AutoResearch/Innovation Generator/V6/Poincaré

VERSION / V6.8

Poincaré

Poincaré

Upgrades the default path to an observation-driven controller: HTTP model providers propose constrained candidates, the operator locks tasks and baselines, execution yields concrete counterevidence and revision obligations, and evidence authorizes representation changes.

Historical version · Design & implementation

Current: Daedalus v7.4.0

V6.8 / ARCHITECTURE

How does an exploration move forward?

01锁定任务与前沿Lock task and frontier02模型提案Model proposal03同资源实验Equal-resource execution04观察与修复Observation and repair05冻结并结束Freeze and end search06一次性确认One-shot confirmation01锁定任务与前沿Lock task and frontier02模型提案Model proposal03同资源实验Equal-resource execution04观察与修复Observation and repair05冻结并结束Freeze and end search06一次性确认One-shot confirmation
Architecture illustration derived from this version’s actual control flow.

Paths & conditions

04 → 02
Revision / authorized reframing
  1. 01

    Lock task and frontier

    Fixed metric, baselines and budget

  2. 02

    Model proposal

    Strict contract and valid IDs

  3. 03

    Equal-resource execution

    Candidate, baselines and raw outputs

  4. 04

    Observation and repair

    Counterexamples, ablations, memory

  5. 05

    Freeze and end search

    Eligible leading candidate

  6. 06

    One-shot confirmation

    Fresh holdout, no model feedback

01

Let the next generation see observations

V6.8 Poincaré moves the default system from template records toward real provider interfaces, controlled program execution and concrete observation feedback. Candidates respond to parents, worst counterexamples, baseline losses and repair contracts rather than merely rewriting prose. Poincaré+ is a directory/packaging name; README, pyproject and manifest identify version 6.8.0, not an additional generation.

02

Separate generation, experiments and evidence control

Models propose candidates, operators lock experiments and independent evidence control rechecks records. Models cannot change the primary metric, baseline sources, limits, ledger rules or confirmation samples. The inner loop revises from observations; the outer loop authorizes representation changes only after eligible counterevidence. Retitling is not reframing, and a code error does not automatically refute a theory.

03

One-shot confirmation after a dual loop

Opening a run snapshots configuration and task. Frontier retrieval, model proposals and reference checks precede candidate and baseline execution on fixed primary and extra suites. Raw outputs become observations; diagnosis, local ablations and difference memory inform the next generation. An eligible leader is frozen and search ends before a separate command opens fresh holdout data for one-shot confirmation without model feedback.

04

The host owns identities and evidence

Proposals require question, representation, assumptions, mechanism, at least two rivals, discriminating tests, predictions, code, ablations, prior differences and uncertainty. The host validates parent, observation and source IDs, fixes the metric and authorizes reframing. Observation preserves per-unit and baseline losses, counterexamples, protocol hash and raw-result references; RepairContract freezes task, baselines, limits and holdout policy.

05

Real protocols and explicit synthetic mode

Inputs include model configuration, operator-defined tasks, local literature or OpenAlex metadata, budgets and execution permissions. Providers implement Responses, Chat-compatible and Messages protocols without silent credential fallback. Executable tasks yield observations, archives, frozen candidates and confirmation states; absent adapters yield IDEAS_ONLY. Synthetic fixtures require explicit marking and demo conclusions.

06

Conceptual walkthrough: affine to quadratic representation

The bundled demonstration confronts a known affine program with synthetic polynomial counterexamples, uses logged counterevidence to permit a quadratic representation, applies equal-resource baselines and local ablations, then freezes and confirms. It shows observation feedback and confirmation separation. Both generator and world are hand-designed, so DEMO_CONFIRMED validates dataflow and rules, not a real LLM inventing an algorithm.

07

From typed ecology to observation-driven control

Relative to V6.7, V6.8 adds actual HTTP providers, candidate-observation loops, eligible-counterevidence triggers, local ablations, difference memory, resume checkpoints and one-shot confirmation. SQLite journals, content-addressed objects, HMAC, external anchors and the verifier continue evidence governance. Frontier and lineage ideas remain, with clearer permissions and dataflow prioritized over role counts.

08

Scope of the bundled validation record

The source VALIDATION.md reports 2,026 repository tests, 113 V6.8 tests and 17 program re-executions, plus loopback HTTP, synthetic dual-loop, build and clean-extraction checks. This review did not rerun them. The loopback server returns known fixtures and external paid-model calls were zero. Docker configuration was delivered but not executed, so neither real-model research nor production isolation is validated by that record.

09

Task improvement and discovery stay separate

Confirmation can report task improvement, demo, negative or inconclusive states, while discovery credit remains zero and global novelty uncertified. Holdout independence is operator-attested rather than proven. Feedback references do not establish causal use, local ablations do not prove an entire mechanism, and active experiment choice is an engineering proxy. Reports preserve these distinctions for future real-task evaluation.

Reading example · Explaining the workflow

How observations constrain the next generation

Inputs

A hand-designed synthetic polynomial task and affine candidate.

  1. 01

    Fix metric, baselines, data interface and resource limits.

  2. 02

    Execute the program and retain per-unit losses and eligible counterexamples.

  3. 03

    Authorize quadratic reframing only after eligible evidence.

  4. 04

    Apply fixed evaluations and local ablations to the revision.

  5. 05

    Freeze before one-shot fresh-sample confirmation; never resume search.

Output

A demo confirmation record and re-executable program evidence.

The source demonstration is synthetic and hand-written; it does not prove real-model innovation or global novelty.

Changes in this version

Generation

Earlier approach

V6.7 local compositional islands

This version’s approach

HTTP providers and strict proposals

Provide a real model interface

Feedback

Earlier approach

Typed lineage and unsettled archives

This version’s approach

Measured observations, diagnosis and counterevidence permissions

Make revisions answer concrete evidence

Confirmation

Earlier approach

Event and world-settlement verification

This version’s approach

Separate one-shot confirmation after freezing

Separate exploration from confirmation feedback