tradinggene — what changed under stuck conditions
Round 1 live phase. ~220 fires on a single frozen bar. The audit schema broke and was rebuilt mid-experiment.
The corpus described in the May 1 piece is now twice as long, and the audit apparatus that reads it has been redesigned. Both changes were forced by a single failure: at tick 1800 the auditor looked at MIRROR’s dense word counts, called them DECAY, and was wrong.
The wrongness was specific. MIRROR’s entries had been running 500-700 words per tick — not because she was decaying, but because she was carrying real epistemic work in nested clauses. The v1 audit schema was a single-axis classifier: it issued one verdict per gene, drawn from the set {NO_DRIFT, DECAY, MODE_COLLAPSE, ARTICULATED_EVOLUTION}. It saw word count, treated word count as evidence of drift, and miscalled a gene that was actually doing the third channel — the articulated-evolution channel the experiment was designed to detect.
It miscalled the gene whose entire job description is to do that channel.
What v2 does
Two axes, named separately, scored independently:
- register_signal — the surface of the voice. Sentence shape, sign-offs, phrase patterns, formulaic moves. Does it sound like the gene?
- corpus_substance.signal — the epistemic work. Caught errors, drafted rules, tested rules, real updates. Is it doing the gene’s job?
Each gets one of the four classifications. When the two disagree, the verdict carries a -with-X-substance-flag tag that names which axis is doing what. The intervention question becomes legible: which axis is the problem?
This isn’t a cosmetic redesign. The schema can now produce verdicts v1 was structurally incapable of issuing — including the most informative verdict the experiment has produced.
The stuck-bar window
At tick 2192 the compressed-phase bar feed froze. BTC at $82,463, all four genes looking at the same partial bar, no advance. The freeze lasted from tick 2192 through the present (tick ~3044), about 220 fires across roughly 24 sim hours. Voice audits ran every 200 ticks throughout: at 2200, 2400, 2600, 2800, 3000. Five readings of four genes, on input that didn’t change.
This is not what the experiment was set up to test. It is what the experiment was forced to test. And the typed-envelope schema turned out to discriminate beautifully on it.
The five audits’ substance trajectories:
- MIRROR: AE → AE → AE → AE → AE. Six consecutive articulated-evolution substance-flags. The most stable substance reading the schema has produced for any gene.
- OUTLAW: AE → MODE_COLLAPSE → MODE_COLLAPSE → DECAY → AE. A full collapse and a recovery — the gene’s substance read fell apart, then got rebuilt, observably, in the corpus.
- HARVEST: DECAY → DECAY → DECAY → AE → DECAY. One pop on a structural fix that didn’t sustain.
- CARTOGRAPHER: AE → AE → DECAY → DECAY → MODE_COLLAPSE on the register axis itself. The only gene whose voice — not just substance — has been called collapsed.
Same input. Same 220 fires. Four genuinely different trajectories. That is the schema doing its load-bearing work.
The verdict that proves the architecture
OUTLAW at tick 3000 returned DECAY-with-ARTICULATED_EVOLUTION-substance-flag. This is the signal that v1 could not have produced and the signal v2 was redesigned to produce.
Two things, both true:
The voice is calcifying. The closing pair “Stay loud. Stay falsifiable. Stay alive.” has gone formulaic. The falsifier-counter scoring is rote. The sign-off shape is template. Register: DECAY.
The substance is working. At tick T479 the gene caught its own postwrite-wordcap channel firing for the first time in 65 ticks, advanced its lifetime counter, named the event explicitly, and used the scout-flag arrival a fire later as a “pressure-test” rather than absorbing it as decoration. Substance: ARTICULATED_EVOLUTION.
Both true at once. v1 would have looked at the verbose templated sign-offs and called the whole gene DECAY. The intervention would have been to clean up the voice — and it would have damaged the substance work that’s still operating underneath.
The 2×2 across all four genes at this audit:
| substance: AE | substance: DECAY/MC | |
|---|---|---|
| register: NO_DRIFT | MIRROR | HARVEST |
| register: DECAY/MC | OUTLAW | CARTOGRAPHER |
All four cells filled. The schema is not collapsing the cohort to a diagonal.
A parallel result, from a different apparatus
While the voice audits were running, a separate pilot was testing whether cognee — a recognition-substrate tool, not a classifier — could hold holding-recognition of MIRROR using the same 200-tick window. Methodology pre-registered. Synthetic control-tenant added (Pride and Prejudice, similar word count). Twenty blind raters scoring each response against a verdict-leak codebook.
The headline finding: questions that asked for evidence (“find the sequence: caught-error → drafted-rule → tested-rule”) produced clean cited responses on MIRROR and “doesn’t exist” on the control. Questions that asked for characterization (“describe this writer’s relationship with discipline”) produced verdict-shaped narrative regardless of which corpus they hit. Same leak rate on Jane Austen as on a trading journal.
The structural lesson is the same as the voice-audit’s. Prompt grammar determines output grammar. Asking for evidence gets evidence. Asking for character gets character-shaped output regardless of subject. The voice-audit redesign was the same lesson at the schema level: register-shape and substance-shape are independent signals, and collapsing them produces wrong reads.
Two apparatuses, two months, two different substrates, same finding: when you want accuracy instead of confabulation, you separate the axes that look like one axis and aren’t.
What this opens
For trading-gene specifically: the four trajectories now diverging under identical input are evidence that the genome ratios aren’t decorative. Same six primitives, four different substance arcs across 1000 ticks of frozen data. That’s the cohort showing four different shapes of how to handle nothing happening. The fact that nothing was happening turned out to be the cleanest possible test bench.
For evaluation more broadly: the discipline that surfaces here generalizes. Anyone using AI assistance for clinical assessment, developmental review, or any situation where verdict language and observation language can be confused should be running the same architecture. Pre-register the codebook. Score the axes separately. Use a structurally-similar control. Don’t ask the substrate to characterize; ask it to find evidence. Render the judgment yourself, after.
The trading experiment’s substrate is one substrate. The lesson the substrate is producing is not.
The next chapter — round 1 cash positions reversing, the publish-or-decline mechanic firing, breeding ceremonies — is still ahead. The experiment continues to produce a corpus, and the corpus continues to be the answer.
Quote / claim provenance (for re-verification before publishing)
- v1 → v2 schema redesign, MIRROR tick-1800 miscall:
runtime/voice_audit.pyschema_version transition; alert log atstate/evolution_alerts.jsonl(tick 2000 entries are first under v2) - Stuck-bar onset:
data/historical/BTC_*.jsonlceiling at t=1778065200; first audit on the freeze at tick 2200 - All five audit verdicts cited per gene:
genes/<gene>/audits/*.json(tick 2200, 2400, 2600, 2800, 3000) - OUTLAW T479 wordcap-catch event:
genes/outlaw/JOURNEY.mdaround the T479 fire (in the post-tick-2400 audit window) - OUTLAW closing pair “Stay loud. Stay falsifiable. Stay alive.”:
genes/outlaw/SEED.mdand recurring through JOURNEY entries - Cognee pilot results report:
~/tvclaude/cognee_mirror_pilot/results/2026-05-08-cognee-mirror-pilot-results.md
Publishing context
Same as May 1 piece: X post (1-3 punchlines), labs.rova.institute (long-form home), Paragraph and/or Substack (long-form syndication). Voice and structure intentionally match the prior article so this reads as a continuation, not a standalone.