Context Algorithm Portfolio

Tracked by pm-g8t7b9, pm-4k6b, and pm-vvib.

This is a navigation index of the historical algorithm ledger and cognitive proposals. Individual pm items own scope, decisions, experiments, and current evidence. A shared substrate or an earlier defect fix does not establish that a proposed ranking mechanism has been implemented. Read each owner live before claiming its state.

Bounded discovery

pm list --all --tags algorithm-proposal --limit 10
pm search "<stable-slug>" --limit 5
pm get <owner-id> --full
pm history <owner-id> --full

Use the result cursor and omission receipt before making a complete-catalog claim. The four agent skills route to generated capability references; this portfolio does not add a new command family or a second execution backlog.

Algorithm ledger

The 76 mechanisms below retain their original note timestamps and numeric aliases. An ordinal is ambiguous without its source note; new proposals use stable slugs. Where two descriptions express one experiment, both slugs resolve to its owner. Existing acceptance owners retain their original delivery boundary.

Cognitive subsystem mapping

The 18 candidate compositions from pm-vvib retain the architecture decision gate. Individual experiment drafts record the ideas without accepting that decision.

Shared experiment contract

Each proposal names its source, bounded SDK contract, inputs, evaluation population, costs, deterministic fallback, and promotion or rejection rule. Compare against an unchanged baseline on held-out judgments. Report task success, critical-evidence omission, ranking quality, tokens, latency, memory, and rebuild cost separately.

Use tiny, real-shaped, sparse/dense, returning-session, concurrent, and long-history fixtures. A smoke run cannot certify million-item performance or fleet convergence. Read and write through the public SDK; CLI and MCP remain adapters.

Historical descriptions may contain hypotheses about optimality, independence, or constant-time behavior. Those are claims to test. State the assumptions required by an approximation guarantee, and preserve missing evidence as unknown.

Promotion requires an explicit decision and passing current mandatory gates. A proposed experiment does not authorize a production policy, adaptive learning, or workspace-data export. Rejected experiments remain searchable evidence.

Mathematical checks before promotion

• RRF's published offset of 60 was an empirical choice. Retain a deterministic baseline and test alternatives; it is not a universal calibration constant. Cormack, Clarke, and Büttcher (2009).
• A plain value-per-token greedy heuristic does not inherit every submodular approximation guarantee. Sviridenko's knapsack result assumes a nonnegative, nondecreasing submodular objective and uses partial enumeration with greedy completion. Representation exclusivity and negative penalties need their own analysis. Sviridenko (2004).
• WL subtree features scale with edges and refinement rounds in the published formulation. Our proposed use still needs bounded neighborhoods and an evaluation of whether structural similarity predicts useful project evidence. Shervashidze et al. (2011).