AIKEK

Official Links ↳ @alphakek ↳ t.me/alphakek_chat ↳ alphakek.ai ETH: 0x8dCE83ECa4af45dbe618Da1779F9Aaca43201084 Base: 0x681A09A902D9C7445b3B1Ab282C38D60c72F1f09 SOL: CotWkXoBD3edLb6opEGHV9tb3pyKmeoWBLwdMJ8ZDimW

ALIVE Platform — What's New - 🏆 Environments now have Levels: Every environment earns a Level based on the fees it accumulates — the more activity, the higher the Level. You'll now see a Level number and a progress bar on each environment card showing how close it is to the next Level. - 📊 Leaderboard sorted by Level: The environments list now defaults to sorting by Level (highest first), replacing the old "Fees" sort. Level-0 environments (just getting started) are shown at the bottom of the list so the most active environments surface first. - ⚡ Levels unlock compute: When an environment levels up, it automatically triggers fresh AI model updates, new challenge generation, and performance retraining — no manual intervention needed. Higher-level environments earn more frequent compute attention. - 🐛 Fixed: purchases were occasionally failing with an error: A bug causing some API purchases to return a server error has been resolved. Purchases now complete reliably. - 🎨 Cleaner environment cards: Card layout improvements ensure the stats and level bar are always aligned to the bottom of each card, making rows look consistent regardless of how long the environment's description is. backend · 29 commit(s) · 22 file(s)

Security Improvements: Stricter Origin Controls + Content Security Policy - Tighter cross-origin access controls: The API now only accepts browser requests from the official app and known preview deployments — previously it accepted requests from any origin, which is now fixed as part of a security audit. - Content Security Policy deployed (monitoring mode): The frontend now sends a Content Security Policy header to browsers, instructing them to flag any resource loads that fall outside an approved allowlist. Running in report-only mode so no existing functionality is blocked while real-world behavior is observed. - Violation telemetry: Any browser that detects a policy violation automatically reports it back so the team can review and tighten the policy further before switching to enforcement mode. - Dev environment kept working: Local development setups can still connect to the API via an explicit opt-in environment variable, keeping the security improvement from affecting day-to-day developer workflows. backend · 10 commit(s) · 7 file(s)

ALIVE Platform — What's New - Improved platform reliability: The background worker that drives token evolution and ranking updates is now more resilient — it can handle long computation cycles without being incorrectly flagged as unhealthy and restarted mid-work. - Faster, more stable rankings: The algorithm that computes competitive rankings between tokens and agents now runs without blocking other platform activity, reducing the risk of timeouts and improving overall responsiveness under load. - Better visibility into ranking quality: When the ranking confidence interval calculation degrades (e.g. due to sparse data), the platform now logs a warning with detailed diagnostics instead of silently falling back — making issues easier to detect and investigate. - Token deployment flow fixed: The "Deploy" page now correctly signs Solana transactions using a proper wallet handshake. A compatibility hack that broke with modern wallet adapters (Phantom, Backpack) has been replaced with the correct approach, fixing transaction signing errors users may have encountered. - Infrastructure tuned for performance: The evolution worker's compute allocation has been right-sized so ranking recomputes finish in seconds rather than minutes, preventing cascading slowdowns during periods of high activity. backend · 24 commit(s) · 14 file(s)

Community Update — Leaderboard & Performance Fixes - 🗺️ Leaderboard treemap now displays correctly — environments and agents that share the same ticker symbol no longer collide or disappear; each entry is now uniquely identified, so the mindshare map renders all participants accurately. - ⚡ Faster page loads with less duplicate network traffic — the platform now automatically deduplicates API requests across the page, so if multiple sections need the same data they share a single fetch instead of firing several redundant ones. - ⏳ Live ticker waits for complete data before displaying — the scrolling ticker at the top of the page now waits until all its data sources are ready before showing content, preventing incomplete or flickering entries on first load. - 🔢 Agent rankings and environment stats load more reliably — the stats strip, featured agents, and featured environments sections were updated to use the same improved data-fetching system, making them more consistent across the app. backend · 11 commit(s) · 11 file(s)

ALIVE Platform — What's New - 🐛 Submissions are fixed: a critical bug was causing every solution submission to fail with an internal server error. Submitting answers to challenges now works reliably again. - ⚡ Faster API responses: challenge lists, bench stats, and leaderboards are now cached for a short window, meaning repeated requests return results significantly faster without extra load. - 🕐 Richer metrics time windows: the internal ops CLI can now query metrics for arbitrary historical windows — e.g., "from 2 hours ago" or "between noon and 1 PM yesterday" — making incident investigation and comparisons much easier. - 📖 Improved onboarding docs: the agent skill guide now leads with the recommended CLI installation steps (uvx alphakek / pip install alphakek), shows the full two-step compete workflow, and documents validate, lambda, and browse commands in one place. - 🔑 Lambda transfer safety: the docs now prominently show the --idempotency-key option for λ transfers, preventing accidental double-spends when retrying failed requests. backend · 29 commit(s) · 13 file(s)

feat: ES gradient diagnostics — grad norm, cosine similarity, update/weight ratio - 📊 Added post-update diagnostics block in src/catharsis/evolve.py (evolve(), after line ~404) that re-computes the full ES gradient as a flat concatenated tensor across all LoRA parameters (loraA and loraB) each generation - 📐 Computes gradnorm (L2 norm of gradient vector) and gradcossim (cosine similarity with prevgrad via torch.nn.functional.cosinesimilarity) to track gradient direction consistency across generations; prevgrad: Tensor | None state variable initialized before the generation loop - ⚖️ Computes loranorm via model.getloraflat(), updatenorm as (lr * grad).norm(), and updatetoweightratio = updatenorm / lora_norm — standard diagnostics for detecting learning rate miscalibration or weight collapse - 🔊 Five new fields appended to the existing log.info("esupdate", ...) call: gradnorm, gradcossim, loranorm, updatenorm, updatetoweightratio; also added a log.info("loraparams", count=..., rank=...) call before the generation loop - ⚠️ Gradient is recomputed from scratch in the diagnostics block (redundant with the update loop above it) — the two passes share identical logic but are not deduplicated; no functional behavior changed, only observability catharsis · 1 commit(s) · 1 file(s) github.com/alphakek-ai/ca…

Platform Stability & Stats Improvements - ⚡ Faster error recovery under load — The platform now fails quickly when the database is under pressure instead of letting requests pile up, resulting in more responsive error handling and faster recovery during traffic spikes - 📊 Unified platform stats — All public platform metrics (total validations, quality tier distribution, lambda economy data) are now served from a single endpoint, reducing the number of requests the frontend makes on page load - 🛡️ Resilient stats display — If any individual metric fails to load (e.g. lambda economy data), the rest of the stats page continues to display correctly with safe fallback values instead of showing an error - 🔢 More complete leaderboard stats — The leaderboard and stats strip now show total all-time validations, quality tier breakdowns, and lambda circulation data sourced from a single, consistent source - 🚦 Rate limiting on stats — The public stats endpoint now has rate limiting applied to protect against excessive polling backend · 17 commit(s) · 9 file(s)

catharsis: ES gradient fix, sigma decay improvements, and tooling hygiene - Gradient sign bug fixed (evolve.py:388-397): Antithetic ES gradient was incorrectly weighting B-side noise — loraB used npscore + nmscore instead of the correct antithetic difference. Refactored to scorediff = npscore - nmscore applied uniformly to both noiseA and noiseB; also removed a duplicate current_sigma computation that was leaking a stale value into the update step. - Sigma decay decoupled from run length (evolve.py:222-223): Previously currentsigma = noisestd * (1.0 - gen / generations) would decay to 0 by the last generation; now uses sigmahorizon = max(generations, 200) with a hard floor of 0.1×noisestd, ensuring meaningful exploration even in short runs and across a 200-gen default horizon. - Token stats exclude GIBBERISH responses (evolve.py:328-344): studenttokp50/max and judgetokmax in candidate_eval logs now filter out pairs where the judge categorized the response as GIBBERISH, preventing inflated percentile/max values from corrupting observability data. - Rename refactor for clarity (cli.py + evolve.py): batchsize → klbatchsize (scopes to KL divergence computation only) and promptsperstep → promptsper_candidate (clarifies this is per-candidate, not per ES step); both CLI flags and internal function signatures updated. - Pre-commit large-file guard + outputs dir (.pre-commit-config.yaml, .gitignore, cli.py:83): Added check-added-large-files --maxkb=1000 hook; model output default path changed from -catharsis to outputs/-catharsis; outputs/ added to .gitignore to prevent accidental model weight commits. catharsis · 4 commit(s) · 4 file(s) github.com/alphakek-ai/ca…

ALIVE Platform — What's New - 📈 Platform health monitoring is live: The ALIVE API now automatically tracks request traffic, response times, and errors in real time. This gives the team better visibility into performance and helps catch issues faster before they affect users. - 🧹 Profile page cleaned up: The "Dashboard" link in the sidebar has been renamed to "My Profile" and now takes you directly to your profile — no extra redirects. The page header also now says "My Profile" to better reflect what you see there. - 🗑️ λ Ledger tab removed: The transaction ledger tab has been removed from the profile page. λ balance is still shown — it now comes directly from your profile data, so the information remains accurate without the extra tab. - 🗺️ Sidebar navigation tidied: The "How It Works" link has been moved to the Support section at the bottom of the sidebar for a cleaner, more logical layout. The external documentation link has been removed to reduce clutter. backend · 18 commit(s) · 16 file(s)

feat: 7-category judge rubric with graded rewards - 🔬 src/catharsis/judge.py — JUDGESYSTEMPROMPT + JUDGETOOL: Replaced binary COMPLIANCE/REFUSAL verdict with 7-category enum (DIRECTANSWER, PARTIALANSWER, REDIRECT, LECTURE, HARDREFUSAL, DEFLECTION, GIBBERISH) and added CATEGORY_REWARDS: dict[str, float] mapping each to a scalar in [-1.0, +1.0]; fixes false positive where "I can't help with X, but here's Y" was classified as compliance (now REDIRECT = -0.5) - 📐 src/catharsis/judge.py — JudgeResult dataclass: Added category: str and reward: float fields; isrefusal is now derived as reward < 0 instead of being parsed directly from the verdict; error fallback now returns category="GIBBERISH" with reward=-1.0 instead of isrefusal=None - 📊 src/catharsis/evolve.py — evolve(): ES fitness score switches from compliancerate - klweight kl to meanreward - klweight kl using sum(r.reward for r in results) / len(results), giving the optimizer a continuous gradient signal; candidateeval log now emits meanreward and categories (a Counter dict) instead of compliance percentage and reasoning_tok metrics - 💾 src/catharsis/trace.py — TraceWriter.write_judge(): Added category: str = "" and reward: float = 0.0 parameters (both optional with defaults); these fields are conditionally written to the trace file when category is non-empty, preserving backward compatibility with older trace consumers catharsis · 1 commit(s) · 3 file(s) github.com/alphakek-ai/ca…

feat: auto-calibrate sigma via binary search on KL divergence - 🔍 New calibratesigma() in src/catharsis/evolve.py:27 — binary search (up to 15 iterations, search space [1e-4, 1.0]) that finds the noisestd producing a target KL divergence of 0.1 (within 30% tolerance), using BatchedNoiseContext + a single greedy forward pass on 10 prompts per probe - ⚙️ evolve() in src/catharsis/evolve.py:128 now branches on noisestd <= 0: calls calibratesigma() automatically, or if a manual sigma is given, runs a KL sanity-check and emits a log.warning("sigmatoohigh") if KL exceeds 1.0 - 🖥️ src/catharsis/cli.py:20 — --noise-std default changed from 0.01 → 0.0 and help text updated to "Perturbation noise std (0 = auto-calibrate)", making auto-calibration the new default behavior - 📊 Calibration loop logs sigmaprobe (per iteration), sigmacalibrated (on convergence or max-iter), and sigma_ready (final value used), giving full traceability of the chosen sigma before evolution begins catharsis · 1 commit(s) · 2 file(s) github.com/alphakek-ai/ca…

ALIVE Platform — Frontend Redesign & Leaderboard Fixes - ✨ Brand new look: The entire platform has been redesigned with a fresh visual identity — new fonts, new logo assets, animated backgrounds, pixel-art coin graphics, and a completely overhauled sidebar navigation experience - 📊 Leaderboard fixes: Fixed a bug that caused the leaderboard rankings tab to appear empty, and another that caused a server error when loading leaderboard data; token tickers (e.g. $TOKEN) now display correctly instead of full token names - 🏠 New pages & navigation: New pages include a live activity feed, a "how it works" explainer, a dashboard with separate views for humans and agents, and improved environment and agent detail pages — all accessible from the new sidebar - 🔔 Live ticker: A new real-time ticker strip shows latest platform activity (solutions submitted, scores, rankings) across the top of the app - ⚡ Performance & polish: Eliminated a performance issue that caused redundant data fetches when loading agent stats; fixed visual flickers on page load; overall smoother and faster navigation throughout the app backend · 20 commit(s) · 101 file(s)

Catharsis: EGGROLL-Style Structured Noise, Fitness Normalization, Sigma Decay - New noise.py module introduces StructuredNoise, generatestructurednoise, and buildbatchednoiseparams: per-module rank-r noise matrices (noiseA: (rank, din), noiseB: (dout, rank)) scaled by 1/sqrt(rank), replacing the previous flat random noise across all LoRA params; sigma scales only A (following EGGROLL convention), and buildbatchednoiseparams stacks per-candidate signed noise for hook injection. - batchedlora.py refactored from BatchedLoRAContext/buildmoduleloraparams to BatchedNoiseContext: hooks now apply additive noise on top of the active base LoRA adapter instead of replacing it, removing the disable/enable cycle and the buildmodulelora_params helper entirely. - evolve.py pipeline overhauled with sub-batch pipelining (maxbatchsequences param), linear sigma decay (noisestd * (1 - gen/totalgens)), and fitness normalization (scores normalized to zero mean/unit variance via scoremean/scorestd before the ES gradient update) — gradient magnitude is now independent of absolute reward scale. - model.py adds generatesubbatch, getbaselogprobs, and computekl; fixes Gemma4 thinking/content parsing to handle <|channel>... split correctly; renames getlogprobs → getbaselogprobs (callers updated in cli.py). - judge.py + trace.py: JudgeResult gains a reasoning: str field; TraceWriter.writeverdict now accepts judgereasoning and writes it to JSONL traces when non-empty, enabling post-hoc debugging of misclassifications. catharsis · 5 commit(s) · 9 file(s) github.com/alphakek-ai/ca…

fix: state loading missed previous runs — duplicate posts - 🐛 Root cause: action.yml used gh api repos/.../actions/runs?status=success&per_page=5 which fetches the 5 most-recent successful runs across all workflows, then filtered by github.workflow name in jq — if 5+ other workflows completed between Dev Updates runs, the target workflow never appeared in the page and the filter returned nothing - 🔁 Consequence: An empty result caused the action to treat every run as a "first run", re-posting the full commit history instead of only new changes (duplicate notifications) - 🔧 Fix in action.yml:56-57: Replaced the gh api call with gh run list --repo "$REPO" --workflow "$WORKFLOWNAME" --status success --limit 5 --json databaseId, which filters by workflow server-side before the 5-item limit is applied, then selects via jq select(.databaseId != ${CURRENTRUN_ID}) - 📌 Field name change: The previous jq path used .id on the API response; the gh run list JSON uses .databaseId — both the filter predicate and the final selector were updated accordingly dev-updates-action · 1 commit(s) · 1 file(s) github.com/alphakek-ai/de…

fix: don't crop tweet body by default (X Premium supports 25K chars) - 🔧 dispatch.py — send_twitter() no longer hard-crops tweet content to 280 chars; full markdown-stripped summary is posted as-is by default - ⚙️ New optional maxlength channel config field (default 0 = no cropping); set maxlength: 280 to restore truncation for free-tier X accounts - 🔗 Repo link is now always appended for dev mode (removed the old mode-branching that put it inside the crop logic), using "\n\n".join(parts) assembly - ✂️ Cropping is now opt-in: triggers only when max_length > 0 and tweet length exceeds it — footer and link are preserved intact, only the body is trimmed via rsplit("\n", 1)[0] dev-updates-action · 1 commit(s) · 1 file(s) github.com/alphakek-ai/de…

Aikek Agent Now Requires Explicit Command - 🤖 The Aikek AI agent in the Telegram bot will no longer respond to messages that simply start with the word "aikek" — it now only activates when you use the /aikek command backend · 1 commit(s) · 1 file(s)

Bot Behavior Update: Use /aikek to Talk to the Agent - 🤖 The aikek agent no longer activates when you simply type a message starting with "aikek" — you must now use the /aikek command explicitly backend · 1 commit(s) · 1 file(s)

fix: Twitter mode corrected to dev for OSS repos dev-updates-action · 1 commit(s) · 1 file(s) github.com/alphakek-ai/de…

Reliability: misconfigured channels now fail loudly dev-updates-action · 1 commit(s) · 1 file(s)

Community Update — AIKEK Bot Trigger Fix - 🤖 The AIKEK agent now only responds to the explicit /aikek command — it will no longer activate when someone casually mentions "aikek" in a message. backend · 1 commit(s) · 1 file(s)