Research: hypar as an ideas lab

Status: research (2026-11). Nothing here is implemented. It maps what exists, what the evidence says, and which questions must stay open while the product is defined. The product intent this research led to now lives in Direction. Settled choices still go to Decisions.

The proposed direction: hypar becomes a lab for ideas. The human plants embryos; an LLM helps research and validate them, using classic creative / continuous-improvement methods and newer human–model processes, so ideas reach a useful outcome faster.

This changes one explicit non-goal today: the agent has no external knowledge (method as process, The Agent). Opening it to the web is the biggest decision below.

---

1. Landscape — what is similar

Family	Examples	What they do	Gap hypar could own
Idea-validation SaaS	ValidatorAI, IdeaProof, Preuve and "AI co-founder" platforms	Score a startup idea: market, competitors, a report	One-shot reports, mostly vendor-graded. No life cycle, no memory of dead ideas, product ideas only
AI scientists	Google Co-Scientist (generate → reflect → Elo tournament via debate → evolve → meta-review), Kosmos (shared world model, every claim linked to code and citations)	Autonomous hypothesis generation and validation	Autonomous, domain-specific (biology). The human is a reviewer, not a co-thinker
Tools for thought + AI	Heptabase (canvas, AI explains and researches), Tana (AI chains actions), NotebookLM	Store and connect knowledge, AI per item	Storage first: AI summarises and organises, it does not challenge. No notion of an idea maturing or dying
Provocation research	Sarkar, AI Should Challenge, Not Obey (CACM 2024); Microsoft "lenses + provocations"	AI as provocateur: short critiques, alternatives, risks	Research prototypes, not a product. Closest to hypar's thesis

Read: nobody combines lifecycle + challenge + memory of failure + validation against the world. Validators are external and one-shot; AI scientists are autonomous; PKM tools are complacent. hypar's spot is a human-led lab where the agent challenges and searches for evidence, and the idea's history (including fossils) is the asset.

---

2. What the evidence says (and why it matters to hypar)

1. Sycophancy hurts thinking; friction helps. Sycophantic replies raise confidence and lower open-minded thinking; neutral replies feel uncomfortable but support flexibility (ACL 2026). Sycophantic models mislead novices (Invisible Saboteurs). Literacy interventions alone do not fix it, so it has to be designed into the system (AIED 2026). → Validates ADR 0002: one challenging question is a product feature, not a style.
2. LLMs homogenise ideas. Individuals get better ideas, but across people ideas converge (Anderson et al. 2024, Moon et al.). Mechanisms: the model fixates on early outputs, and all users share one distribution (Deng, Brucks & Toubia). Better prompts do not escape the shared space (IDEAFix). → The VARIETY stance is necessary but not sufficient. The diversity source must be the user (their seeds, second-brain, fossils), not the model.
3. Research agents narrow exploration. 219k AI ideas were more concentrated, closer to the starting literature and lower-impact than human follow-on work (arXiv 2605.27905). → If hypar adds web research, the agent will tend toward local elaboration. Research should feed contrast and refutation, not "here's what exists, build on it".
4. LLM judges are biased and overconfident. Self-preference is measurable, verbalised confidence is miscalibrated (DINCO), and a self-review pass helps (EMNLP 2026). Standard advice: a judge from a different model family, audited against human labels. → Relevant to the Jev question below.

---

3. Methods — the repertoire

hypar already treats methods as invisible agent moves, not a UI stepper (DEFINE · PROBE · INVERT · VARIETY · SIMPLEST). The proposal is to expand the repertoire, not add a wizard.

Classic (well proven, cheap to encode as moves)

Method	Where in lifecycle	Possible move
5 Whys, problem reframing (Munari)	LATENT	DEFINE (exists)
Pre-mortem ("it failed in a year; why?")	GERMINATING	INVERT (extend)
Assumption mapping / "what must be true" (Lafley/Martin)	GERMINATING	PROBE → assumptions as tensions
Toulmin argument (claim, warrant, rebuttal)	GERMINATING → GROWING	WARRANT: "what links your evidence to your claim?"
SCAMPER, analogies, TRIZ contradictions	GROWING	VARIETY variants
Six Thinking Hats / red team	GROWING → MATURE	rotate the critic's perspective
PDCA / Lean build-measure-learn, falsifiability (Popper)	MATURE	TEST: "the cheapest experiment that could kill it"
Retrospective / post-mortem	FOSSIL	the fossil reason (exists)

New human–model processes (riskier, more distinctive)

• Provocations, not answers (Sarkar): short critiques attached to the idea, not a chat. Fits HITL notes.
• Debate tournament (Co-Scientist): paths compete in pairwise debates; the human is the referee. Fits GROWING with paths[].
• Evidence-linked claims (Kosmos): every agent claim from the web carries a source; unsupported claims are marked as such. Fits "claims match runtime".
• Refutation search: the agent searches the web for reasons the idea is wrong or already done (prior art, counter-evidence), not for support. Directly counters findings 2.3 and 2.4.
• Personal diversity: draw variety from your corpus (second-brain, other embryos, fossils) instead of the model's average. This is the existing saved references as contrast experiment, which the evidence supports.
• Calibrated gates: a separate, fast critic gives typed verdicts with confidence (ready to mature? prior art exists? kind of death?), and the human decides. See Jev.

---

4. Jev (TypeSafe AI) — does it fit?

Source: TypeSafe AI, Introducing System One models and Jev (2026, early access).

What it is: a "System One" model. It takes unstructured state and returns typed decisions with calibrated probabilities, not free text. It is trained with RL for Calibrated Decisions. It answers in roughly 70–500 ms, with input at about $0.04 per million tokens and free output (vendor figures, possibly subsidised), and up to 255 options per decision.

What it is not: it does not generate text, so it cannot ask hypar's question. The blog makes no claims about impartiality or criticism. Its "calibration" claim is self-evaluated, with no public standard benchmarks. Early access only, closed weights.

Where it fits: not as the collaborator, but as a fast, calibrated classifier/judge beside it:

Use	Typed output	Value
Stance selector	DEFINE / PROBE / INVERT / VARIETY / SIMPLEST + p	Picks the move before the LLM writes; tests "stance is theatre"
Solution-shaped seed detector	bool + p	The heuristic method as process Q3 left open
Maturity / fossil gate	ready / not ready, ILL_DEFINED / WRONG_PATH / SUPERSEDED + p	Suggests transitions with confidence; the human still decides
Connection typing	relation enum + p	Cheap triage before the LLM proposes a link
Path tournament judge	A / B + p	Cheap pairwise judge from another family (less self-preference)

Verdict: worth a bounded experiment as the System 1 (gates, routing, judging) beside the LLM System 2 (question, research). Its calibration must be checked against your own labels (evals/embryo-stance.jsonl already exists) before trusting it. Risks: early access, vendor lock-in, and unverified claims. Note too that "impartial" is not guaranteed by a model; it comes from separating roles (the critic is not the generator) and using different families.

---

5. Insight from the conversation: outcome type varies per idea

The maintainer's answer about deliverables: "it depends — an HTML component for a UI idea, a doc or a repo for a knowledge-storage idea, a skill for a human/machine working method".

This is a product statement in itself. An embryo that matures produces an artefact whose type depends on the idea. Today MATURE leads only to fossil or to staying alive. A lab needs an exit that is not death: graduation into an artefact (prototype, doc, repo, skill, experiment).

---

6. Open questions to keep open now

Answered by the maintainer (2026-11-05), pending an ADR:

1. External knowledge — yes. The agent searches the web, both to refute (prior art, counter-evidence) and to extend (evidence, adjacent work). Every web claim carries its source. Still open: how to balance the two so research does not collapse into local elaboration (finding 2.3).

2. Validation is not one process. It depends on the idea, the proposed solution and the end user. So there is no fixed kind → validation recipe: the agent has to infer how this idea should be validated (problem definition already does part of this). Still open: whether that inferred validation plan is stored, and whether the user can edit it.

3. Roles inside, one voice outside (accepted). Collaborator, researcher and critic are separate roles under the hood, but the user hears one voice: the collaborator's single question (ADR 0002). The researcher's output arrives as sourced HITL notes or tensions. The critic's output arrives as typed gate suggestions with confidence (a candidate use for Jev). Impartiality comes from separating roles, with the critic from a different model family than the generator, not from adding voices.

4. A success end state — yes. When an idea is productised or used, it ends in a third terminal state, distinct from MATURE (still alive, undecided) and FOSSIL (path closed). The metaphor is it has already sprouted: the idea left the garden and lives in the world. Name: TRANSPLANTED. Still open:

   • what it stores: a link to the artefact (component, doc, repo, skill), plus what made it work, the success counterpart to the fossil reason;
   • whether it is immutable like a fossil (409 on PATCH and agent) or can still receive notes when the real-world artefact evolves;
   • whether only MATURE can be transplanted, or any living state;
   • whether the agent may propose it (HITL, like fossil proposals) or only the user sets it.

Also decided (2026-11-05), see Direction:

• Audience: a product for people with ideas, not the maintainer's private lab. Bounds stay loose.
• Framing: a creative harness that guides an idea to be discarded or watered as well as possible.
• Where validation happens: all of it — thinking and research, recorded experiments, produced artefacts — chosen by context.
• Providers: bring your own (Anthropic, OpenAI, Ollama local/cloud as the floor). Cursor was mentioned, but it is an editor, not a public model API. It fits better as a destination for transplanted artefacts (for example a skill or a repo) than as a provider.
• Capture: any input (image + text, link + instruction, a sentence). The root idea is extracted from it.
• Lifecycle: MATURE → MATURING (undergoing validation), plus terminal TRANSPLANTED.
• Validation plan: both implicit in the questions and visible/editable.

Keep open (answer with experiments):

5. Does a separate calibrated critic (Jev or another family) improve gates compared with the same LLM self-judging?
6. Does web research narrow ideas, as the evidence predicts? Measure path diversity with and without it.
7. Does variety from the personal corpus beat model variety (VARIETY with and without second-brain)?
8. Which classic methods earn a move, and which are theatre? Use the same test as stance: count move × state × outcome.
9. Does "lab" stay single-user, or do validations need other humans (interviews, reviewers)?

Explicitly not now: multi-user, a methods UI, autonomous agents that run without the human.

---

7. Suggested next experiments (smallest first)

1. Refutation research (one move, GERMINATING): the agent searches for prior art and counter-evidence; the result is a tension with sources, still one question.
2. Jev as stance selector + seed classifier: offline against evals/embryo-stance.jsonl; compare calibration with the current LLM.
3. Graduation: design spike only (state, artefact link). Write an ADR before code.
4. Path tournament in GROWING: pairwise debate, human as referee.

---

Sources were gathered 2026-11-05. Vendor figures (validators, Kosmos, Jev) are self-reported and not independently verified.

Related: Method as process · Saved references as contrast · Open Questions · Direction