Demo + assets
- Synthetic backtests now occasionally fail (regime failures, OOS degradation)
so verdicts span RECOMMENDED / RISKY / NOT_RELIABLE realistically. Phase 1
has ~35% failure rate, Phase 2 AI loop ~10%, OOS has 30% chance of severe
degradation — matches what real markets look like
- screenshots/dashboard.png + best_result_modal.png regenerated against the
current UI; screenshots/apex_demo.gif (6-frame autonomous-run timelapse)
embedded in the README
FEATURE 1 — Live AI token streaming
- AIReasoner._call_claude() now streams via SSE when a callback is
registered. Each text delta forwards to the dashboard as
`ai_thinking_chunk` events
- The Live AI Thinking Feed renders a single growing bubble with a blinking
cursor while text streams in, finalising on `end`. Looks and feels like
watching the AI type
FEATURE 2 — Pre-flight check on /setup
- New /api/preflight endpoint runs 5–7 probes: config readable, API key
set, MT5 paths exist (skipped in demo), EA registered, reports folder
writable. Returns {ok, blocking_count, checks[]}
- Setup page renders a colour-coded checklist on load and refocus.
Replaces "click Start, wait 5s, see generic error"
FEATURE 3 — Hot-reload settings into the running pipeline
- pipeline.reload_config() applies AI model / timeout / API-key swaps to
the live reasoner mid-run. Threshold changes surface for next run
- /api/settings POST detects a running pipeline and calls reload_config(),
returning the changed keys plus a "hot-reloaded into the running
optimization" note
FEATURE 4 — Replay scrubber on Best Result
- Evolution path now renders as an interactive scrubber: range slider +
prev/next/play buttons. Each step shows the run ID, phase, score, full
metrics grid, parameter changes for that step, and the AI's analysis
text — auto-plays at 700ms/step
FEATURE 5 — Compare runs on /reports
- Each card has a checkbox; selecting 2–4 reveals a floating Compare bar.
Compare modal renders a side-by-side table with metric winners
highlighted (Calmar / PF / profit favour higher; DD favours lower)
and a parameter-diff section showing changed values
FEATURE 6 — Discord / Slack / generic webhook on completion
- New `notifications.webhook_url` + `webhook_style` config keys
- Auto-detects Discord vs Slack from the URL host. Posts a one-line
summary on `optimization_complete`: verdict + best run + PF/Calmar/DD/
profit/trades/elapsed
Refresh-during-run no longer wipes the dashboard
- pipeline now keeps in-memory rolling logs of ai_thinking, param_changes,
validation events, and the current early-termination state. Logs reset at
the start of each new run (capped: 200 thinking / 50 param-change records /
40 validation events)
- new GET /api/live_activity returns those logs + current phase + running
state in one shot — the dashboard hits it on page load
- restoreHistory() in dashboard.js now replays each event into addThinking /
addParamChanges / validationRunStart-Complete-Done / showEarlyTermination,
and re-applies the active phase via setPhaseActive. Reload F5 mid-run no
longer shows a fresh empty dashboard
- addThinking() preserves the original ts on replay (was using nowStr() so
every replayed entry got the refresh time)
Reports page (/reports) rebuilt
- previous template crashed with 500 on legacy summary.json files that
pre-date the win_rate / drawdown_pct fields. Server now backfills sane
defaults for every metric the template touches
- runs now sorted by ts (was filesystem-iterdir order)
- new template: search bar + filter chips (All / Exploration / AI Iteration
/ Validation / AI insight / Has .set), phase tags, AI/.set badges, three
action buttons per card (View Params / Download .set / Full Report)
- per-card detail modal shows metrics grid + full parameters table +
AI reasoning + Download .set button — the missing "click to see params
+ download" path the user reported
- has_set detected per-run by globbing run_dir/*.set; AI insight tag shown
when ai_insight.json exists on disk
Major upgrade making the AI loop visible and the project ready for public release.
UI / UX
- Live AI Thinking Feed: streams reasoning, decisions, and outcomes per iteration
- Parameter Changes panel: prev → new + reason for every AI-driven edit
- Validation Activity panel: out-of-sample + sensitivity runs with live metrics
- Early Termination banner: surfaces why optimization stopped (targets met, no profit, budget, stuck, user stop)
- 3-phase tracker renamed Exploration / Iteration / Validation with live N/total
- Best Result modal exposes Evolution Path showing how the AI arrived at the winner
- Run-detail modal accessible from every recent run row
- Setup form validation (dates, walk-forward order, params selection, AI targets)
- Pause button removed; misleading sidebar nav consolidated to Dashboard / New Run / Reports / Source
Backend
- AIGuidedLoop streams ai_thinking, param_changes, ai_targets_met, ai_stuck
- Pipeline emits validation_start / validation_run_start / validation_run_complete / validation_done
- Pipeline emits early_termination on every early-stop path
- /api/best_result returns best run + full evolution chain
- /api/run/<id> + /api/runs sorted by ts
- AIReasoner falls back to ANTHROPIC_API_KEY env var when config is a placeholder
- Demo mode (APEX_DEMO_MODE=1) generates deterministic synthetic backtests so judges can run end-to-end without MT5
Open-source readiness
- README.md with pitch, demo flow, architecture diagram, quickstart, event reference
- LICENSE (MIT)
- config.example.yaml template (config.yaml now git-ignored)
- requirements.txt: added anthropic / requests / psutil / beautifulsoup4, capped majors
- .gitignore: secrets, *.set, scratch screenshots, ea_registry.yaml
- demo/run_demo.py: one-command offline demo runner
- 10 polished screenshots for README + judge review
PROBLEM: Old system repeated identical params, all scores flat at 0.2500,
user had zero control over symbol/TF/dates. Not smart, not dynamic.
NEW ARCHITECTURE:
optimizer/ (NEW package)
├── __init__.py
├── session_config.py User choices (EA, symbol, TF, dates, budget, objective)
├── lhs_sampler.py Latin Hypercube Sampling — diverse exploration
├── result_ranker.py Relative scoring (best in session=1.0, worst=0.0)
├── budget.py Time budget tracker
└── pipeline.py 3-phase orchestrator
Phase 1 — Broad Discovery (LHS, 20-26 runs):
Samples FULL parameter space, not just defaults±tiny step
LHS guarantees coverage: all 17 optimizable LEGSTECH params explored
Relative ranking: profitable configs float top, losers score 0
Phase 2 — Refinement (9 runs):
Neighbor search around top 3 configs at ±20% range (not ±0.5 step)
Keeps best of Phase1 vs Phase2 — never regresses
Phase 3 — Validation (5 runs):
OOS backtest on unseen data period
Sensitivity test: nudge params ±20%, detect fragility
Verdict: RECOMMENDED / RISKY / NOT_RELIABLE
Output: Clean downloadable .set file via /download_set/<run_id>
UI REDESIGN:
ui/templates/landing.html New / homepage (was old dashboard)
ui/templates/setup.html New /setup — EA, symbol, TF, dates, budget, objective
ui/templates/dashboard.html Updated /dashboard with:
- 5-step phase indicator
- Real progress bar per run
- Phase 1 results table (top 5 after phase1)
- Verdict banner with download button
- No-profitable-config warning
ui/static/js/dashboard.js Handles 8 new pipeline SocketIO events
app.py New routes: /, /setup, /dashboard
/api/start accepts full SessionConfig JSON
/download_set/<id> serves optimized .set
ea/registry.py +list_all() for setup page dropdown
ui/templates/reports_index.html Back to Dashboard → /dashboard (was /)
ui/static/css/style.css +dot-warn, dot-done, profit-pos/neg, aliases
FIXES:
Score no longer flat 0.2500 (was: absolute thresholds on losing EA)
User now controls: symbol, timeframe, dates, budget, objective
Parameters now span full range (was: tiny step from defaults)
Verdict is actionable: RECOMMENDED / RISKY / NOT_RELIABLE with reason
TESTED:
8/8 pre-flight checks pass
Browser test: landing ✓, setup form ✓, /dashboard ✓,
phase indicator active ✓, /reports ✓, back link ✓
1. CRITICAL FIX - Reports 500 Internal Server Error:
Root cause: {{}} in onclick was Jinja2 template expression
Fix: Changed to single {} in reports_index.html onclick
2. Score history chart always empty:
Root cause: score_update only emitted inside wfv.passed block
Fix: Emit score_update after every hypothesis run with {promoted:false}
Fix: Baseline run also pushes to chart via run_complete handler
3. IS gate too strict - optimizer produces 0 candidates forever:
Root cause: min_calmar=0.35 but best EA has calmar=0.165
Fix: Two-tier IS check - passes if composite_score improves vs baseline
(absolute thresholds still apply as alternative pass condition)
4. Findings emitted twice (double UI display):
Root cause: baseline analyzed once after run, then re-analyzed in iter 1
Fix: Cache baseline findings, reuse in iteration 1 without re-emitting
5. Chart labels improved:
'Baseline' for first point, 'It1·h1' for hypothesis runs
Calmar normalized -0.5..2.0 -> 0..1 for chart display
6. Best score header only updates on actual promotion (not per-hypothesis)
LEGSTECH Optimizer