Show HN: LoopVera – evidence ledger for pipeline debugging (no code yet)

English · 简体中文

See inside your vision pipeline.Bring industrial-grade runtime debugging to the open-source vision stack — while exploring theruntime evidence ledger for agent workflowsunderneath.

The loop:see every step → tweak one param & rerun → diff this run against the last → human sign-off when stakes are high

. Library-agnostic · replayable · drops into your existing code, replaces no operator.

⚠️ Process open source · not a finished product · and this repo is just getting started.LoopVera is beingopen-sourced step by step, in public— first the reasoning and design, then the specs, then the code skeleton, then working slices. See the[.]RoadmapPhase 1 (Problem & Vision) is done; Phase 2 (Architecture & ADRs) is next.This repo has governance and Phase 1 narrative docs —no runnable code here yet. Starring means"watch this direction,"not"production-ready."

Test accuracy drops from 94 to 87. You sprinkle five cv2.imwrite

calls, your folder fills with debug_003_v2_final.png

, and to compare two versions you just eyeball two windows — a week later you can't recall what you changed or why it "looked fixed."

The algorithms aren't weak; the dev tooling is. LoopVera targets the see → tweak & rerun → diff across runs → human sign-off when stakes are high layer — it does not replace your operators.

Today's agents can edit code, tweak params, and batch runs — but often lack a closed-loop substrate underneath: intermediates scatter across chat and folders, runs don't reconcile, and high-stakes changes have no clear record of who authorized what next. LoopVera aims to put humans and agents on the same evidence ledger; the vision workbench is the first load-bearing skin, not the whole story.

This is the interaction path we are

building toward.pip install loopvera

andloopvera open

arenot published yet.

import loopvera as lv

def detect(img):
    gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
    lv.observe("gray", gray)                                   # one line captures the intermediate
    _, mask = cv2.threshold(gray, lv.param("thresh", 130), 255, cv2.THRESH_BINARY)
    lv.observe("mask", mask)                                   # lv.param makes this tweakable in the UI
    return mask
pip install loopvera              # ⏳ not published — will build from source first
python my_pipeline.py             # run once; a run is recorded automatically
loopvera open                     # ⏳ not delivered — local browser workbench

Read the full pain-point + vision write-up: ** LoopVera Vision** ·

简体中文.

The

LoopVera column is target design— not a claim that everything works today.

imwrite + folders | ad-hoc matplotlib | wire up Rerun yourself | closed commercial vision IDE | LoopVera (target) | | |---|---|---|---|---|---| | every step visible | ✅ | ✅ | ✅ | || | tweak param & rerun | ❌ | ❌ | ❌ | ✅ | ✅ | side-by-side diff across runs + param diff | ❌ | ❌ | ✅ | || | library-agnostic (no operator lock-in) | ✅ | ✅ | ✅ | ❌ own only | ✅ | adds the loop for AI (evidence closed loop) | ❌ | ❌ | ❌ | ✅ | | | price | free | free | free | 💰💰💰 | free / open source |

The differentiator isn't "can it show an image" — it's tweak-and-rerun + cross-run diff, without locking you into anyone's operator library.

Optional. Explains what the vision workbench stands on.

LoopVera stands on a domain-neutral runtime-evidence substrate — the same loop applies to any "behavior is invisible" domain (embedded, signal, medical, robotics):

① see            → ② tweak & rerun → ③ diff across runs → ④ high-stakes only:
   capture any                                                human sign-off
   intermediate                                                     │
        ↑______________________________ next round ________________│

Adding a domain takes three hooks (observe

/ decode

/ placement

); the loop itself is provided by the runtime. Vision is the first load-bearing skin, not the whole story — the substrate is designed for multiple domains (e.g. embedded

alongside vision

) so the loop isn't vision-only on paper; additional domains ship in later roadmap phases.

Depth → AI agent workflows · Why not LangGraph / all-in-one.

LoopVera is released in stages so that every step carries standalone value and invites a different kind of contributor. This is deliberate: the design should be critiqued before the code locks it in.

Full detail, entry points, and definitions of done: ** ROADMAP.md**.

— forLoopVera Visionvision / CV engineers: problem, loop, comparison, honest status.— forAI agent workflowsAI agent developers: the evidence ledger your agents are missing.— forPipeline debugging pain pointsanyone tuning pipelines: a problems-only discussion piece.— forWhy not LangGraph / all-in-oneagent-infra: graph orchestration vs. an evidence/authority ledger.

All docs are bilingual — see the docs index.

This is build-in-public capability exploration, not a commercial launch. There is no fixed ship date.

• ⭐ — follow the direction as it evolves.Star / Watch on GitHub - 💬 — real pipeline pains and architecture RFCs. This is theGitHub Discussionsnormative source of truth for decisions. - 🎮 — chat & build help (not normative; conclusions go to Discussions).Discord · Open Build - 📖 — see labels likeIssuesgood-first-issue

,help-wanted

, anddesign-rfc

. - 🤝 Read before opening a PR, and theCONTRIBUTING.mdCode of Conduct.

Early contributors' feedback directly shapes which happy path we wire first.

MIT © 2026 LoopVera / Proofrun.

See every step. Diff every run. Orchestration swappable. Workers swappable. Ledger not.