Skip to content
BayetoThornbury Fixed Income (demo)
Demo — fictional firm,
real engine output
Analyze your own system

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.

Active constraints: latency <= 1700ms · quality >= 0.75 · 3 recommendations survive this objective · 14 moves rejected as non-Pareto — e.g. “p95 latency would rise to 2910ms > cap 1700ms.

  • latency <= 1700 can reject: highest observed p95 2912ms; worst projected move -5% → 2766ms
  • quality >= 0.75 cannot bind here: lowest measured quality 1; worst projected move 0% → 1

Recommendations

3 generated optimization hypotheses, ranked by business value under each application’s objective. Nothing here is seeded — the engine produces them from telemetry.

Open moves3outstanding
On the table€909direct bill savings / yr
Cost validated0€0 cost-validated
Identified€3,454total value / yr

2 of 3 shown · filters active · Clear filters

Outstanding — ranked by value × confidence × objective alignment

1

Route small-context Portfolio Analytics traffic off claude-sonnet to GLM-5.2

€567
/yr bill saving
plus €968/yr latency value, at 0.05 €/s — a default, not a measurement
Outstandingpayback 142dconfidence 97%evidence Highcontext 4/6effort Low
3

Pin Portfolio Analytics serving to the caller’s region

€21
/yr bill saving
plus €1,031/yr latency value, at 0.05 €/s — a default, not a measurement
Outstandingpayback not economicconfidence 90%evidence Highcontext 4/6effort Medium
Ruled out by your objective · 14 generated, then refused by a constraint — not in the count above · €128,653/yr what these would have saved, if your constraints allowed them

These moves would save money and your objective forbids them. Each says which constraint bound and by how much — change the constraint and they become available.

Right-size Client Reporting: move low-output claude-sonnet traffic to claude-haiku

would save €1,457/yr
  • p95 latency would rise to 2910ms > cap 1700ms

Extend prompt caching across Client Reporting

would save €708/yr
  • p95 latency would rise to 2765ms > cap 1700ms

Compress the un-cached context Compliance Assistant re-sends every call

would save €1,604/yr
  • p95 latency would rise to 2025ms > cap 1700ms

Right-size Compliance Assistant: move low-output claude-sonnet traffic to claude-haiku

would save €5,967/yr
  • p95 latency would rise to 2132ms > cap 1700ms

Extend prompt caching across Compliance Assistant

would save €1,313/yr
  • p95 latency would rise to 2025ms > cap 1700ms

Compress the un-cached context Credit Memo re-sends every call

would save €2,112/yr
  • p95 latency would rise to 2395ms > cap 1700ms

Right-size Credit Memo: move low-output claude-sonnet traffic to claude-haiku

would save €8,255/yr
  • p95 latency would rise to 2521ms > cap 1700ms

Extend prompt caching across Credit Memo

would save €629/yr
  • p95 latency would rise to 2395ms > cap 1700ms

Compress the un-cached context Meeting Summaries re-sends every call

would save €6,952/yr
  • p95 latency would rise to 2766ms > cap 1700ms

Right-size Meeting Summaries: move low-output gpt-4-turbo traffic to claude-sonnet

would save €40,940/yr
  • p95 latency would rise to 2912ms > cap 1700ms

Finish the gpt-4-turbo → claude-sonnet migration: Meeting Summaries missed the rollout

would save €40,940/yr
  • p95 latency would rise to 2912ms > cap 1700ms

Compress the un-cached context Research Copilot re-sends every call

would save €3,240/yr
  • p95 latency would rise to 2393ms > cap 1700ms

Right-size Research Copilot: move low-output claude-sonnet traffic to claude-haiku

would save €11,930/yr
  • p95 latency would rise to 2519ms > cap 1700ms

Extend prompt caching across Research Copilot

would save €2,606/yr
  • p95 latency would rise to 2393ms > cap 1700ms
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/yr

The 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/yr

Examined 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/yr

Examined 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/yr

Examined: 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/yr

Examined 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/yr

Examined 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/yr

Examined 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/yr

Credit 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/yr

Portfolio 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/yr

Examined 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/yr

Portfolio 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/yr

Examined: 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/yr

The 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/yr

Python 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/yr

Examined 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/yr

Python 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/yr

Examined: 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/yr

Examined 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/yr

Examined 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 1b1a8afa88e54575 — identical on every page, export, and decision record of this workspace state · how to verify it

© 2026 Bayeto