AutoResearch/Innovation Generator/V3/Lens-First Active Generation

VERSION / V3.x / V3.4

Lens-First Active Generation

Choose the contribution before generating the idea

This stage organizes research ideation around twelve contribution types, including methods, theory, perspectives and measurement. It fixes the subfield and essential prior work, combines first-principles kernels with transferable anchors, and develops candidates with a clear difference, risk and decisive test.

Historical version · Design & implementation

Current: Daedalus v7.4.0

V3.x / V3.4 / ARCHITECTURE

How does an exploration move forward?

01承诺贡献透镜Commit to a contributionlens02三杠杆探索Explore through threelevers03机制与组合检查Check mechanism versuscombination04研究问题契约Specify the researchquestion05按透镜审查Audit by lens06交付人工审核Hand off for humanreview01承诺贡献透镜Commit to a contribution lens02三杠杆探索Explore through three levers03机制与组合检查Check mechanism versuscombination04研究问题契约Specify the research question05按透镜审查Audit by lens06交付人工审核Hand off for human review
Architecture illustration derived from this version’s actual control flow.

Paths & conditions

05 → 01
Reframe at saturation
  1. 01

    Commit to a contribution lens

    Framing and baseline classes

  2. 02

    Explore through three levers

    Kernels, anchors and literature

  3. 03

    Check mechanism versus combination

    Coupling value and established moves

  4. 04

    Specify the research question

    Inputs, mechanism, outputs and tests

  5. 05

    Audit by lens

    Evidence, weaknesses and reframing

  6. 06

    Hand off for human review

    Recommendation, alternatives and risks

01

Start with the contribution identity

V3 starts by asking what the work would contribute: a method, a theoretical result, a measurement resource or a correction to existing understanding. Twelve contribution lenses organize the generation route, prior-work comparison and evidence requirements around that answer.

02

Input and framing contract

The input is a research direction with optional literature. Broad topics are narrowed before committing to a primary lens, up to two alternatives, confidence and reasons to exclude other framings. Subfield resolution enumerates relevant baseline classes, while claimed, actual, perceived and best-framed contributions are checked for alignment.

03

Kernels, anchors and literature refutation

First-principles kernels supply mathematical, physical or logical structure. Classical anchors supply transferable innovation shapes. Literature refutation asks whether prior work already contains the key move. After these levers converge, component decomposition, coupling value and contemporary prior art determine whether a candidate is discarded, retained or refined.

04

Turn a sentence into a research question

The refinement template spells out inputs, mechanism, output form, a measurable quantity, why the work is timely and who benefits. A cheap decisive test and explicit failure criterion follow, along with residual knowledge if the hypothesis fails. These fields make the proposal discussable beyond its title.

05

Successive V3.1–V3.4 refinements

One archive contains four patch histories: V3.1 calibrates self-evaluation and hybrid baselines; V3.2 checks theorem assumptions and responses to weaknesses and permits lens switching; V3.3 constrains universal and causal claims; V3.4 sharpens measurement criteria, separates promises from evidence and tracks diminishing returns. These are embedded histories, not four independently supplied releases.

06

Deliverables and human judgment

The handoff contains a recommended direction and alternatives with strongest-prior comparisons, risks, self-evaluation and staged validation plans. State resides primarily in the model-facing Markdown protocol and research notes; the later persistent Python search tree is absent. Human checkpoints can revise the scope, deepen a lever or stop an unpromising direction.

07

From active exploration to lens-first generation

Relative to the V2 design described in the archive, V3 moves framing ahead of generation and changes routes and evaluation axes by lens. Evaluation criteria become design conditions: measurement needs measurement evidence, theory needs valid assumptions, and a perspective needs a reason to alter interpretation. Every candidate need not become a new architecture.

08

How to read a V3 proposal

A proposal can be read along three lines. Identity asks whether its title matches its actual contribution. Comparison asks whether the strongest prior solves the same problem in the same subfield. Evidence asks whether the decisive test can overturn the hypothesis rather than merely show a success example. This keeps the lens connected to research judgment instead of turning it into a title label.

A measurement proposal discusses construct validity, coverage, baselines, reliability, diagnostic value and community utility separately. Planned inter-laboratory replication differs from replication data; promising code differs from an available tool. V3.4 embeds those distinctions in self-evaluation so future work remains a plan and current judgments rest on current material.

09

Archive and public-source relationship

The two older ZIPs are byte-identical and form one V3.x archive containing V3.4 patches. The public repository whose name includes v3 is the assessment system, with a README already describing later internal versions; the public generator repository contains an earlier three-lever protocol. They can be linked as related tools without claiming that the local V3.4 snapshot itself was published.

Protocol revisions within the archive

Four protocol revisions retained in the same V3.x archive.

V3.1

Calibrate self-evaluation and added complexity

V3.1 responds to generation–evaluation feedback on long-tail knowledge recall: self-evaluation was high, and added components introduced stability, comparison and reproducibility risks. The patch separates complexity from originality and reassesses the actual contribution instead of continuing to raise the previous self-score.

Rules add originality tiers, cross-class hybrid baselines, three added-weakness questions, source-domain coverage deduplication and stability evidence for gradient-sharing components. Beyond a proposal, the input must describe new parameters, resource overhead, component interactions and baseline combinations. Risks enter Killer Risk; empty answers return to refinement.

These requirements enter self-checks and internal acceptance, letting evaluation feedback shape regeneration. The document limits the evidence mainly to one direction and the method lens; further error convergence remains a target awaiting data. This is an embedded V3.x protocol revision, not evidence of calibration across all fields or an independent release.

V3.2

Audit theorem conditions and permit reframing

V3.2 addresses formal compliance left by the previous rules: a stability theorem can be cited without checking its assumptions, and a weakness response can look complete without a concrete test. Repeated refinement within one method framing can also reach structural limits, making another kernel change an unhelpful next step.

The theorem audit lists assumptions, checks applicability, falls back to empirical monitoring or analogous priors and states dependencies. Weakness answers require quantified overhead, a falsifiable conditional and a test path. Framing saturation, dominant baselines or repeated low-originality tiers trigger lens reselection and a restart after human confirmation.

Control shifts from answered questions to testable answers and from endless component additions to reconsidering contribution identity. Evidence still mainly concerns the method lens and long-tail recall; projected error levels and cross-lens calibration remain unverified. This records a protocol change within the archive, without adding a separate historical release.

V3.3

Set evidence criteria for empirical correction

V3.3 follows a shift from method contribution to empirical correction. Earlier calibration mainly served methods, while the new framing requires belief update, experimental design, confound control and reproducibility. Reusing the old criteria can inflate confidence. Lens switching must change evaluation criteria, not merely the title.

New rules add evidence tiers for six empirical-correction axes, audit universal wording and require identifiability, intervals and sensitivity assumptions for causal decompositions or attribution proportions. Insufficient evidence requires narrower claims or additional tests. A local counterexample is not a universal law, and numerical precision does not establish a decomposition.

These checks enter pre-handoff validation and align claims with available material. The document acknowledges that complete anchoring still covers only some lenses; subsequent calibration and expansion are plans. This account preserves that scope without equating internal review gains with experimental validation of a corrective conclusion or inventing a release date.

V3.4

Separate measurement evidence from promises

V3.4 turns to measurement and benchmark research, refining its anchors through proposal feedback. It asks whether measures match constructs, coverage spans domains, baselines are adequate, reliability has data, errors are diagnosable and tools have community utility. Sub-suite counts and future replication plans cannot substitute for those answers.

Rules tier evidence across six measurement axes, limit self-evaluation based only on promised replication, leaderboards, audits or releases, and inspect unsupported thresholds and unspecified aggregation weights. Diminishing refinement gains or structural remaining weaknesses can trigger reframing. Inputs must distinguish available evidence, planned validation and future maintenance.

These are protocol revisions entering self-checks and control decisions, not a conclusion that a measurement tool was successfully deployed. The document lists incompletely covered lenses and contains inconsistent score-gap notation, so convergence percentages are not used as effectiveness claims. V3.4 and the preceding revisions share one archive; later version concepts remain plans.

Reading example · Explaining the workflow

Reading a measurement contribution

Inputs

Suppose a researcher wants to compare evidence use in language-model reasoning, rather than invent an architecture.

  1. 01

    Choose the measurement lens and specify the measured construct, intended users and baseline classes.

  2. 02

    Use kernels and anchors to draft a measurement mechanism and search for priors defining the same quantity.

  3. 03

    Apply V3.4 to separate validation plans from available data and state a test that could falsify the measure’s value.

Output

Deliver a discussable measurement proposal with unresolved literature and evidence questions visible.

A conceptual reading example. No generation, retrieval, review or experiment was run, and no performance result is reported.

Changes in this version

Generation order

Earlier approach

Explore before inferring the contribution type.

This version’s approach

Commit to a lens before generation and reframe at saturation.

Let the idea and evidence criteria develop together.