bijux-proteomics-intelligence

bijux-proteomics-intelligence turns scientific evidence and program constraints into inspectable decisions. It owns candidate ranking, scenario analysis, skeptical challenge, recommendation posture, and refusal. It does not own the evidence it evaluates.

python -m pip install bijux-proteomics-intelligence

Diagnose recommendation movement

Two recommendation records can disagree without either being corrupt. Compare the dimension that moved before interpreting the decision change.

Changed dimension	Record to inspect	Defensible interpretation
evidence snapshot	Knowledge bundle identities, source versions, contradictions, and sufficiency decision	the decision responded to different scientific support; do not describe this as policy learning
candidate universe	included, excluded, invalid, and unavailable candidate identities	the winner changed inside a different choice set; rankings are not directly comparable until the shared set is isolated
decision policy	objectives, hard constraints, thresholds, weights, tie-breakers, and policy version	the value judgment changed; evidence did not become stronger merely because the ranking moved
scenario burden	counterfactuals, perturbations, sensitivity ranges, and falsifiers	the recommendation faced a different challenge envelope
feasibility or lab burden	capacity, assay readiness, cost, safety, and consequence assumptions	an analytically attractive action became operationally weaker or stronger
authority state	human review, escalation owner, approval, hold, or refusal	the permission to act changed; the scientific inputs may be identical

Record multiple causes separately when they move together. A later recommendation supersedes a prior decision only for its declared evidence, candidate, policy, scenario, and authority context; it does not rewrite why the earlier record was produced.

Decision pipeline

flowchart LR
    evidence["versioned evidence bundle"]
    snapshot["evidence fingerprint"]
    candidates["candidate set\nfilter · validate · fingerprint"]
    policy["declared policy\nconstraints · objectives · thresholds"]
    burden["feasibility and lab burden"]
    rank["ranking\npolicy · quality · selection"]
    challenge["challenge\ncontradictions · falsifiers · scenarios"]
    judgment["judgment\nsensitivity · confidence · regret"]
    result{"supported?"}
    recommend["recommendation record"]
    refuse["refusal with reasons"]
    review["human and domain review"]
    evidence --> snapshot --> challenge
    candidates --> rank
    policy --> rank
    rank --> challenge
    burden --> challenge
    challenge --> judgment --> result
    result -->|yes| recommend
    result -->|no| refuse
    recommend --> review
    refuse --> review

Every recommendation is a policy output, not a fact. Its candidate set, policy, evidence posture, sensitivity, alternatives, and reason codes remain recoverable so a reviewer can reproduce the judgment without treating the result as new evidence.

Recommendation record anatomy

Field	Why it matters
candidate universe and exclusions	prevents a winning candidate from being presented without the alternatives it defeated
evidence references and fingerprints	ties the decision to immutable review inputs without copying or rewriting them
policy and constraints	exposes the values, thresholds, feasibility limits, and objectives that shaped the ranking
score components and ordering	makes aggregation and tie-breaking inspectable
contradictions and falsifiers	records evidence that weakens or could overturn the action
scenarios and sensitivity	shows whether small plausible changes reverse the ranking
confidence and regret	separates certainty language from the estimated cost of being wrong
posture and human-review flag	distinguishes advisory output, downgrade, escalation, and refusal

Analytical capabilities

Surface	Responsibility
candidates	typed records, validation, filtering, quality, fingerprints, ranking, selection, storage, and lifecycle
interpretation	quantitative, contrast, pathway, PTM, contaminant, structure, and run-level readings
claims, contradictions, falsifiers	support checks and skeptical pressure against an interpretation
judgment	policies, scenarios, recommendations, blinded challenges, counterfactuals, sensitivity, confidence, regret, and flagship decisions
posture	explicit evidence posture and skeptical review
reviews	benchmark reviews, review boards, decision briefs, outsider packets, independent reruns, and public scrutiny
learning	adaptation, refinement convergence, and stagnation detection
next_steps, query, refusal	action handoff, interrogation, and unsupported-claim refusal

The package root lazily exposes these fourteen owner modules, keeping import cost and accidental coupling low while making the supported capability families discoverable.

What makes a recommendation defensible

A recommendation is strongest when:

1. the candidate set and exclusions are explicit;
2. the ranking policy and input evidence are fingerprinted;
3. plausible contradictory evidence and falsifiers were evaluated;
4. ranking stability survives threshold and scenario sensitivity;
5. competing actions and the cost of error are visible;
6. confidence is calibrated against benchmark and observed outcome evidence;
7. a refusal remains possible when the support is inadequate.

Benchmark review modules cover DDA, DIA, LFQ, multiplex, PTM, and targeted workflow families. LFQ and multiplex review both live in the quantification surface, but they do not share an evidence posture: LFQ is outsider-auditable bounded while multiplex remains internal support only. Module existence does not grant authority; the recommendation inherits the weakest relevant ceiling in its input evidence chain.

Decision output states

Intelligence returns a disposition with an explicit next authority. A score or rank is incomplete until this state is recorded.

Output state	Meaning	Required next authority
recommend	one action remains defensible under the declared evidence, policy, and challenge set	human and domain review before commitment
recommend with conditions	the action survives only inside named thresholds, prerequisites, or stop conditions	owner verifies every condition before handoff
downgrade	evidence or policy sensitivity does not support the requested confidence or consequence	reviewer narrows wording, scope, or action
escalate	the decision depends on unresolved expertise, safety, cost, or contradiction	named domain or operational owner decides
hold	additional evidence can resolve a material gap and waiting is an admissible action	evidence owner supplies the named record or closes the route
refuse	no candidate satisfies the hard constraints or support burden	requester changes the question, evidence, or constraints

These states are not confidence synonyms. A high-scoring candidate can still be held, escalated, or refused when the action exceeds the available authority.

Decision stability

Observed behavior	Interpretation	Required posture
ordering survives plausible thresholds and evidence removal	locally stable under tested pressure	bounded recommendation with tested conditions
top candidates exchange rank under small changes	policy-sensitive	expose alternatives and require review
recommendation depends on one contested source	evidence-fragile	downgrade until contradiction is resolved
feasible action changes when assay burden is included	consequence-sensitive	return cost and burden to the decision record
no candidate satisfies hard constraints	unsupported action	refuse with unmet conditions

Stability applies only to the tested candidate universe, evidence snapshot, policy, and scenario set. It does not imply that an omitted candidate or future evidence could not change the result.

Challenge before action

flowchart TD
    R["ranked candidates"] --> B["blinded evidence challenge"]
    B --> C["counterfactual scenarios"]
    C --> S["threshold sensitivity"]
    S --> G["regret analysis"]
    G --> D{"ranking remains defensible?"}
    D -->|yes| P["bounded recommendation"]
    D -->|weakens| W["downgrade or escalate"]
    D -->|no| F["refuse"]

A recommendation that changes under a plausible threshold, withheld evidence pattern, or feasible alternative must expose that instability. Explanation is not a substitute for challenge; it reports how the declared policy behaved under challenge.

Ownership boundary

• Core owns scientific calculations and benchmark contracts.
• Runtime owns what executed and whether it can be replayed.
• Knowledge owns sources, claims, provenance, and contradiction state.
• Intelligence owns how reviewed inputs become a ranked or refused action.
• Lab owns whether that action is feasible and what happened after execution.

Intelligence may consume all of those signals, but it must not rewrite them. Outcome-aware learning creates a new policy or calibration record rather than editing the historical recommendation.

Compare decisions without erasing history

When a recommendation changes, compare immutable decision records. Do not rewrite the earlier record to match the current evidence or policy.

Comparison dimension	Meaning of a difference	Required interpretation
candidate universe	candidates were added, removed, or newly excluded	isolate selection effects before comparing scores
evidence fingerprint	the support or contradiction snapshot changed	attribute the change to evidence custody in Knowledge
policy fingerprint	weights, thresholds, constraints, or objectives changed	report a policy change, not scientific discovery
scenario set	the tested uncertainty envelope changed	compare only shared scenarios or label the new burden
feasibility input	cost, capacity, safety, or assay burden changed	distinguish operational consequence from evidence strength
ranking and posture	ordering, confidence, escalation, or refusal changed	state which upstream difference caused the disposition change

flowchart LR
    O["earlier decision record"] --> CP["compare fingerprints and inputs"]
    N["current decision record"] --> CP
    CP --> EC{"what changed?"}
    EC -->|evidence| ER["evidence-attributed explanation"]
    EC -->|policy| PR["policy-attributed explanation"]
    EC -->|feasibility| FR["consequence-attributed explanation"]
    EC -->|multiple| MR["separate effects and disclose ambiguity"]

If the responsible difference cannot be isolated, the correct output is an attribution gap. A plausible narrative is not a substitute for matching record identities.

Human authority boundary

Intelligence can rank, challenge, downgrade, or refuse an action. It does not approve clinical use, spend resources, authorize an executable laboratory handoff, or override biosafety and operational custody. A human_review flag is a required decision state, not a claim that human review has occurred.

flowchart LR
    R["recommendation record"] --> H{"human and domain review"}
    H -->|revise| I["new policy or evidence input"]
    H -->|reject| X["closed or refused action"]
    H -->|accept advisory| L["Lab readiness assessment"]
    L -->|not ready| N["revised or refused plan"]
    L -->|ready and authorized| E["executable handoff"]

Challenge A Recommendation

Review a recommendation as an argument that can fail. Begin with the action, then recover the alternatives, evidence snapshot, policy, challenge burden, and authority boundary that produced its posture.

Challenge	Evidence to demand	Honest response when missing
was the winner selected from a complete declared universe?	candidate ledger and exclusion reasons	disclose selection uncertainty or refuse comparative language
would another reasonable policy change the ordering?	policy fingerprint, component scores, tie-breaks, and alternative policy result	label the decision policy-sensitive
does one contested source control the outcome?	source attribution, leave-one-source-out result, and contradiction state	downgrade until the dependency is resolved
does a plausible threshold reverse the action?	sensitivity surface and rank crossings	expose alternatives and require review
is the cost of error acceptable?	regret estimate, falsifiers, and stop conditions	narrow the action or refuse
is the action feasible and authorized?	Lab readiness and human decision	keep the output advisory

flowchart TD
    recommendation["recommendation"] --> universe["recover candidates and exclusions"]
    universe --> evidence["resolve evidence snapshot"]
    evidence --> policy["replay policy and ordering"]
    policy --> pressure["apply contradictions · sensitivity · regret"]
    pressure --> posture{"posture survives?"}
    posture -->|yes| advisory["bounded advisory action"]
    posture -->|weak| review["downgrade or human review"]
    posture -->|no| refuse["refusal with unmet conditions"]

Use Workflow Consequence Maps for family-specific effects, What Changed The Recommendation to compare decisions, and Lab Consequence for operational authority.

Continue By Decision Question

Need	Read next	Review is complete when
understand module ownership and artifact flow	package overview	candidates, evidence, policy, challenge, disposition, and next authority resolve separately
pressure-test a ranking	recommendation challenges	leave-one-source-out, counterfactual, threshold, and burden pressure expose every rank reversal
interpret confidence, calibration, or regret	recommendation confidence	confidence language matches the tested evidence snapshot and cost-of-error record
inspect policy and dependency boundaries	architecture	Intelligence consumes owner records without rewriting scientific, execution, evidence, or laboratory truth
choose Python, data, or artifact contracts	interfaces	the recommendation can be reproduced from identified candidates, evidence, policy, and scenarios
determine when Intelligence must downgrade or refuse	known limitations	every unsupported action ends in a named downgrade, escalation, hold, or refusal