Explore a vague goal through competing archetypes. Inspect the semantic weight of a single word, test counterfactual changes, and revise the goal without pretending a word is a separable neuron.
01 / Goal laboratory
Ambiguous inputs are intentional. The goal may change after each round.
Without a valid key the app uses a clearly labelled local illustration. The key is not persisted.
Ready to explore.
Mode: model-backed when available; labelled local fallback otherwise.
02 / Word-level deliberation
Choose any word. The system evaluates it within the whole statement rather than treating frequency as importance.
Selected word significance—
Select a word to inspect.
Illustrative semantic editing only. This does not modify model weights or residual activations.
Original
Counterfactual
Micro-debate
The selected word's role, objections, and preservation decision appear here.
03 / Archetype council
0 rounds
No deliberation yet. Enter an objective and begin the first round.
04 / Counterfactual Shadow Space
Server geometry
Headless controller maps competing concepts on a signed opposition axis. Other axes encode independent sources: the user goal, archetypes, and stored revisions. The shadow pool holds routes explicitly proposed but not selected.
Enter the private AI key above to load stored research.
Conceptual geometry, not transformer residual activations. Opposing labels are model proposals; grouping uses embeddings where available.
Route ledger & rejected alternatives
No route evidence loaded.
Headless micro-deliberation
Select any word above or enter one here, then evaluate its relationship to the recorded routes.
Provenance/source axes
No source references loaded.
05 / Measured decoder candidates
Top-5 observed
Each headless debate runs a separate, deterministic next-text probe. These are actual conditional token probabilities, not probabilities of the philosophical routes.
No evidence loaded
Select saved generation—
Conceptual route connections
No links inspected.
An unchosen top-five token was available but not emitted, not philosophically rejected. Unlisted tokens are not observed. No complete alternative branch was executed.
Research status: this prototype provides genuine LLM-generated archetype responses when the hosting inference service is available and a transparent deterministic fallback when it is not. Word-level scores are heuristics, not measured token probabilities. It does not perform conceptor projection, KV-cache rollback, hidden-state probing, or validated latent steering. Private Cloudflare D1 stores up to 64 debate records, objectives, proposed routes, and geometric counterfactuals. The browser-only preview and headless backend have different evidence levels. Implementation notes.