+
+
+
+
+
+
+
+
+
+ ![]()
+ STAMPED principles and their role in the age of
+
+ Live slides/Sources: + https://datasets.datalad.org/centerforopenneuroscience/talks/2026-reproAI-MIT-STAMPED.html
+
+ +![]()
+ ![]()
+ ![]()
+ ![]()
+ ![]()
+ ![]()
+ ![]()
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+ ![]()
+ Seven properties a research object should have —
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+ ![]()
+
+
+
+ Tracked:
+
+
+
+
+
+
+ ![]()
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+ So I tried to
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+ ![]()
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+ Structure your objects, and
+ ![]()
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+ 
+ STAMPED principles and their role in the age of
AI-driven scientific software development
+
+ Agents often fail not for lack of "intelligence".
They fail for lack of structure to act on.
+
+
+
+
+
+ LLM Workshop — MIT — October 16th, 2026|
+ Yaroslav O. Halchenko +  @yarikoptic
+  @yarikoptic@fosstodon.org
+ |
+ + |
|
+ Center for Open Neuroscience + Department of Psychological and Brain Sciences + Center for Cognitive Neuroscience + Dartmouth College + New Hampshire, USA + + 
+ |
+ + | +
+ Live slides/Sources: + https://datasets.datalad.org/centerforopenneuroscience/talks/2026-reproAI-MIT-STAMPED.html
+
+ +
+ 
+ 
+ 
+ 
+ 
+ 
+
+
+
+
+
+
+ Reality Check
+ +
+ 
+
+
+
+
+
+ + Disclaimer: I am not here to sell you a model, a vendor, or a workflow. +
+
+ I am here to argue that the boring infrastructure work we have been doing for 20 years
+ is exactly what decides whether agents help you or hurt you.
+
+ Counts and names on these slides are links — follow along, click through later. +
+Standing on two talks
+|
+ Mine — ReproNim webinar, Feb 2026 + ReproFlow & YODA: structure your studies + (video) +
|
+
+ Oscar Esteban — OHBM 2026 Educational + Using LLMs in Neuroimaging Research: LLMs for Coding +
|
+
+ Oscar showed how to build the harness.
+ This talk is about what the harness stands on — and how harnesses themselves become shareable. +
+
+
++ This talk is about what the harness stands on — and how harnesses themselves become shareable. +
It’s 2026. Agents write the code.
+ +So the bottleneck moved
+| writing the code | +→ | +verifying and situating it | +
| can I express this? | +→ | +is this right, and does it fit what already exists? | +
+ 
+
+
+ + HBR, Feb 2026: “AI doesn’t reduce work, it intensifies it” +
+ +Without structure, the agent has to guess
+
+ What an agent finds in most labs
+
|
+
+ What an agent finds in a BIDS dataset
+
|
+
+ Same model. Same prompt. Different affordances.
+ On the left the agent improvises confidently and is wrong.
+ On the right it can act, and the validator catches it when it does not.
+
+ + Not just rhetoric: a structured neuroimaging harness moved first-choice tool-selection accuracy from + 23.3% to 93.6% across seven models — + Chen, …, Poldrack, + Bringing analytic rigor to agentic AI for science, arXiv:2608.19902, 2026 + (brain-researcher.com). +
+ +This problem already has a name
+
+ 
+ YODA, 2018 — for humans composing studies + |
+ → | +
+ 
+ STAMPED, 2026 — generalized, gradeable, domain-agnostic + |
+
+ We did not invent these for AI. That is the point — they were already the right properties. +
+
+ Seven properties a research object should have —
whoever, or whatever, picks it up next.
+
+ STAMPED in one table
+| Property | +In one line | +What it buys an agent | +
|---|---|---|
| Self-contained | +Everything under one boundary | +A bounded context window. Nothing it needs is “on Bob’s laptop”. | +
| Tracked | +Version control & provenance | +A diff to review, and a record of who/what produced each change — including the agent. | +
| Actionable | +Executable & documented workflows | +The load-bearing one. A command it can run, not a paragraph it must interpret. | +
| Modular | +Independent & composable modules | +A unit small enough to reason about, swap, and verify in isolation. | +
| Portable | +Runs anywhere with explicit specs | +The agent’s sandbox is not your laptop. If it only runs on your laptop, it does not run. | +
| Ephemeral | +Transient compute environment | +Throw-away containers — exactly how agents already work. No accreted state to corrupt. | +
| Distributable | +Shareable, persistent, retrievable | +It can get the inputs itself instead of asking you for a Dropbox link. |
+
+ + A. Macdonald, C. Baker, I. To, Y. O. Halchenko. + STAMPED: principles for reproducible research objects. + Preprint, 2026. doi: 10.31222/osf.io/f3h82_v1 + · + sources: stamped-principles/stamped-paper + · + paper.stamped-principles.org +
+ +Actionability is the one that agents live or die on
+
+
+
+ An actionable object answers, without a human in the loop:
+ +
+ And standards are what make Actionability cheap.
+ You do not negotiate the layout per project — you inherit it.
+ “You’ve seen one BIDS dataset, you’ve seen them all” was a convenience for humans. + For a model that would otherwise have to infer the layout, it is the difference between acting and hallucinating. +
+
+ + “You’ve seen one BIDS dataset, you’ve seen them all” was a convenience for humans. + For a model that would otherwise have to infer the layout, it is the difference between acting and hallucinating. +
+ “All standards are ‘bad’, but some are used.” — D. Clunie, MICCAI 2017. + A used standard beats a correct one — and now it has a second constituency. +
+ +… and AI changed what counts as actionable
+| + | “Actionable” required… | +… so what could act | +
|---|---|---|
| Before | +An exact command, API call, or documented CLI. Prose was not enough. | +Tools, and humans who already knew the tool. + Everything else stayed a paragraph in the Methods. | +
| Now | +A procedure described in human language — + with scripts dropped in exactly where precision matters. | +Any agent, on first contact. + The interface moved from syntax to language, so vastly more of what we + already wrote down became executable. | +
| Next | +Routines that fire without being asked. | +The harness itself — CI, PR monitors, scheduled agents. | +
+
+
+
+
+
+ 1. Skills — logic written in language
+
|
+
+ 2. Routines — the harness runs them for you
+
|
+
+ AI did not only make more things actionable. It made automation on top of
+ actionability cheap — which is the real multiplier.
+
+
+ So grade them too: a skill is a research object
++ A skill that only runs in your repo, on your paths, undocumented and unversioned, + is a silo with a friendlier interface. +
+| Property | +… applied to a skill or an automation | +
|---|---|
| Self-contained | +Carries its own references and scripts. Not “see our internal wiki”. |
| Tracked | +In git, with a metadata.version and a history — so a change to the procedure is reviewable. |
| Actionable | +Runnable commands and scripts, not only prose. Recursive, and the point. |
| Modular | +One skill, one job. Composable with the others instead of a mega-prompt. |
| Portable | +No hardcoded paths, no “works on our cluster”. Declares its dependencies and versions. |
| Ephemeral | +Runs in a fresh sandbox. No accreted state it silently depends on. |
| Distributable | +Published where someone else can find and install it — not pasted into chat. |
+ Good automation and good skills pass the STAMPED checklist.
+ That is not a purity test — it is precisely what makes them reusable
+ beyond the person who wrote them.
+
+
+
+ Walk it yourself: checklist.stamped-principles.org
+ (MUST / SHOULD / MAY, item by item) ·
+ examples ·
+ principles ·
+ paper source
+ (main.tex
+ — small enough to paste into a context window, which is how we actually grade a study today).
+
What we built to make this real
+Twenty years of it, in neuroimaging, before anyone said “agent”.
+Standards are the grammar — BIDS as the worked example
+
+
+ 11 modalities · 25 active BEPs ·
+ a machine-readable schema ·
+ a validator ·
+ BIDS-Apps with one calling convention
+
+
+ The schema is the part that matters here: the standard is not prose an agent reads,
+ it is data an agent queries. And the validator is the type-checker that
+ turns a confident wrong answer into a failing build.
+
+
+ But a layout only tells you what the files are.
+ It says nothing about how they came to be, or what it cost to make them.
+ Standard →
+
+
+ datalad run → duct:
+ what it is → how it was made → what it took.
+ Three layers, and an agent needs all three.
+ Tracked: datalad run writes the record for you
+$ datalad run -m "run brain extract workflow" \
+ --input "inputs/ds*/sub-01/anat/sub-01_T1w.nii.gz" \
+ --output "sub-*/anat" \
+ bash -e code/brain_extraction.sh inputs/ds*/sub-01/anat/sub-01_T1w.nii.gz$ git show --stat
+commit 1223bc4bead3c7e6be15b78546c894c6465dd7a4
+
+ [DATALAD RUNCMD] run brain extract workflow
+
+ === Do not change lines below ===
+ {
+ "cmd": "bash -e code/brain_extraction.sh inputs/ds000001/sub-01/anat/sub-01_T1w.nii.gz ...",
+ "dsid": "b656ac80-973c-45ef-887e-1b8d6f6a51dd",
+ "exit": 0,
+ "inputs": ["inputs/ds*/sub-01/anat/sub-01_T1w.nii.gz"],
+ "outputs": ["sub-*/anat"],
+ "pwd": "."
+ }
+ ^^^ Do not change lines above ^^^
+
+ sub-001/anat.nii.gz | 1 +
+ sub-001/anat_mask.nii.gz | 1 +-
+
- Machine-readable, in the commit message — no sidecar database to lose +
- Clean git state in, clean git state out: no surprises +
datalad rerunreads it back and re-executes
+
+ Example from the DataLad Handbook. + Found in the wild in ~15% of indexed DataLad repositories on GitHub. +
+Provenance does not care who wrote the code
+
+ Agent-written or hand-written, the [DATALAD RUNCMD] record is the same shape.
+
+ The tests do not know the code was AI-generated either. +
+
+ This is why Tracking is the principle that makes AI disclosure
+ mechanical instead of aspirational — and we will come back to that at the end.
+
+
+ The missing piece: con/duct
+
+
+ Wrap any command. It records, next to the outputs:
+ wall time, CPU, peak RSS, per-process-tree samples, exit code, environment, full
+
+
+ stdout/stderr — as plain, self-contained files.
+ … and that log is the provenance you need later
+
+ “Why did this get OOM-killed on the cluster?”
+
|
+
+ 
+
|
+
+ Why an AI audience should care: when you hand a failing pipeline to an agent,
+ it needs numbers and text it can read, not a human recollection.
+
+ duct turns “it crashed, I think it ran out of memory” into a file.
+ Self-contained (S), tracked (T), immediately actionable (A).
+
+ Already wired in: set REPRONIM_USE_DUCT=1 with
+ ///repronim/containers
+ and every datalad containers-run gets a duct trace for free.
+
All three layers, one real artifact — built with AI
+
+ dandi/compression-comparisons-t261-study
+ — a reproduction of
+ Buccino et al. 2023 on ephys compression,
+ plus a comparison against the new ITU-T T.261 codec.
+
+ 1. The layout (what)
+BIDS-shaped, YODA-composed: data and tool are + subdatasets pinned by commit, not copies. + |
+
+ 2. The record (how) — and 3. the cost (what it took)
+The recorded command is a |
+
| commits | [DATALAD RUNCMD] |
+ of those, wrapping duct |
+ cells | recorded compute | +AI disclosure | +
|---|---|---|---|---|---|
| 1265 | 1216 | +1095 | +912, 0 errors | +3838 h (peak RSS 120 GB) | +26 Co-Authored-By: Claude |
+
So I tried to rerun it. Live, on this laptop.
+
+ |
+
+ Did the science come back?
+
+ Different machine, different kernel, six weeks later — + every scientific number came back bit-for-bit. + Only timings and the environment record moved. + + |
+
+ And the checklist earned its keep: the one thing that broke was
+ Distributable — a relative submodule URL that resolved only on
+ the author’s own disk. S, T, A and
+ M all held. You do not find that by reading your own repo;
+ you find it by being someone else.
+
+
+ So what do we build next?
+Not more code. Harnesses and skills.
+Agent = Model + Harness
++ formulation borrowed from Oscar Esteban, OHBM 2026 +
+| Harness piece | +What it is | +STAMPED property | +
|---|---|---|
AGENTS.md / CLAUDE.md |
+ The project’s operational contract: conventions, commands, what not to do | +Self-contained | +
| Skills | +A named, reusable procedure with its own references and scripts | +Modular, Actionable | +
| Hooks / CI | +Guard rails that run whether or not anyone remembered to | +Tracked | +
| MCP servers | +The port to external systems — archives, trackers, databases | +Distributable | +
| Container / sandbox (con/yolo) |
+ Where it is allowed to run, and how much it can break | +Portable, Ephemeral | +
+ The harness is itself a research object. Which means it can be
+ self-contained, tracked, modular, portable — or it can be a pile of prompts in your notes app.
+
+ Where are you on the ladder?
+
+ 
+ + Ladder figure & framing: + Jo Van Eyck, + “You are Not Ready: Agentic coding in 2026” + + |
+
+ L4–L5 is exactly where a bad harness stops being survivable —
+ which is why the harness, not the model, is the thing to invest in.
+
+ |
And the loop that gets you up it
+|
+
+%%{init: {'theme': 'base', 'themeVariables': {'fontSize': '17px'}}}%%
+flowchart TD
+ IDEA["Idea"] --> S1["1. Overall design / spec"]
+ S1 --> S2["2. Aspect design / prompt"]
+ S2 --> S3["3. AI work"]
+ S3 --> S4["4. Review / test"]
+ S4 -->|"iterate"| S2
+ S4 --> S5["5. Compound into spec / SKILL"]
+ S5 -->|"next aspect"| S2
+ S5 -->|"adjust / memorize"| S1
+
+ style IDEA fill:#9c27b0,color:#fff,stroke:#7b1fa2,stroke-width:2px
+ style S1 fill:#1565c0,color:#fff,stroke:#0d47a1,stroke-width:2px
+ style S2 fill:#2e7d32,color:#fff,stroke:#1b5e20,stroke-width:2px
+ style S3 fill:#e65100,color:#fff,stroke:#bf360c,stroke-width:2px
+ style S4 fill:#c62828,color:#fff,stroke:#b71c1c,stroke-width:2px
+ style S5 fill:#283593,color:#fff,stroke:#1a237e,stroke-width:3px
+
+ |
+
+
+ Stage 5 is where a skill is born.
+ Vibe coding throws the lesson away; compound engineering writes it down —
+ and what it writes down is exactly the artifact we just said must be STAMPED.
+
+ + Tooling for stage 1: spec-kit · + OpenSpec · + compound-engineering · + LAD + (in use in dandi-cli) + + |
+
That was the individual ladder. Here is the team one.
+
+ + Figure reproduced from: + E. C. Johnson, T. T. Nguyen, B. K. Dichter, …, + M. E. Martone, F. Pestilli, D. Yatsenko. + SciOps: Achieving Productivity and Reliability in Data-Intensive Research. + arXiv:2401.00077, 2024. +
+
+ The preprint names BIDS, NWB, DataLad/git-annex, DANDI and brainlife.io as its own
+ Level-3 exemplars — and reserves Level 5 (AI + human in the loop) for the top.
+
L5 presupposes L3 underneath it. You do not get closed-loop discovery + on top of
+
+ L5 presupposes L3 underneath it. You do not get closed-loop discovery + on top of
study_final2/.
+ Right now, everyone is rebuilding the same private harness
+
+ A hundred labs, each writing their own “how to work with BIDS” prompt.
+ None of them shared. None of them versioned. None of them tested.
+
+ We have seen this failure mode before. It was called
+ “every lab has its own analysis scripts”, and we fixed it with standards and packages.
+
+ Skills are the new reusable module.
+ Apply Modularity and Distributability to the harness,
+ or watch AI recreate the silo problem at machine speed.
+
+
+ We are already doing this — upstream, in public
++ K-Dense-AI/scientific-agent-skills + — 177 scientific skills, + Agent Skills spec, MIT-licensed, scanned and reviewed. +
+| Contribution | +Who | +What it carries | +
|---|---|---|
skills/bids+ PR #125 — merged |
+ ours (CON) | +Entity grammar & 35 entities · the + machine-readable schema shipped as a reference · + all 25 BEPs · + HeuDiConv/ReproIn/dcm2bids/BIDScoin conversion paths · + validation · derivatives & BIDS-Apps | +
skills/datalad+ PR #227 — merged |
+ Dylan Pulver, with our upstream review |
+ Clone/get semantics & the “pointers are not data” failure ·
+ datalad run / rerun / containers-run as the provenance spine ·
+ publishing to siblings · STAMPED landed in
+ references/provenance.md on our suggestion |
+
| PR #265 — open | +Brent McPherson (McGill) |
+ Maps all seven STAMPED properties onto concrete DataLad commands,
+ plus a read-only evidence probe per property
+ (datalad subdatasets -r for S/M, git log --grep '[DATALAD RUNCMD]' for T,
+ datalad rerun --report for A, datalad siblings for D…) |
+
| issue #217 — closed | +ours (CON) | +“Should STAMPED be its own skill?” — and the still-open question underneath: + where should domain skills live? An upstream monorepo, or the project’s own repo, with discovery on top? | +
+ Note the shape: a community member wrote the DataLad skill, we reviewed it, and a third party + then extended it with STAMPED. That is exactly how a standard spreads — and it is + now happening to harnesses. +
+And domain-agnostic ones: con/skills
+| Skill | +What it automates | +
|---|---|
| introduce-reuse-compliance | +SPDX / REUSE licensing metadata — including for BIDS datasets |
| credit-contributions | +CRediT author-contribution statements from git history |
| tinuous-analyzer | +CI-failure investigation across archived con/tinuous logs |
| bisect-and-patch-git-annex | +Bisect a git-annex regression from CI logs and produce a red/green PR |
| issue-triage · pr-feedback-review | +Triage and review work, cross-referenced against the codebase |
| introduce-codespell · introduce-intuit-auto | +Bootstrap spell-checking / automated releases into any project |
+ Each of these started as a thing we did by hand three times and got tired of.
+ Writing it down as a skill is how the n-th time gets cheaper — for us,
+ and for whoever clones the repo.
+
+ + Deeper dive on the workflow around these: + “A few words of intro into AI assisted coding” + (CA Origami Retreat, 2026) — also in the appendix of this deck. +
+Two regimes — and be honest about which one you are in
++ Either the system checks itself against goals and spec — + or the work items stay small enough that a human still can. +
+|
+ A. Scaffolded self-verification + the agentic system keeps itself in sync +
+ Only then are large autonomous batches defensible.
+
+ |
+
+ B. Small, human-reviewable, hand-tried + where almost every research project actually is +
+ Not a lesser mode — the correct default until A is actually built.
+
+ |
+
+ There is no third regime where you ship large, unverified agent output safely.
+ And the way out of B is not a better model — it is
+ Actionable specs, Tracked provenance and
+ Modular units: the same STAMPED properties, again.
+
+ + “Well, less is more” — Robert Browning, Andrea del Sarto, 1855; + later the motto of Mies van der Rohe, who credited his mentor Peter Behrens rather than claiming it. +
+ +We are not the only ones saying this
+| Source | +What it found / argues | +
|---|---|
| DORA 2025 + (Google Cloud) | +Higher AI adoption raises throughput and delivery instability. + Their prescription is literally regime A+B: + “working in small batches is a critical countermeasure to the risks of AI-assisted + development”, plus test automation and fast feedback — because + “verification is a fundamentally different cognitive task than creation”. | +
| METR RCT, 2025 | +16 experienced OSS developers, 246 real issues, on repos they already maintained: + 19% slower with AI tools — while estimating they had been 20% faster. + The perception gap is the hazard, not the tooling. + (METR now labels this historical — tools moved on. The gap is the durable lesson.) | +
| Addy Osmani | +“Proof over vibes.” Verification, not authoring, is the bottleneck; + break agent output into digestible commits; no PR without tests or a demo; + humans stay non-negotiable for security and architecture. | +
| Simon Willison + ▶ Pragmatic Summit talk + ▶ Lenny’s Podcast |
+ “Your job is to deliver code you have proven to work.” + The skill is instructing and then confidently verifying; a model cannot take accountability, you can. + And on regime B specifically — even with tests, he reports that + “I haven’t used it yet” still undermines his confidence in a change. + Trying it by hand is not superstition. | +
| Faros telemetry + (~22k devs) | +PR size +51%, median time in review +441%, + incidents per PR +243%. Big agent diffs do not get reviewed — they get waved through. | +
+ Converging, from very different vantage points:
+ the safe size of an AI-assisted work item is set by the strength of your verification loop.
+ Build the loop, or keep the items small. Preferably both.
+
+
+ HI ↔ AI — every project picks its own policy
++ Rows 1–2 live in regime B; rows 3–4 are only honest once + regime A is actually built. +
+| Stance | +OSS exemplars | +CON / adjacent example | +Leans on | +
|---|---|---|---|
| Reject AI-generated content | +Zig, Krita, Clojure, QEMU | +git-annex upstream — pure-HI commits | +AI stays ephemeral; never enters the tracked artifact | +
| Accept with disclosure | +NumPy, Kubernetes, Linux kernel, Django | +DataLad, DANDI —
+ Co-Authored-By: Claude… trailers, @pytest.mark.ai_generated |
+ Tracking — provenance is non-negotiable | +
| Spec-driven AI-generated (HI specs, AI writes, HI reviews) |
+ Common in practice; rarely a declared policy | +AnnexTube, + citations-collector, + parts of dandi-cli | +T + A + M — the spec is the contract | +
| Autonomous agents in the loop | +Brain Researcher + (Poldrack lab) — agentic neuroimaging analysis; + K-Dense AI — “AI co-scientists” + (and the host of the skills library we contribute to) | +con/skills + + con/yolo for triage, review, dep updates | +All of STAMPED — the harness becomes the type-checker | +
+ Common ground, all four: AI cannot be an author + (ICMJE, Jan 2026); + humans retain full responsibility; AI use is disclosed inside the artifact. + STAMPED Tracking is what makes that mechanical rather than a promise. +
++ Catalog of declared OSS stances: + melissawm/open-source-ai-contribution-policies. + — Note what the row-4 exemplars are: harnesses. + Brain Researcher constrains the agent to + acceptable analyses and required validations rather than trusting it — + autonomy is bought with structure, not with a bigger model. +
+Take home — what to do Monday
+-
+
- Make one object actionable. Pick one study. Put it in a declared standard + (BIDS, NWB) + and run the validator. The agent gets a grammar; you get a type-checker. +
- Record one command. Wrap your next analysis step in
+
datalad run+ — or at minimum induct. + A provenance record you did not have to remember to write.
+ - Grade one thing against STAMPED. + checklist.stamped-principles.org, + fifteen minutes, and you will know which property is your weakest. +
- Write down your harness. An
+
AGENTS.mdin the repo beats a prompt in your notes app. + Declare your AI policy there while you are at it.
+ - Size the work to the loop you actually have. Until the project + self-checks, keep AI-assisted items small, reviewable, and tried by your own hand — + and say so out loud, so nobody mistakes regime B for regime A. +
- Push one skill upstream. The next person who needs your procedure should
+
cloneit, not re-derive it. + scientific-agent-skills, + con/skills, or your own repo — but somewhere public.
+
+ Standards and provenance were always the right answer.
+ AI did not change the answer — it changed the deadline.
+
+ Acknowledgements
+ + +Thank you!
+Structure your objects, and
let both intelligences work on them
+
+ + stamped-principles.org · + checklist · + examples · + con/skills · + con/duct +
+Appendix
+Material cut for time — press ↓ to walk it, or follow the links.
+STAMPED and FAIR do different jobs
+| FAIR | +STAMPED | +
|---|---|
| About sharing: can someone else find this and get it? | +About the object being operational on its own: can someone — or something — re-run it? | +
A FAIR dataset can still be a .zip of .mat files with a DOI. |
+ A STAMPED object carries its own layout, environment, and execution record. | +
| Graded by a repository. | +Graded by a checklist you can walk, item by item. | +
+ Not competitors — different lenses on the same stack. FAIR tells you whether the world can reach it.
+ STAMPED tells you whether anything useful happens once it does.
+
+ Portable + Ephemeral: the environment travels with the record
+
+
|
+
+ 
+ ReproNim/containers — auto-rebuilt, versioned collection + |
+
+ “Depend on where you run, you should not.” An agent’s sandbox is not your laptop — + which makes portability a correctness property, not a courtesy. +
+