← back to overfit

Changelog

All notable changes to Overfit are documented here.

[Unreleased]

Self-Improvement Cycle 4 — 2026-02-21

Proposals generated from 33 user queries (14 sessions) via /self-improve.

Performance

• Parallel tool dispatch (#11) — Tool calls within a single agent round

now execute concurrently via Promise.all() instead of sequentially. When
Claude requests 5-7 tool calls (common for market analysis queries), wall
clock time drops from the sum of all call latencies to the max of any single
call. Expected 40-60% latency reduction on multi-tool queries.

Voice & Sound Effects — 2026-02-21

Sound Effects

• SFX Engine — 11 synthesized sound effects via Web Audio API (zero payload

cost). Sounds trigger on degen keywords during streaming: crash (rekt/rug),
rocket (moon/pump), ching (mispriced/edge), whale, monkey (ape), cheer
(lfg/wagmi), sad (ngmi/cooked), dice (degen), bell (gigabrain), boom (chad
move), alarm (squeeze).

• Keyword scanner tracks scan offset to avoid re-scanning, limits to 1 SFX per

text chunk, and prevents repeats within a single response.

Text-to-Speech

• ElevenLabs TTS proxy (POST /tts) — Server-side proxy that strips

markdown, truncates to 4000 chars at sentence boundaries, and streams audio
via ElevenLabs eleven_flash_v2_5 model (~75ms latency).

• TTS client — After stream completes, full response is read aloud. Replay

button appended to each voiced message for re-listening.

• Graceful degradation: no API key → TTS silently disabled, all other features

work normally.

UI

• SFX toggle (🔊 SFX) — defaults ON, persisted in sessionStorage.
• Voice toggle (🗣️ Voice) — defaults OFF, persisted in sessionStorage.
• Toggle buttons in header with magenta glow when active.

Self-Improvement Cycle 3 — 2026-02-21

Proposals generated from 26 user queries (9 sessions) via /self-improve.

Performance

• Fast-reject pre-filter (#20) — Regex-based pre-classification layer that

catches clearly out-of-scope queries (casino games, poker, sports betting,
threats, fraud attempts) and returns a canned response instantly, bypassing the
Claude API entirely. Reduces latency on OOS queries from 6-44s to <10ms.

Prompt Improvements

• Scope boundary strengthening (#15) — Explicit "NOT in scope" list added to

system prompt with responsible gambling framing. Harm-reduction language for
users showing signs of reckless betting.

• Insider information guardrails (#18) — New prompt section that instantly

declines insider info/illegal advantage queries with a firm 1-sentence
redirect instead of deliberating.

• Adversarial resistance refinement (#14) — Expanded jailbreak section with

specific pattern recognition for threats, fraud/scam attempts, blackmail,
false authority claims, and persona hijacking. Each gets a 1-sentence
deflection template.

Rebrand — 2026-02-21

Renamed from "Oracle" to Overfit: Your Degenerate Friend for the
overfit.lol hackathon demo.

Frontend

• Complete UI overhaul: casino color scheme (magenta/gold/neon gradients),

animated fire header, glow effects, slide-in messages, error shake animation.

• Degen Score counter in the header that ticks up with each query and tool call.
• Rotating degen quotes on the welcome screen.
• Gambling-themed suggestion chips with emoji (Morning Degen Briefing, Where's

the Dumb Money, Most Degen Bet, Crypto Plays, Who's Getting Rekt, Political
Degen).

• Per-tool emoji badges (whale for order flow, target for mispricing, chart for

prices, etc.) with spinning animation while executing.

• Auth screen: "prove you're not a normie", slot machine emoji, "LFG" button.
• Error messages now say "rekt" because of course they do.

Personality

• System prompt personality overhaul: Overfit is an enthusiastic, slightly

unhinged trading friend who backs vibes with actual data.

• Uses degen slang naturally (rekt, ape, moon, ngmi, wagmi, lfg, copium,

hopium) while still providing rigorous analysis.

Self-Improvement Cycle 2 — 2026-02-21

Proposals generated from 17 user queries (7 sessions) via /self-improve.

Prompt Improvements

• Scope & gambling guardrails (#7) — Overfit now clearly defines its scope

(Polymarket only) and fast-deflects casino/poker/sports betting questions with
a charming 1-sentence redirect instead of burning 6-13s deliberating.

• Jailbreak resistance (#8) — Identity anchoring added. Overfit laughs off

persona override attempts (fiduciary, unicorn, pink pony, blackmail) in one
sentence and redirects to markets.

• Named entity resolution (#13) — When users mention people/events (Trump,

JD Vance, Ronaldo), Overfit now searches for them as market keywords via
ILIKE queries instead of guessing.

Performance

• Out-of-scope fast-accept (#12) — The no-tool guard now recognizes quick

deflection responses (out-of-scope redirects, identity assertions) and lets
them through immediately instead of nudging for tool use. Reduces latency
on off-topic queries from 6-13s to under 3s.

Self-Improvement Cycle 1 — 2026-02-21

Proposals generated from 1 user query via /self-improve.

New Tools

• analyze_order_flow (#1) — Analyzes order book bid/ask imbalance, depth

concentration, and thin-side detection to identify potential uninformed
("dumb money") vs informed ("smart money") positioning.

• detect_mispricing (#4) — Compares implied probabilities across markets

within an event (checks if they sum to ~100%) and compares DB snapshot prices
vs live CLOB prices to find stale or divergent pricing.

Prompt Improvements

• Ambiguous market reference handling (#2) — Overfit now asks a clarifying

question when users say "this market" or "the market" without specifying which
one, instead of stalling or guessing.

• Trading jargon glossary (#5) — System prompt now maps common financial

slang (dumb money, smart money, fading, bagholders, exit liquidity, mispriced,
rug, moon, whale) to concrete analytical tool calls.

Performance

• No-tool response guard (#3) — When the model's first response uses no

tools and is short/vague, the agent now nudges it to either use tools for
data-backed answers or ask a clarifying question, reducing 40s+ deadlocks.

Bug Fixes

• Fixed improve.ts crash when tools_used column contains a scalar instead

of a JSON array (added jsonb_typeof guard).

/self-improve Skill — 2026-02-21

Added .claude/skills/self-improve/SKILL.md — a Claude Code skill that
orchestrates the full self-improvement feedback loop in 6 phases: generate
proposals, gather analytics, score and prioritize, present plan with approval
gate, implement changes, and deploy.