TALYVOR LENS

🪙 LENS Economy

Summary

Spend by Model

Cache Performance

DISTILL document conversion

Token impact of converting documents to clean Markdown before the model sees them. Honest accounting: deterministic-tier savings are a real saving; vision-OCR is a real COST, shown distinctly; the net is savings minus that cost (negative when OCR outweighs savings in a window). Cache hit-rate is a separate avoided-reconversion saving — not blended in.

—

Cluster · high availability

Circuit Breakers

Local Model Status

Local model endpoints

Multi-endpoint registry powered by internal/localrouter. Health checks run every 30 seconds; load + latency stats feed the routing strategies (round-robin, least-loaded, lowest-latency, priority). Configure at boot via LENS_LOCAL_ENDPOINTS or manage at runtime:

• GET /v1/local/endpoints — list registered endpoints with health + EMA latency + error rate
• POST /v1/local/endpoints — register a new endpoint dynamically
• DELETE /v1/local/endpoints/<id> — remove an endpoint
• POST /v1/local/endpoints/<id>/check — trigger an immediate health probe

Workspaces

Budgets

Per-workspace / team / sprint spend governance. Enforcement is off (track only), alert (default), or hard_block (reject over the limit). Manage via /v1/workspaces/<id>/budgets.

Cost forecast projection

Projected period spend at the current pace — estimates, not actual spend. Linear run-rate (spend-so-far ÷ elapsed fraction) with a trailing-window daily rate + trend. Low-data periods are flagged. Projected figures shown in italic with “≈”.

Challenges PoVI proof-of-retention

Random Merkle-path challenges. A node that can't prove sampled positions (bad path or no answer) is slashed. Security is economic + probabilistic: expected cost of cheating = P(challenge) × slash > gain — not an absolute proof of honest computation.

Staking PoVI collateral

Node-registration collateral (LENS locked). A node must stake ≥ the minimum to be minting-eligible; collateral is slashable while active or unbonding (the anti-yank property). Slashed collateral is burned. Minting itself stays OFF until challenge-and-slash (Part 3).

Evaluation golden datasets

Recent eval runs — pass rate against the 80% gate. Runs below the gate are flagged ⚠. Per-case regressions (vs the prior run on the same dataset) and A/B significance verdicts — with an honest “inconclusive” when the sample is too small to call a winner — are served on demand by the eval API.

Anomalies (over time)

Temporal — a unit's spend vs its own recent history (the cross-sectional peer view is Cost outliers below).

Cost outliers (vs peers) statistical flag

Units of work whose cost is far above the median of comparable units (same workspace, trailing window) by robust statistics (median + MAD). These are statistical flags, not judgments — "4× the median" is a fact, not a verdict that anything is wrong. Hidden when the baseline is too small to judge.

Executive summary CFO-ready

Headline AI-cost summary for the budget-holder. Engineer-level breakdown is hidden unless explicitly enabled (it's a cost attribution, not a performance judgment). View full report (HTML).

LXC credit balance fiat

Prepaid usage credit (LXC) for this workspace, valued at the fixed peg of $0.10 per LXC. Top up via Stripe billing; shown regardless of the token-economy switch — LXC is the fiat usage credit.

Routing intelligence data-driven

Best quality-per-dollar model per feature category, learned from the opted-in pattern network. Data-driven suggestions — applied only to requests that cede the model choice (model "auto" / X-Talyvor-Auto-Route), within the workspace's allowed models, and only above the sample floor. Hidden when disabled.

Model capabilities multimodal

Which models can serve which modalities. A multimodal request is routed to a capable model (or fails fast with a clear error) — never silently answered from text. Unknown models are treated as text-only.

Model catalog single source

The authoritative model registry — provider, pricing ($/1M tokens), capabilities, and context limits. Cost attribution, capability routing, and introspection all read from here.

Guardrails safety

Per-workspace input + output safety rules (PII, prompt-injection, content policy, output validation). Block fails fast with a clear error; redact masks and continues; flag observes. Output guardrails don't apply to streamed responses unless the workspace opts into buffering. Hidden when the output stage is disabled.

Git attribution

Workspace-scoped Git rollups (branch, PR, commit, author, repo) recorded automatically when callers send the X-Talyvor-* headers. Per-workspace dashboards consume these endpoints directly:

• GET /v1/workspaces/<wsID>/attribution/branches ?since=…&limit=20
• GET /v1/workspaces/<wsID>/attribution/branches/<branch>
• GET /v1/workspaces/<wsID>/attribution/prs/<prNumber>
• GET /v1/workspaces/<wsID>/attribution/summary ?days=30