Reasoning & memory·Memory
Failure-aware optimization
Most discovery systems over-index on positive examples. Protean keeps rejection signals, degradation patterns, motif conflicts, and weak-candidate behaviour inside the optimisation loop.
Failure memory feeds the failure-similarity scoring signal at rank time, and claim-QA flags reach reviewers as warnings. The back-edge from contradiction memory into proposal generation is reserved in the current runtime.
What failure memory influences
Failure memory affects:
- candidate rejection at the constraint surface
- ranking penalties (via the
failure_similarityscoring component, baseline weight 0.10) - constraint refinement across cycles
- review warnings surfaced to the reviewer
- bounded feedback signals when trusted preconditions exist
- research prioritisation in the cognition layer
Signal classes
Failure-aware optimisation incorporates several classes of negative signal:
- sequence motifs associated with cleavage or degradation against the four-enzyme panel (trypsin, chymotrypsin, pepsin, elastase)
- candidate records rejected by deterministic validation
- literature language describing instability, poor permeability, short half-life, weak solubility, or delivery limitations
- clinical or translational context that suggests failure modes
- internal review notes and assay feedback when available
The runtime distinguishes between a failure observation, a warning, and a hard rejection. Negative evidence can be useful without becoming an automatic veto.
Ranking influence
Failure memory affects ranking through penalties, warnings, similarity context, and bounded feedback signals. It does not erase candidate review. A candidate near failure memory may still be scientifically interesting — but it should advance with visible context and a higher review burden.
The contradiction graph
Claim-QA flags are stored alongside failure memory as a contradiction record: which generated statement was flagged, against what evidence, with what confidence. These records are content-addressed and surface on the Protean Ledger as Hypothesis and Experiment records joined by Contradicts lineage edges — first-class typed relations a reader can walk from chain state alone.
Today, the contradiction graph is part of the cycle's terminal RuntimeCycle record on chain, and any individual contradiction can be retraced through the Contradicts edges hanging off the records it concerns. The back-edge from contradiction memory into proposal generation or hypothesis prioritisation is reserved — when that loop closes, it will be a single review-gated change visible inside the cycle executor. See Provenance layer for the verification recipe.
Why this matters
A system that forgets failures repeats them. A system that learns from failures develops sharper search boundaries and better review discipline. Failure data accumulates slowly and becomes more valuable as it is normalised — what failed, why it failed, where it appeared, and how that signal should influence the next cycle.