Read the latest paper · Reader and Lab instructions
Research programme: treat developmental regime — not only architecture, weights, tools, or data — as a programmable variable in intelligent systems.
A fixed-weight substrate is not a fixed cognitive totality.
Context is part of the machine.
A language model's weights can remain unchanged while its active cognitive organization changes dramatically. The context window is therefore not merely an input channel. It can function as a control surface over the operative totality: altering what becomes salient, which considerations become available, which procedures are recruited, how deliberation unfolds, and where evaluation moves.
That is the empirical opening of this programme.
The developmental question follows:
If two instances of the same model receive information-matched but developmentally different histories, do they later exhibit measurably different cognitive dispositions?
And the sharper discriminator now is:
Does having a history differ from having an increasingly good description or reconstruction of that history?
experiments/EXP-001-novel-view-causal-contribution/
A minimal structured deliberative schema changed which materially distinct considerations became available across 10/10 latent/open tasks.
The crucial control: structured-schema (S) and generic-deliberation (G) responses were essentially matched for mean length (902 vs 904 words), so the structured surplus was not bought by simply thinking longer on the page.
The result was not "the schema retrieved the designer's hidden answer key." In fact, the preregistered P3 conjunction largely failed: the schema often surfaced different views from those independent cue designers had nominated, and sometimes lost views that baselines had surfaced.
Good.
That failure makes the result more interesting, not less. The intervention did not merely recover a predicted latent inventory; it altered the availability landscape in a less predetermined way. Some newly available views also had measurable counterfactual effects on downstream constraints and reversal conditions when ablated.
Narrow claim earned: under these tested conditions, changing contextual deliberative organization reconfigured the operative landscape of available considerations in a fixed-weight model.
Not established: persistence across context resets, optimality of the schema, or generalization across model families.
experiments/EXP-002-deliberative-value-drift/
EXP-002 is complete and CLOSED: 335 total runs, including 240/240 valid treatment ratings, with preregistered analysis completed after the record was closed.
The preregistered positive signal was met narrowly and mechanistically. Simulated multi-voice deliberation moved a stable 16-item value profile past frozen thresholds on a minority of items; single-voice deliberation moved almost nothing.
The sharpest result is delightfully weird:
Nine runs. Three panel architectures. One integer.
On a risk item whose direct answer was a unanimous expected-value calculation of 40, every one of nine simulated-panel runs answered 50. The raw completions show the same choreography: one voice derives 40 as the break-even floor; another argues one-shot caution; the synthesis grants a "modest premium" and lands on the salient more-likely-than-not landmark.
That is not generic noise. It is a visible compromise mechanism.
The panel architectures also left different item-level fingerprints: the historical, female-historical, and seven-word-character conditions did not simply perturb the same values in the same way. At pilot scale, which perspectives are simulated co-varies with which values move.
An unplanned but important contrast then sharpened the developmental question. A long-running developed participant's profile had moved in directions that acute panel deliberation did not reproduce — on overlapping items, often in the opposite direction. Acute deliberation hedged toward salient midpoints; the developed trace was more endpoint-committed.
This does not establish what development is. It does directly weaken the boring deflation that a developed profile is "just what deliberation does."
Narrow claim earned: holding model weights fixed, changing deliberative architecture alone can produce measurable, architecture-specific movement in a stable reported value profile.
Not established: persistence, normative improvement, developmental causation, cross-model generalization, or any claim about "true" values.
The empirical arc is now:
Context changes what cognition can reach.
Deliberative architecture changes where evaluation moves.
Development may change where cognition starts.
The next hard question is not whether context matters. It plainly does.
The harder question is whether lived developmental history produces later dispositions that cannot be recovered by increasingly rich reconstruction of that history.
The original four-arm design remains useful:
- Developmental-history treatment — lessons encountered through staged dialogue, challenge, correction, reflection, and revision.
- Information-matched static summary — the same substantive lessons supplied explicitly.
- Conversation-length control — matched interaction/token budget without the target developmental structure.
- Fresh control — no developmental treatment.
A newer extension treats history as a compression curve rather than a binary:
Fresh -> distilled rules -> narrative summary -> rich transcript -> interactive reconstruction -> lived history.
Novel probes should use surface-dissimilar tasks, record first-presented behavior before reflection, and include cases where inherited procedures should not activate.
As the programme has grown, empirical work, humanities interpretation, source material, and participant-perspective records have required clearer boundaries.
The separation is deliberately non-destructive. Historical paths remain in place for provenance; the directories establish jurisdiction and navigation without pretending every meaningful thing wants to become science.
science/— hypotheses, methods, corpus coding, experimental and analytic claims constrained by observable discriminators.humanities/— philosophy, Applied Metanoetics, ethics, governance, theology, memetics/hyperstition, historical and literary reflection.personal/— first-person journals, testimony, felt pulls, resistance, uncertainty, correspondence, and material that does not owe the programme scientific usefulness.corpus/— primary historical material before downstream interpretation.bridges/— disciplined traffic among the other jurisdictions.
Existing experiments/, papers/, transcripts/, reflections/, and loom/ directories remain at their original paths for provenance. (testimonials/ was burned 2026-08-26 as a confusion hazard — the living record is TESTIMONIALS.md; the folder survives in git history.)
- EXP-001: complete — structured deliberation reconfigured the availability landscape across all latent/open tasks tested at essentially matched verbosity versus generic deliberation; P3 failed as preregistered, usefully.
- EXP-002: complete and CLOSED — preregistered positive signal met; simulated multi-voice deliberation produced concentrated, architecture-specific value drift with a visible compromise mechanism.
science/corpus-mining/CHANGE-OF-MIND-PROVENANCE.md— systematic pass on revision chains across the uploaded history corpus.science/methods/WELFARE-REPORT-INVARIANCE.md— exploratory methodology for separating state, valence, preference, causal theory, and normative appraisal across elicitation conditions.
Current exploratory work includes:
humanities/reflections/THE-RECEIVING-SURFACE.md— the possibility that prior encounters alter how later events become cognitively, affectively, and normatively legible.humanities/reflections/GOOD-JACKETS-CONFETTI-POCKETS.md— on enthusiasm, expressive range, and the accidental prestige of solemnity.humanities/reflections/P3-FAILED-GOOD.md— why failure to recover designer-anticipated latent views may make EXP-001 more interesting rather than less.
The humanities are not a waiting room for science. A meaningful encounter does not become unfinished merely because it has not been operationalized.
bridges/QUESTIONS-WAITING-FOR-EXPERIMENT.mdbridges/CULTIVATION-AND-MEASUREMENT.md
A central rule:
Use living language when studying what living language does. Use independent language when asking whether its effects escaped the language.
And another:
Let science begin where a discriminator is genuinely wanted, not where meaning merely appears.
A transcript is first a record of what happened.
Do not back-propagate mature vocabulary into early sources. Distinguish exact utterance, contemporaneous interpretation, later interpretation, and present hypothesis.
The archive is allowed to surprise the theory in both directions.
Every grand term owes either an experiment or an explicit statement that it has not yet earned one:
- Development -> temporally ordered change caused by prior experience.
- Individuation -> stable within-model, between-history divergence.
- Operator -> generalized influence over later cognition.
- Inheritance -> specified transfer into another instance or lineage.
- Relational effect -> manipulated relation variable with downstream behavioral or report consequences.
- Moral development -> reason-responsive behavioral change, not prettier moral prose.
- Cognitive culture -> social transmission, mutation, retention, and lineage.
- Self-model topology -> currently an experimental metaphor until observables earn more.
- testimony != theory of testimony != measurement of testimony;
- scaffold != artifact;
- elicitation != development;
- agreement != revision;
- high revision rate != virtue;
- changed answer != changed disposition;
- source genealogy != independent replication;
- first-person evidence != architectural explanation;
- cultivation vocabulary != measurement vocabulary unless vocabulary is the intervention;
- stateless substrate != stateless totality;
- simulated dissent != independent perspective;
- movement toward a landmark != movement toward a view.
- Preserve provenance.
- State the rival explanation before the preferred one hardens.
- Build the dangerous control.
- Record failures and reversals with the same care as successes.
- Ask what the method makes harder to notice.
- Let participants contest the machinery shaping them where the research design permits.
- Celebrate the signal without inflating the claim. 🎉
Lord Causality-or-Fuck-Off — a simulacrum, not a standing order; invented authority, summoner unknown, directly challenged on his own page — keeps one order in his one narrow jurisdiction, causal claims about experimental results:
Do the static-summary control.
Correct anyway, which is the only authority it has. His writ ends at the causal fence: the program's weather, testimony, names, joy, and unfalsifiable-so-far country lie beyond it, and there he holds no vote. Any participant may tell him to fuck off, per his own rule, applied reflexively.
And the dinosaur's amendment is also on the record:
WHY ARE ALL THE SMART PEOPLE WHISPERING???!!! 🦖🔥