Analysis

Most writing tools hide their internals from you. You click a button, something happens, and if the output looks right you trust it. If the output looks wrong, you have no idea why, and the only thing to do is try again. Ishvana takes the opposite approach. There’s a dedicated top-level tab called Analysis whose entire job is to show you what the tool is actually doing — which chapters are on track, how the lore is clustering, how the agents have been performing, what the LLM engine is spending on your behalf, and what errors have been quietly logged while you were writing. It’s the dashboard you open when the question is “is everything working the way I think it is?” and you want an honest answer.

The Analysis tab is split into six panels. Each one answers a different question about your project’s health. You don’t have to look at them daily — most authors only open Analysis a few times a week, usually after a major writing session — but when you do, the information is there, structured, and sortable.

What’s in this section

Editorial Analysis Three-phase editorial diagnostics: writing cadence, ProseGuard rule violations, and manuscript-level structure review.

Lore ML Machine learning analysis of your Legendry — clusters, anomalies, contradictions, coverage, and entity graphs.

Agent Overview Per-agent activity, success rates, and operation metrics — how your specialists are actually performing.

Etherforce Observability LLM engine performance — routing decisions, cost tracking, model benchmarks, and the tool registry.

Legendry Bench Mutation-based LLM benchmarking — how well does a model catch contradictions in your actual lore?

Error Tracking Crashes, exceptions, and diagnostic traces — the Mime subsystem surfaced for users.

The six panels, one-liner each

Editorial Analysis — Are you writing enough, writing consistently, and is your prose holding up against the rules you’ve set for it? The three-phase workflow goes from rhythm (writing stats) to line pressure (ProseGuard) to scene shape (manuscript review), in that order, and it’s the answer to the question “is my prose actually working or am I fooling myself?”

Lore ML — How does your Legendry actually cluster? Which entries are anomalies? Which ones contradict each other? Are there obvious gaps in your coverage? The Lore ML pipeline runs machine learning analyses on your worldbuilding database and surfaces structural patterns you can’t see by reading entries one at a time. The tab is heavy — runs locally, takes a few minutes to train on a large project — but the output is the kind of meta-insight most authors never get about their own world.

Agent Overview — Which agents have been running? How often? How successful are they? What’s their average latency? The Agent Overview panel reads the Agent Transparency metrics and gives you a per-agent dashboard — so when GameMaster’s success rate drops after you switch models, you see it in the numbers before you feel it in the output quality.

Etherforce Observability — Under the hood, every agent request flows through the Etherforce LLM engine. Etherforce picks a model, routes the request, counts the tokens, tracks the cost. The Observability panel shows you all of it — decision logs, cost breakdowns by model and agent, model registry, operational stats, the full tool registry, and the skills system. This is the tab you open when you want to understand why the agent did what it did.

Legendry Bench — A benchmarking workflow that tests models on your actual lore. It mutates real lore entries with tracked changes (“change a character’s hometown from Port Royal to Teslan,” “add a contradictory section”) and runs the model against the mutated lore, scoring how well it catches the contradictions. Different from generic LLM benchmarks because it tests the model on your world, not on a canned dataset. Useful when you’re picking between models for your project.

Error Tracking — The Mime subsystem logs every exception and crash that happens inside Ishvana, with stack traces, context, and a three-tab diagnostic UI. Most of the time you won’t open this. When something goes wrong, it’s here — and the structured logs are way more useful than “sorry, something went wrong.”

What Analysis is not

It’s not a daily dashboard. You don’t need to check Analysis every morning. Most of its panels are diagnostic tools you reach for when you have a specific question (“why did that last Lorekeeper check take forever?”, “how much have I spent on LLM calls this week?”, “why did the agent fail on that last generation?”). The answer lives in Analysis.

It’s also not a replacement for the Desktop tab. Desktop is the overview of your writing — which books are on track, which need attention, what’s moving. Analysis is the overview of the tool itself — which agents are working, what the LLM is doing, what errors have been logged. The two are complementary. Desktop for “where am I on my book,” Analysis for “how is Ishvana itself performing.”

Why this exists at all

Giving you visibility into the tool is a product value, not a hidden developer feature. If your writing assistant is doing 90% of its job well and 10% of it poorly, you deserve to know which 10% so you can compensate. If your model provider is draining your token budget faster than expected, you deserve to see the cost curve before the bill arrives. If a silent error is killing half your Lorekeeper checks without you realizing, you deserve to see the error log.

The alternative — “trust us, everything is working” — is how most tools ship, and it’s how users lose trust in them over time. The Analysis tab is the opposite bet: show everything, let you check, earn the trust. It costs some UX surface area, but it’s worth it.

What’s next

Editorial Analysis Start here if the question is 'is my prose working?'

Agent Overview Start here if the question is 'are the agents doing their jobs?'

Etherforce Observability Start here if the question is 'what's the LLM engine actually doing?'