Date: 2026-08-28
Repository: MSKazemi/idkmesh
After merging the Mathematical Evolution Kernel and proving its Bayesian checkpoint persistence across two real main iterations, the next useful step was to apply the mathematics to live repository work rather than add more isolated formulas.
The owner asked for a more solid, smarter mathematical/algorithmic foundation implemented through GitHub Actions and GitHub-native capabilities.
The repository already had reusable implementations for:
The missing operational question was:
How should the live open issue/PR portfolio be observed with these algorithms without granting the observer repository authority?
Added a read-only Repository Mathematical Portfolio layer.
The Action uses GitHub CLI under read-only token permissions to snapshot up to 200 open issues and 200 open pull requests, including public metadata such as title/body/labels/comments/author/timestamps.
The normalized snapshot is retained with the resulting report so every ranking can be replayed against the exact observed repository state.
Open work is deterministically classified into:
The vocabulary is versioned in state/repository-portfolio-policy.json rather than hidden in workflow code.
The parser creates graph edges only for explicit phrases:
blocked by #N;depends on #N;requires #N;blocks #N.A generic #N mention does not create a dependency.
This rule is regression-tested to prevent the system from inventing graph structure from ordinary cross-references.
Each issue/PR receives transparent bounded proxy features:
They are fed to non-dominated Pareto sorting and NSGA-II crowding. A small scalar opportunity score exists only as a deterministic explanatory/tie-break diagnostic, not as the primary optimizer.
Open issues use the canonical mathematical kernel’s discounted directed unlock value. Only explicit dependency edges contribute.
The observer computes strategy entropy and Jensen-Shannon divergence between the live open-issue portfolio and the previous attention mixture.
The latest trusted Bayesian evolution checkpoint is downloaded on trusted main runs.
Distance from homeostatic targets becomes a strategy attention need, which updates strategy mixture weights with multiplicative weights and an exploration floor.
This is explicitly not described as causal reward.
The portfolio state counts how often each strategy has received exploration focus. Current live opportunity plus historical attention count drives a UCB exploration choice.
An unseen strategy receives exploration priority, preventing the repository from permanently focusing only on historically dominant work classes.
Again, UCB selects a recommended exploration focus, not an issue mutation or merge action.
The workflow uses the proven artifact-checkpoint pattern:
It separately restores the latest trusted Bayesian evolution checkpoint for health signals.
The workflow requests only:
contents: read;issues: read;pull-requests: read;actions: read.It cannot label, assign, comment, approve, close, merge, or write repository contents.
This is intentional while main remains externally unprotected under canonical issue #35.
The most important output is not the current ranking itself. It is a replayable longitudinal dataset:
repository snapshot
+ policy version
+ Bayesian health state
+ dependency graph
+ feature vectors
+ Pareto fronts
+ diversity state
+ UCB/attention state
That dataset can later be joined to delayed real outcomes and used to calibrate the proxy model empirically.
The long-term direction remains:
transparent hypotheses
-> replayable observations
-> delayed outcomes
-> predictive calibration
-> bounded experimental policy updates
-> independent governance