Objective lens — the root object, switchable
The same telemetry under a different objective yields different decisions. The default view evaluates each application under its own declared objective; a lens re-evaluates every application under one. Ranking always recomputes; a move is only rejected where a constraint can bind on your data — so the lens states which of its constraints can, and which cannot.
Recommendations
17 generated optimization hypotheses, ranked by business value under each application’s objective. Nothing here is seeded — the engine produces them from telemetry.
1 of 17 shown · filters active · Clear filters
Outstanding — ranked by value × confidence × objective alignment
Route small-context Python Quant traffic off claude-sonnet to GLM-5.2
Also considered — overlapping alternatives (not double-counted) · 2 of the generated hypotheses above · €82,458/yr already counted in the leads, not additional
0 matching
Held — surfaced, not led · 7 of the generated hypotheses above · €18,883/yr released if you declare these surfaces low-stakes
These fired, but Bayeto won’t lead with them: either the value is trivial next to the bill, or they carry an unmeasured quality risk on a surface you’ve declared critical. The reason is shown on each. Where a move competes with one already leading the same spend, the figure above is what declaring would gain by swapping — never the two added together.
0 matching
Considered & not surfaced · 19 examined and never generated — not in the count above · €0/yr if the telemetry had supported them
Candidates the engine examined and declined — it shows its work on the moves it didn’t make, not only the ones it did.
Batch the offline Client Reporting jobs
would save €0/yrThe pinned catalog carries no batch tariff for Client Reporting's models — every schedule in it is synchronous, on-demand. Batch pricing is published per provider; Bayeto has not captured it, so there is no discount to quote and this move is not priced.
Cache economics for Client Reporting (claude-haiku)
would save €0/yrExamined Client Reporting's claude-haiku cache: reuse 12.4× vs break-even 0.28× — the cache nets in your favor. Healthy; nothing to change.
Cache economics for Client Reporting (claude-sonnet)
would save €0/yrExamined Client Reporting's claude-sonnet cache: reuse 12.5× vs break-even 0.28× — the cache nets in your favor. Healthy; nothing to change.
Response length for Client Reporting (claude-sonnet)
would save €0/yrExamined: output is 55.7% of this segment's cost, but responses average only 798 tokens — already short, so there is no padding to cap.
Cache economics for Compliance Assistant (claude-haiku)
would save €0/yrExamined Compliance Assistant's claude-haiku cache: reuse 12.5× vs break-even 0.28× — the cache nets in your favor. Healthy; nothing to change.
Cache economics for Compliance Assistant (claude-sonnet)
would save €0/yrExamined Compliance Assistant's claude-sonnet cache: reuse 12.5× vs break-even 0.28× — the cache nets in your favor. Healthy; nothing to change.
Cache economics for Credit Memo (claude-sonnet)
would save €0/yrExamined Credit Memo's claude-sonnet cache: reuse 12.3× vs break-even 0.28× — the cache nets in your favor. Healthy; nothing to change.
Cache economics for Credit Memo (gpt-4o-mini)
would save €0/yrCredit Memo's gpt-4o-mini bills cache-writes at or below the input rate ($0.15/1M vs $0.15/1M) — churn carries no premium, so there's nothing to fix.
Prompt caching for Portfolio Analytics
would save €0/yrPortfolio Analytics's prefix (~490 tokens) is below the provider's 1024-token minimum cacheable length — caching cannot engage.
Cache economics for Portfolio Analytics (claude-sonnet)
would save €0/yrExamined Portfolio Analytics's claude-sonnet cache: reuse 12.5× vs break-even 0.28× — the cache nets in your favor. Healthy; nothing to change.
Cache economics for Portfolio Analytics (gpt-4o-mini)
would save €0/yrPortfolio Analytics's gpt-4o-mini bills cache-writes at or below the input rate ($0.15/1M vs $0.15/1M) — churn carries no premium, so there's nothing to fix.
Response length for Portfolio Analytics (claude-sonnet)
would save €0/yrExamined: output is 65.3% of this segment's cost, but responses average only 190 tokens — already short, so there is no padding to cap.
Batch the offline Python Quant jobs
would save €0/yrThe pinned catalog carries no batch tariff for Python Quant's models — every schedule in it is synchronous, on-demand. Batch pricing is published per provider; Bayeto has not captured it, so there is no discount to quote and this move is not priced.
Prompt caching for Python Quant
would save €0/yrPython Quant's prefix (~289 tokens) is below the provider's 1024-token minimum cacheable length — caching cannot engage.
Cache economics for Python Quant (claude-sonnet)
would save €0/yrExamined Python Quant's claude-sonnet cache: reuse 8× vs break-even 0.28× — the cache nets in your favor. Healthy; nothing to change.
Cache economics for Python Quant (gpt-4o-mini)
would save €0/yrPython Quant's gpt-4o-mini bills cache-writes at or below the input rate ($0.15/1M vs $0.15/1M) — churn carries no premium, so there's nothing to fix.
Response length for Python Quant (claude-sonnet)
would save €0/yrExamined: output is 66.1% of this segment's cost, but responses average only 160 tokens — already short, so there is no padding to cap.
Cache economics for Research Copilot (claude-haiku)
would save €0/yrExamined Research Copilot's claude-haiku cache: reuse 12.5× vs break-even 0.28× — the cache nets in your favor. Healthy; nothing to change.
Cache economics for Research Copilot (claude-sonnet)
would save €0/yrExamined Research Copilot's claude-sonnet cache: reuse 12.5× vs break-even 0.28× — the cache nets in your favor. Healthy; nothing to change.
ledger version 22c35e346df765c3 — identical on every page, export, and decision record of this workspace state · how to verify it