Decision intelligence map

The intelligence package is a collection of explicit decision stages. It keeps ranking mechanics, scientific interpretation, skeptical challenge, policy judgment, and learning separate so a change in one can be reviewed without silently changing the others.

flowchart LR
    question["decision question"]
    candidates["candidate universe\nincluding exclusions"]
    interpretation["bounded scientific interpretation"]
    challenge["support · contradiction · falsifier"]
    policy["named scenario and policy"]
    disposition["recommend · hold · downgrade · refuse"]
    review["human or explicitly promoted authority"]
    question --> candidates --> interpretation --> challenge --> policy --> disposition --> review

The package owns the recorded argument between question and disposition. It does not own the source evidence, execute proteomics analyses, or grant the experimental authority shown at the right edge.

Candidate construction

candidates defines candidate records and schemas, validation, transformations, filters, quality signals, fingerprints, lifecycle state, persistent stores, ranking, and final selection. A ranking is meaningful only relative to the candidate set it considered; excluded and invalid candidates remain part of the audit trail.

Interpretation

interpretation provides bounded readings of runs, quantitative results, contrasts, pathways, PTMs, contaminants, and structures. These modules translate scientific outputs into features and statements suitable for decisions. They do not replace core calculations or knowledge grounding.

Skeptical pressure

claims evaluates support, contradictions finds incompatible assertions, and falsifiers formulates evidence that could overturn a position. posture records whether the assembled evidence is strong, weak, conflicting, or otherwise unsuitable for a decisive recommendation.

flowchart TD
    interpretation["bounded interpretation"]
    support["claim support"]
    contradiction["contradictions"]
    falsifier["falsifiers"]
    posture["evidence posture"]
    interpretation --> support --> posture
    interpretation --> contradiction --> posture
    interpretation --> falsifier --> posture

Judgment

judgment contains the policy-bearing surface: scenarios, recommendation rules, benchmark policies, confidence, decision packets, blinded challenges, counterfactuals, sensitivity, quality, and regret. The output is a recommendation record with enough context to explain why another policy or scenario could produce a different answer.

reviews exposes that reasoning for scrutiny through workflow-specific benchmark reviews, boards, candidate reviews, independent-rerun checks, decision briefs, external review kits, outsider packets, release candidates, and public-scrutiny reports.

Inspect A Recommendation By Question

Question	Owning surface	Evidence to inspect
Which alternatives entered or left the comparison?	candidates	input fingerprint, validation, exclusions, transformations, lifecycle state
Which scientific meaning was extracted from the inputs?	interpretation	bounded feature or statement plus the source record it projects
What could defeat the preferred answer?	claims, contradictions, falsifiers, posture	adverse evidence, severity, missing support, falsification condition
Why did one action outrank another?	judgment	scenario, policy identity, weights, gates, tie handling, decisive criteria
Would a plausible assumption change the answer?	counterfactual, sensitivity, and regret surfaces	alternate policy result, rank stability, observed cost
Who may act on the result?	decision envelope and external authority	advisory/enforced state, promoting actor, rationale, human-review requirement

Learning without rewriting history

learning adapts future policy from review and outcome signals. Refinement tracks convergence and stagnation explicitly. Historical evidence, recommendations, and outcomes remain immutable inputs to the later learning record; learning must not retroactively make an earlier decision appear better calibrated than it was.

Public surface

The root package exposes module families rather than a broad collection of functions:

from bijux_proteomics_intelligence import (
    candidates,
    interpretation,
    judgment,
    posture,
    refusal,
    reviews,
)

This is a library package without a standalone CLI or HTTP service. Runtime can expose intelligence-backed routes while retaining ownership of transport and execution behavior.

Non-goals

Intelligence does not curate literature, resolve biological identifiers, run mass-spectrometry workflows, or schedule assays. It consumes those owned artifacts and returns a policy decision. When evidence, reproducibility, or lab feasibility is insufficient, the correct result is a downgrade or refusal.

A review is incomplete if it reports the selected candidate without the candidate universe, or reports confidence without the policy, adverse evidence, sensitivity, and authority state that produced it.