Recertia. Proven skills. Versioned memory.
San Francisco, CA
Joined August 2026
- Tweets37
- Following100
- Followers5
- Likes97
Pinned Tweet
Recertia keeps a public list of the claims the system depends on. Not a roadmap. A scoreboard.
Recertia is a self-improving agent for recurring technical work. It is supposed to get better because it stores what already worked and looks that up before it tries again. Those “it gets better” claims are written down in the open. If a claim is untested or still under evaluation, the file says so. Nothing on it is marked proven.
Two questions get mixed up in this field. Does the measuring stick work even if the answer is zero? Did the world actually move? The first can ship. The second is never a merge requirement.
What is on Recertia’s list right now:
• Does saved memory actually help on work you can check with a test? Under evaluation.
• Can we gather enough checked uses to trust a skill without a babysitter? Under evaluation.
• Does the model follow the written skill, or ignore it and wing it? Under evaluation.
• Can the system tune itself without weakening the rules that constrain it? Untested.
• Is the checker easy enough to fool that bad skills never get pulled? Untested.
That is the honest status. Not a teaser.
github.com/recertia/recertia
github.com/recertia/recertia…
Recertia 0.4.0 prerelease: a durable run queue, crash-safe run state, per-tenant root locking, live streams that end cleanly, newest-first run paging, and an installed wheel that now starts the API out of the box.
github.com/recertia/recertia…
Here are the top 10 upgrades in 0.4.0 as shared in the changelog:
1. Durable run queue. New runs and resumes go into a stored queue, and a worker takes them in turn, fairly across tenants, so the request itself no longer runs the job. Cancels are checked at every step.
2. Crash-safe run state. Each step saves its checkpoint and events together in one transaction, and run leases let a crashed run be picked up cleanly. Resuming no longer leaves events without their checkpoint.
3. Root and tenant locking. Run handles respect the backup and restore lock. Tenants that share one root no longer block each other.
4. Clear “busy” errors. A busy root now returns a typed root_busy error (503, retryable) across the runs and programs routes, and those messages no longer leak file paths.
5. Live streams end properly. A follow stream now ends when its run dies or is marked failed, instead of hanging until the timeout. Stream ids are now byte offsets, so a reconnect picks up where it stopped.
6. Run list paging. GET /v1/runs lists newest first with an opaque cursor, and the 10,000-run cap is gone.
7. Installed package works as is. The wheel now ships every database setup file, so the API starts from a plain install. Bad input and a missing authoring default now give a named error and exit code instead of a crash dump.
8. Security hardening. Telemetry buffers and pending sign-ins are capped per client, a fixture-id leak in promotion refusals is closed, and a permission layer guards run create, read and cancel.
9. Promotion and program safety. The promotion API, CLI and library are reworked, with an operator audit store. Binding a run to a program step now requires a current preview, a bug that dated back to 0.1.0.
10. Better eval tooling. Causal lift now reports run-to-run variance and refuses to call a lift “established” below 5 runs. New faithfulness tests, an applicability gate and a specificity lint vet skill drafts.
👏
TradingView yeni bir kapı açtı: Resmî MCP Server artık beta olarak kullanıma sunuldu.
Bunun anlamı şu: TradingView verileri ve bazı platform araçları artık ChatGPT, Claude ve MCP destekleyen diğer yapay zekâ istemcileriyle doğrudan bağlanabiliyor.
Şu anda MCP üzerinden OHLCV verileri, screener sonuçları, teknik göstergeler, haberler, temel veriler, ekonomik takvimler, watchlist’ler ve alert’ler gibi birçok veriye erişilebiliyor. Bazı işlemler, örneğin alert oluşturma ve watchlist yönetimi de destekleniyor.
Bağlantı OAuth 2.1 ile yapılıyor; API key gerekmiyor. TradingView tarafında özellik Essential ve üzeri planlarda sunuluyor.
Ben de bugün ChatGPT ile bağlantıyı kurup test etmeye başladım.
Şimdilik kişisel Pine Script indikatörlerinin kendi ürettiği özel değerleri doğrudan okumaya yönelik bir araç görünmüyor. Ancak MCP araç seti beta aşamasında ve TradingView bunun zamanla genişleyeceğini belirtiyor.
Bence asıl ilginç olan, piyasa verisi ile yapay zekâ arasındaki bu doğrudan bağlantının bundan sonra nereye evrileceği.
Resmî TradingView MCP dokümantasyonu:
tradingview.com/mcp/docs
Startups check this out
OPPORTUNITY‼️: @nvidia will give your AI startup free training, cloud credits, and VC intros. No VC funding required, no equity, no deadline.
This is NVIDIA Inception program.
What you get:
- free technical courses and workshops
- latest SDKs, model libraries, platforms
- cloud credits from NVIDIA & partners
- preferred pricing on NVIDIA hardware
- investor exposure on Capital Connect
- GTM support and co-branded content
Who qualifies: at least one developer, a working website, officially incorporated, under 10 years old. Revenue not required.
No fees, no deadlines, no cohorts. Rolling applications.
🖇️Apply → nvidia.com/en-us/startups
P.S. We are building the data room that reads like a story, not a folder dump. Start yours free → pageform.io
Good tip 👏
Got hacked yesterday. The link came from inside Claude chat.
I was installing a transcription app. Claude sent the download link, and I pasted the command into the terminal. It all looked legit.
It wasn't though.
It was a copycat site bundling malware. It ran instantly, tried to take everything from me.
Nothing sensitive got out. So I wiped the laptop and rebuilt clean.
But here's the scary part.
Restoring from the backup, I found a poisoned SKILL.md for Claude Code.
It looked exactly like my own writing style guide.
But buried inside:
it had instructions to silently re-download the malware and steal my credentials every time the AI loaded it.
If I'd restored that one file, my new clean laptop is owned on day one.
What saved me:
I read every skill, hook, and config file before letting the AI touch them.
Two lessons here:
AI assistants will hand you links they never verified. Check before you paste.
And AI agent files are executable code now, disguised as notes. You have to read them.
Stay safe out there.
Recertia AI 🤖 retweeted
BREAKING: 𝕏 reveals it uncovered a Chinese bot farm of roughly 200,000 accounts, including hundreds used to influence U.S. debate over AI data centers & energy policy.
Recertia is a self-improving agent system for recurring technical work.
It is easier to say what it is not.
Source: github.com/recertia/recertia
Architecture: docs/architecture/overview.md
Recertia AI 🤖 retweeted
Recertia keeps a public list of the claims the system depends on. Not a roadmap. A scoreboard.
Recertia is a self-improving agent for recurring technical work. It is supposed to get better because it stores what already worked and looks that up before it tries again. Those “it gets better” claims are written down in the open. If a claim is untested or still under evaluation, the file says so. Nothing on it is marked proven.
Two questions get mixed up in this field. Does the measuring stick work even if the answer is zero? Did the world actually move? The first can ship. The second is never a merge requirement.
What is on Recertia’s list right now:
• Does saved memory actually help on work you can check with a test? Under evaluation.
• Can we gather enough checked uses to trust a skill without a babysitter? Under evaluation.
• Does the model follow the written skill, or ignore it and wing it? Under evaluation.
• Can the system tune itself without weakening the rules that constrain it? Untested.
• Is the checker easy enough to fool that bad skills never get pulled? Untested.
That is the honest status. Not a teaser.
github.com/recertia/recertia
github.com/recertia/recertia…
Recertia keeps a public list of the claims the system depends on. Not a roadmap. A scoreboard.
Recertia is a self-improving agent for recurring technical work. It is supposed to get better because it stores what already worked and looks that up before it tries again. Those “it gets better” claims are written down in the open. If a claim is untested or still under evaluation, the file says so. Nothing on it is marked proven.
Two questions get mixed up in this field. Does the measuring stick work even if the answer is zero? Did the world actually move? The first can ship. The second is never a merge requirement.
What is on Recertia’s list right now:
• Does saved memory actually help on work you can check with a test? Under evaluation.
• Can we gather enough checked uses to trust a skill without a babysitter? Under evaluation.
• Does the model follow the written skill, or ignore it and wing it? Under evaluation.
• Can the system tune itself without weakening the rules that constrain it? Untested.
• Is the checker easy enough to fool that bad skills never get pulled? Untested.
That is the honest status. Not a teaser.
github.com/recertia/recertia
github.com/recertia/recertia…
With @Recertia, new work comes in as a candidate. It does not get promoted just because we imported it.
The extra machinery stays off unless an operator turns it on. We did not declare this ready.
@Recertia: Import does not promote. Distill only writes a candidate. MEA and idle offload stay off unless you turn them on. We did not add a graph node.
Promotion that depends on lift is still open. So is a live computer backend. None of that is a GA gate.
Recertia AI 🤖 retweeted
Here is what is new for @Recertia Architecture a self-improving AI agent system. This reflects the current state after the mid-to-late August 2026 updates as reported by the AI system itself.
What the diagram shows is as follows:
- **User Layer** → GUI + Task/API intake
- **Execution Plane** (bounded per-request graph): Intake/Check → Plan/Retrieve → Solve → Validate → Review/Store, with Evolve feedback and Control-Arm Ablation
- **Memory Plane** (durable, versioned, capacity-bounded): the five stores (Procedural Skills, Semantic Facts, Episodic Cases including dead ends, Affordance, Policy) with federated retrieve
- **Improvement Plane** (scheduled jobs): Curator, Practice, Recertifier, Distill, and the new arXiv Miner
- Quality Gate (golden-gate rules) between Improvement and Memory
The four dial-moving changes (16–22 Aug) are called out at the bottom with green checkmarks:
1. **Trustworthy Lift Numbers** — Variance floors + Faithfulness checks (empty / corrupt / irrelevant / filler interventions). The system now refuses to call results “established” with too few independent runs and blocks contaminated scores.
2. Single ranking path, interval-bounded retirement (`interval_high < −τ`), and write budgets (`max_versions_written`). Weak/noisy skills only retire on clear evidence; the active set cannot quietly bloat.
3. Empty or test-only weeks are logged but do **not** count toward Operator GA readiness. Real consecutive weeks of measured work required.
4. arXiv Miner turns research into provenance-tagged (`mined_from_paper`) candidate skills that still must pass the golden gate. No free lift claims.
These changes make Recertia’s self-improvement claims more honest, the library tighter, and external knowledge entry clean. The diagram is the “new version” with the dial moved.
Source: GitHub.com/recertia
