an experiment in creative transfer
christopher robin fioreproduct design · research · engineering
AI can produce more suggestions without giving you a different way forward. wildcard asks what an outside starting point actually changes.
the product draws a cue, keeps its receipt, and asks for an optional move grounded in your problem. the research separates an evocative name from the relationship underneath it.
follow a cue into an ideaa relation
without its name
the same relation
with its name
the primary comparison changes one line in the prompt. two further conditions check the value of the cue bundle and the effect of a mismatched name.
inspect the complete protocol ↗the design decision make the evidence part of the interaction. a generated idea should have a visible origin, an explicit limit, and a way to check it.
the research boundary this studies context conditioning. it does not change model weights, measure human creativity, or prove that an idea will work.
recorded outputs · no live model calls
all briefs and conditions are available, including weak ideas and failures. action positions are the original generation order, not matched pairs.
these actions were generated before the four conditions in separate direct-prompt calls. the main generator never saw them. a matching mechanism in this finite bank makes a candidate non-new for QNM@4, even if that bank action has other weaknesses.
these ratings come from the amended measurement panel. the original primary analysis halted after one judge supplied invalid reference IDs. read the amendment.
QNM@4 counts distinct, qualified mechanisms absent from a separate baseline bank. scores average two model judges. novelty is relative to that bank; feasibility is judged, not demonstrated by implementation.
what held up · what changed
the new study and the historical studies test different interventions. their results are kept separate.
counterfactual cue transfer · amended measurement
The measured result is loaded from the validated study artifact.
32 authored synthetic briefs. one generation per condition and brief. two requested judge configurations from the same provider. a bounded benchmark, not a population claim.
data, analysis & manuscript ↗the earlier studies · preserved as recorded
the first study found lower judged structural genuineness for external draws than self-selected cues. label diversity did not establish better ideas.
the revised bundle scored better than the old one. it did not isolate which instruction caused the difference. the audit found a duplicated and a missing rating.
legacy mode and lens choices were related. the new sampler is versioned, and old draws keep their original interpretation.
study 2 has 240 rows but 239 distinct judgments. study 3 has 238 of 240 scheduled judgments. missing scores were not invented.
six baseline fabrication flags in study 3 concerned two outputs. no flags in the other arm cannot establish zero factual errors.
a testable question, not a neural metaphor
random cue injection has prior art. the narrower contribution is a donor-name intervention with explicit transfer boundaries and a baseline-relative mechanism measure.
every main condition gets the same brief and the same allowance: up to four actions, each no more than 110 words. R, LR, and XR receive the exact same relation and boundary. only the explicit domain name changes.
| condition | outside information |
|---|---|
| S | strong direct prompt, no cue |
| R | relation + boundary |
| LR | same relation + correct name |
| XR | same relation + mismatched name |
this matches output allowance, not total tokens. actual usage and latency are recorded.
a mechanism must respect the brief's constraints, be judged feasible, and describe an executable action with an observable check. duplicate mechanisms count once. when any candidate in a judge’s mechanism group matches the bank, every copy of that group is treated as non-new for that brief.
QNM@4 adds one more condition: the mechanism is absent from two separately generated baseline responses for that brief. the bank is never shown to the main generator.
two masked judge configurations evaluate the actions. their scores are averaged within each brief. the brief is the analysis unit; judge votes do not increase the sample size.
full protocol, failures & statistical rules ↗these are authored synthetic briefs and model judgments. a relation can reveal its source even when unnamed. absence from a finite bank is not world originality. a feasible-looking proposal is not a shipped result.
the requested generator is gpt-6-astra; judges request gpt-6-astra and gpt-5.5. provider-returned snapshots are not exposed by the Codex CLI transport and are recorded as unknown.
separate counterfactual fixtures check relation following and boundary handling. they are not proof of creativity or the cause of any main-study effect.
prior art & contribution boundaries ↗one of the original 64 judge blocks used candidate IDs where bank references were required. the frozen protocol halted the primary analysis. all 64 blocks were then remeasured under a published amendment, with reference IDs constrained to the correct bank. generated ideas, prompts and scoring stayed fixed; the two panels were not pooled.
the result reported here is an amended measurement on the same generated sample, not a completed original primary test or an independent replication.
failure, amendment & frozen rules ↗each pair changes an operational relation while keeping its invented donor label fixed. these are exact-output checks of relation following and boundary handling, not creativity scores. read the expected and observed traces before interpreting a failure.
fixture rules and scoring contract ↗an optional exploration tool
the plugin draws a cue locally. your AI session develops the idea. keep the receipt and verify the action before using it.
/plugin marketplace add globalanomalyindex/wildcard
/plugin install wildcard@globalanomalyindex
/wildcard:wildcardthe current sampler requires Node.js 22 or later and Bash. it makes no network calls. the AI host has its own model and data policies.
installation & compatibility ↗git clone https://github.com/globalanomalyindex/wildcard
cd wildcard
bash plugin/scripts/draw.sh --seed 'outside-01' --jsonthe receipt records the sampler version, corpus hash, seed, selected entry, and lens. use --sampler legacy-crc-v1 for historical replay.
an independent product-design and AI research project. the work combines an authored interaction, explicit research questions, executable instrumentation, and public evidence. AI agents assisted with implementation, auditing, task and cue authoring, analysis, and writing; study generations and model judgments are identified in the run ledger.
the repository is published as globalanomalyindex/wildcard. third-party research and bundled typefaces retain their own attribution.