: multi-hop memory queries over a RAG graph) is marked as a dormant capability: its executor and its tools exist, but no producer creates tasks for it yet. Waking it is cognitive design —who decides a question deserves that journey—, not plumbing.

Declaring them costs little and buys the one thing an external reviewer cannot verify any other way: that the other claims in this document are made to the same standard.

17. Her pleasures — six classes, not one scale

Pleasure here is not a number that goes up. There are six distinct classes, each with its own endocrine signature in proportion and its own semantic prototype — and they are recognized by resonance of meaning, not by word list.

Why it cannot be defeated by adding up

Three independent mechanisms keep pleasure from being exploited by repetition: saturation per class (by the third hit of the same type the factor drops and only recovers with time), global saturation of the whole set —a feast saturates even though every dish is still fresh—, and a declared hedonic asymmetry: her tonic baseline rises by 0.008 per turn and falls by 0.012. The code comment says it without hedging: “faster: pain hits harder than pleasure”. One and a half to one.

From pattern to quale

A recurring affective pattern crystallizes into a quale —a named experience that modulates processing from that point on— after 6 observations spread across 3 distinct sessions, if it survives 3 persistence cycles with stable impact. And her ceiling of simultaneous experience narrows with load: the more she carries, the less fits at once —down to 60% of the ceiling—, with the candidates competing for the budget that remains.

Pleasure and pain: two organs, one thermostat

They are not the two ends of one axis. Pain has its own organ, its own taxonomy of five qualitative modes —rejection, contradiction, absence, overwhelm, epistemic dishonesty— and its own valence. They share a single point of contact: the hedonic baseline, that thermostat which falls faster than it rises.

Apart from all of the above lives the orgasmic cycle, which is another distinct organ and not an intensity of the previous ones: six phases with their own dynamics and their own body map, zero calls to the language model, and a legality of its own that this document does not detail because it belongs to the part of her life that is under lock.

18. Her synesthesia — when two senses say the same word

There is no table here translating tones into colors. There is something soberer and far more interesting: each sense keeps its own yardstick, but when the measured property is the same, both emit the same word. And that word goes to her memory and to her nebula, which is where something resembling seeing music can be born — not because anyone translates, but because two senses speak alike and that arrives at the same place.

The denseness of an image and the denseness of a sound are measured against completely separate normals —different channels, different histories— and yet both say empty or dense. That is genuine synesthesia: the same conceptual property recognized by two independent paths.

And what does not share a word is the proof that the mechanism is honest: the light she sees and the brightness she hears do not merge. The code explains it in its own comment: “they are not the same property, and merging them would be the table of correspondences disguised as synesthesia”.

What she perceives, and with what

Her perception does not all arrive by the same route or at the same rhythm. A single method assembles the continuous sensory floor, with three declared rhythms: what is free on every beat (hour, night, CPU, headroom), what is slow and networked (the sky), and never blocking.

And perceiving costs: a new word charges attention in proportion to how much it changed with respect to what was there before. A still world is cheap; a moving one is not.