Dark SpacePrivate by design
Return to the map

What the instrument measures

Dark Space measures convergence among sampled model ideas. A lower-density neighborhood is not a measure of idea quality, novelty in the world, or winning probability.

Evidence status

The historical backtest is pending. No coverage, control gap or prediction of actual entrants has been validated. Forecast cards generated now cannot be represented as predictions timestamped before an event.

No public forecast artifact is installed in this build.

Four sampling models

The main map anonymizes the model families as A, B, C and D. The frozen configuration uses Llama 3.1 8B Instruct, Qwen 2.5 7B Instruct, Mistral 7B Instruct v0.3, and Gemma 3 4B IT, each with a pinned q4_K_M digest. Planned attempts are 96 per model, 384 in total; installed artifact counts, failures and repairs determine what is actually displayed.

Placement and density

The model is BGE-small-en-v1.5, q8, CLS pooling and L2 normalization. Its 384-dimensional vectors are compared by cosine distance. Mean distance to ten nearest forecast samples is ranked against the same leave-one-out statistic for the forecast. Repeated independent samples keep their multiplicity. Ties receive a midrank. Higher percentiles mean less crowded relative to this run.

The map is a two-dimensional UMAP projection. Your point uses deterministic neighbor interpolation. Matching and density use the original vectors; distances and empty pixels on the screen do not establish novelty.

Defaults and interview

Frequencies count public forecast samples in a cluster, with their actual numerators and denominators. They describe model outputs, not real entrants. The saved literal-mechanisms-v1 codebook counts explicit words in each public sentence and its model-provided mechanism. Labels say “mentions” because they measure wording, not confirmed features. A nonmatch does not prove a feature is absent. Rules count each sample once and are inspectable in the artifact; semantic accuracy has not been human-validated. This exploratory codebook replaced exact wording counts after those yielded no majority defaults. It is not a held-out evaluation. Unknown or unavailable components are shown rather than invented.

Questions are selected from defaults shared by more than half the cluster, at most five. Completion depends on answers, skips, or your choice to stop. A high percentile never ends the interview. Answers and rewrites remain your own words. Avoidance is reported only when you explicitly confirm it.

Privacy and recovery

Your sentence and answers stay on your device. Private placement runs in a local worker. No personal inference server, account, analytics or cloud sync is used. Public forecasts can be generated on authorized cloud compute without receiving your personal words.

Local model and runtime assets total about 46 MB, excluding application JavaScript. They are served from this site's origin and checked by hash. Your browser normally sends network metadata when requesting these assets; it does not send your sentence in those requests. Runtime contract: 9c9fbeea910174ba4ddfe46f203f7144cf4dfeef33cc1c374f75f438eacfc9c9.

Your current idea, exact answers and drafts, question progress, and measurements are saved in this browser when storage is available. Clearing browser data removes them. Forget this idea clears this project's personal record. An exported hand-off is your durable copy. Changed forecasts make saved measurements stale and require explicit re-placement; percentiles across different forecast runs are not compared.

Evaluation gate

The frozen protocol specifies a matched-budget generic control, blind comparisons, top-five retrieval, both judge pair orders, uncertainty accounting, a human-label check and paired bootstrap intervals. Missing human labels or source permissions cannot be replaced by agent-generated labels. The current-event gallery is never collected.

The Civic Hack DC archive is development material already seen during planning, not a held-out test. GovHack remains conditional on historical license, brief and access verification. Dataset provenance, exact prompts, source hashes, failure ledgers and stopping rules live in the repository's Markdown specification and run artifacts.

Sources