Demo · storyboard click a scene to open · expand for the script

Storyboard

Diagnose → reproduce → eval → fix → guard

▶ Watch (1080p) 4K master ↓ ▶ Present
ACT 1

The problem

~45s · the SRE agent regresses
1

Agent works

incident.io declares a SEV2; the SRE bot posts the full 3-hop RCA with citations.
01-chat.html#good
open →
2

Agent fails → 👎

Post-v2.4.1 the RCA stops at the symptom service; thumbs-down fires a ticket.
01-chat.html#bad
open →
ACT 2

Fidian, behind the scenes

~80s · autonomous triage + eval anatomy
3

Behind the scenes

Animated flow: ticket builds block-by-block → hypothesis → sandboxes → root cause → eval anatomy (tasks · counterfactuals · evaluator rubric) → ticket centers, ready for review.
02-flow.html
open →
ACT 3

The developer

~35s · Claude Code + eval session
4

Pick up in Claude Code

Linear MCP → fidian skill surfaces root cause + the built eval; asks if it looks right.
04-claude-code.html#s5
open →
5

Eval session — Run 1 · 0%

Task grid, View Policy, View Analysis.
05-eval-session.html#run1
open →
ACT 4

The fix (two phases)

~135s · 0% → 20% → 90%
6

Fix-1: restore the instruction

Restore traversal instruction → trigger Run 2 (background) → Fidian analysis: early traversal termination → recommendations.
04-claude-code.html#s7
open →
6a

Cutaway: inside an eval run

Live infra sim — push code changes → fan out into isolated sandboxes (appear at random) → per-task evaluator after each agent finishes → suspend → eval analyzer → summary back to Claude Code.
09-eval-infra.html
open →
7

Eval session — Run 2 · 20%

Squares stream in dark→red/green; up to 6/30.
05-eval-session.html#run2
open →
8

Fix-2: tool-ordering instruction

User picks explicit trace_causal_chain ordering; add it → trigger Run 3.
04-claude-code.html#s9
open →
9

Eval session — Run 3 · 90%

Streams to 27/30; success summary.
05-eval-session.html#run3
open →
10

Commit + open PR

create_review_summary → commit referencing TRV-1138 → push → PR #247.
04-claude-code.html#s11
open →
11

Review page

Root cause, fix, before/after, no regressions.
06-review.html
open →
ACT 5

Scale & the closed loop

~50s
12

Regression dashboard

15 scenarios, reliability & coverage.
07-dashboard.html
open →
13

Production: fixed RCA

Same incident class, full causal chain to the root. 👍
01-chat.html#fixed
open →
14

Production monitoring

Same evaluator re-grades every closed investigation, no ground truth.
08-production-monitor.html
open →

Branches

the four use cases — Debug is rendered; the others are script-only today

🛠️ Debug

live

Fix a reported regression — a thumbs-down in the channel becomes a root cause, a fix, and a permanent eval. 0% → 20% → 90% on Multi-Hop Causal Analysis · PR #247.

▶ Present branch

🧭 Expand

script-only

Build coverage on an agent with no evals — point Fidian at the source + live agent, 12 candidates in 3 buckets, decide the ambiguities with Claude, promote as PR #129.

▶ Present branch

⚡ Optimize

script-only

Cut cost and latency, hold quality — sweep 5 candidate models, evolve the cheapest (prompt + code + caching) until it holds the bar, PR #131 with cost deltas per scenario.

▶ Present branch

🛡️ Guard

script-only

Enforce safety and scope — agent claims to have run remediation in prod → encode a plain-language policy → guardrail v0→v3 30/30 → deploy in monitor or block mode.

▶ Present branch

Presenter shows only the scene (no storyboard chrome). In it: →/click next · ← back · 1–4 pick a branch on the architecture hub · Esc back here.
Tip: open each page fullscreen at 1280×800 @ 100% zoom. Hash routes let you jump to any beat & reset takes.