The investigation

A chronological log of attempts.

This is a record, not a roadmap. Each moment carries the question that was asked, what happened, and a link to its evidence. Failed approaches and reframes stay in the log, with the same weight as the results that held.

Status key Supported Signal Completed Inconclusive Retired / reframed In progress
  1. Path A Retired

    The attractor that ended Path A

    QuestionCan a hand-built artificial nervous system, with no language model at the centre, sustain open-ended behaviour?

    ResultNo. It collapsed into a single deterministic attractor and stopped changing. The approach was retired.

    Notebook entry 002 →
  2. Path B Reframed

    A body the model can feel

    QuestionIf the model stays at the centre, can the substrate (body, world, memory) carry continuity instead?

    ResultThe project was reframed around a stateless model kept generating without stop, with sensation injected into the running stream rather than asked as a prompt.

    Notebook entry 003 →
  3. Cycle 1 Supported

    A phrase that outlived its organism

    QuestionDoes anything persist across lives?

    ResultA phrase propagated across roughly eight lives with mutation. Transmission with drift, found by grep, not open-ended evolution.

    Notebook entry 004 →
  4. Cycle 1 Completed

    Moving the life inside the model

    QuestionCan the held KV-cache itself be the identity, carried in-process rather than re-fed as text?

    ResultA native runtime was built: a life is the model's own held cache, saved to disk and restored.

    Notebook entry 005 →
  5. Cycle 2 Supported

    The same thought, twice

    QuestionIs a restored cache faithful, or does restoration quietly change the organism?

    ResultToken-identical: 8/8 decoded actions matched across a fresh process. Restore is deterministic under fixed seed and temperature.

    Canary log →
  6. Cycle 2 Supported

    A structural attractor, under control

    QuestionIs the cache attractor a real structural effect, separable from the prompt text?

    ResultControlled discontinuity breaks it (0/3 vs 3/3 final-window locks); a text-only cue alone does not. The effect requires the real held cache.

    Finding →
  7. Cycle 2 Inconclusive

    The claim we did not make

    QuestionDoes the organism actively maintain itself: a first test toward operational self-maintenance?

    ResultA small enabled-versus-baseline asymmetry that did not survive scrutiny, and was later found to be confounded by a copyable instruction in the cue. Not published as a finding.

    Notebook entry 007 →
  8. Cycle 3 Completed

    Dropping the deadline

    QuestionPublish to a conference deadline, or self-pace the investigation?

    ResultThe deadline was dropped. Results are posted here when they are undeniable, with guardrails set against overclaiming.

    Notebook entry 008 →
  9. Cycle 3 Inconclusive

    A self-maintenance signal, withdrawn on review

    QuestionWith the copyable instruction removed from the cue, can the organism close a self-maintenance loop on its own: notice a decayed self-belief and act to restore it?

    ResultOn first analysis a loop appeared to close, and a belief was refreshed (rehearsal 0 to 1, confidence 0.29 to 0.80). On review the closed-loop credit did not hold: a scorer clock-ordering coincidence, and a control that matched a trivial repeated-probe latch. What stands is narrow, and not on its own a self-maintenance signal.

    Finding F-04 →
  10. Path C Reframed

    The model demoted to an organ

    QuestionCan the language model itself be the organism, its running text the mind of the creature?

    ResultNo. A transformer is inert between calls and reads text, not a world; trying to make it be alive was a category error. We demoted it to a bounded brain inside a mortal body, and retired the held-cache-as-mind line.

    Notebook entry 010 →
  11. Path C Supported

    Does the brain actually do anything?

    QuestionWith the model demoted to an organ that only nudges the body, does it do anything, or could noise replace it?

    ResultIts choices beat all four controls, including randomized-weights and valid-random, on a survival task, across two model families (p ≈ 0.006). The learned prior is load-bearing for survival. A finding about the body's dynamics, not a life-property.

    Finding F-06 →
  12. Path C In progress

    Generalizing the result

    QuestionDoes the load-bearing prior hold on a structurally different task, and against a stronger randomized-weights control?

    ResultIn progress. A second task where the body must forage rather than just repair, and a stronger weights control. This is the current work.

    Findings →
  13. Path C Signal

    Online weights begin to change outcomes

    QuestionDoes letting the brain's weights slide online change the body's outcomes, beyond a frozen brain and beyond no brain?

    ResultIn Dolphin (n=12) the online arm beat matched frozen-brain and no-brain controls on final energy (p ≈ 0.003 and p ≈ 0.0002), reproduced directionally in Qwen (n=3, p = 0.125). A final-energy advantage, not a survival claim, and not artificial life.

    Finding F-07 →
  14. Path C Inconclusive

    Expressible, but not yet selectable

    QuestionCan a memory-bearing drive express richer heritable behaviour, and can selection maintain it?

    ResultYes to the first: behavioural dimensionality rose from about 2 to 5, genome-driven and heritable. No to the second: under priced selection it collapsed back to 2. Expressibility and selectability are separate bottlenecks, and the ecology presents only about two resource channels.

    Findings F-08, F-09 →
  15. Path C Refuted

    The falsifier that caught us

    QuestionIs the live-contrast advantage from stall-coupled bursts adaptive, or could matched-volume random churn reproduce it?

    ResultRandom churn matched it: 41 advances stall-coupled, 42 matched-random. Burst timing is not yet a defensible adaptive signal. The falsifier stopped a tempting over-read; both arms still failed the open-ended-evolution gate.

    Finding F-10 →
  16. Path C Signal

    A signal that survived its falsifiers

    QuestionIs the substrate novelty tied to real geometry and resource structure, and does it hold under a longer horizon?

    ResultA stricter geometry-retained readout rose from 1 of 12 seeds to 3 of 12 as the horizon lengthened, while shuffled and destroyed-structure controls stayed at 0 of 12. It strengthened rather than washing out, and broke under shuffled or destroyed structure. A broader novelty metric still leaks, the open-ended-evolution gate still fails closed, and nothing was promoted. No artificial-life claim.

    Finding F-11 →

The log is the argument: a sequence of falsifiable questions, answered in public.

Not: a long log is not progress toward life. Each cycle is gated on the one before it surviving its controls, and the current experiment has not cleared yet.

How each result is judged →   The evidence behind them →