Strategy learning per pass

Every strategy is scored per objective: observed change minus the change its doctrine expected. The expectation itself learns from completed uses -- "learned" is what it actually does, "authored" is only where it started. A residual is the gap between the two.

Strategy interactions per pass

Pairs of strategies whose five-pass windows overlapped: what the pair achieved together against the sum of both strategies' own expectations.

Plain-action baselines per pass

What each ordinary action, with no strategy chosen, has achieved in each active situation's objectives -- the yardstick a strategy's learned expectation is really being compared against.

Stall recognition per pass

A stall is a gap since the last objective improvement longer than 90% of the gaps play has recorded. The threshold is learned, not set.

Advice standing per session

Every monster tactic and doctrine line Jev is shown carries its record inline, as "[shown N times: X died, Y survived]" -- plain counts, never a weight or ranking. Loaded once at session start; only changes once a run ends. Text is recovered from the current knowledge base and matched by the same hash the payload itself uses, so a record whose advice has since been edited or removed shows its counts with no text rather than a guess.

Learned lessons per session

A death pattern (same monster or hazard, at least twice) becomes a sentence Jev is actually shown. Below it, the raw deaths it was built from.

Recent mistakes (source material for the lessons above)

Bandit posteriors per session

Which value of each tunable threshold (e.g. how low HP has to fall before a flee rule's advice fires) is sampled at session start, weighted by which values have produced better-scoring runs so far.

Promoted interpretations per pass

Pilot (2026-09-23): one hand-written derivation rule, proving the full lifecycle -- derive, corroborate, promote, contest -- end to end. Corroboration is a for/against count pair, never a single confidence number. Once promoted and matching the active situation, the claim reaches Jev's actual payload, not just this dashboard.

Coverage trend

Of the item/monster/hazard kinds actually encountered in a run, what fraction match an entry in the hand-authored knowledge base. Not a measure of how much of the game is known -- the denominator is only what showed up that run, so a quiet run and a catalog gap can both read 100%.

Reward-shaping weights

Learning measurement

Three separate claims -- a growing store is not the same as better play. Read together, not as one score.

Accumulation

Reachability

Contributions by goal

Outcome improvement

Knowledge ledger static

Not learned. This is the hand-authored reference catalog itself and how much of it has been transcribed and verified -- a documentation audit, not a report on play.

Knowledge discovered

First-time-seen items, monsters, and hazards -- a log, not itself read back into any decision.

Run history

Past training generations