Adversarial walkthrough: four skeptical newcomers
25 September 2026. This is an agent’s close-reading exercise using four fictional reader perspectives, not observed participant behavior, a usability study or four independent human reviews. I inspected the blog, both human onboarding documents, all domain cases, dissemination plan and prepared invitations. Edits were limited to the two onboarding documents; publication and domain prose were not changed in this pass.
Physics/astronomy newcomer: “I can check a number, but I cannot certify your instrument”
The blog’s calibration story gives a recognizable reason to care. The initial human path then asked me to pick a domain without a direct link and fill a blank card. As an early-career researcher, I could mistake “who can decide” for a request to appoint myself catalog editor, or think I needed a colleague before doing anything. Reading the full statistical chapter is an unnecessary prerequisite for contributing a scope correction.
Repair made: a linked astronomy starter now supplies raw 10.5 and both gains, with a sentence recording the two calculations while excluding physical calibration. A complete plain-language card shows a permitted personal classroom use without asserting catalog authority. “Version,” “dependency” and “rely” receive short definitions. Solo first contact is explicit. The paired exercise now assigns its roles to two people and warns that combining fictional roles does not model independent professional review.
What this reader can now do: reproduce 10.5 and 10.0, preserve the old numerical statement, and say that a new gain changes the current estimate. A reader who cannot perform the arithmetic can still identify an unclear phrase. No agent team, external calibration expert or appointment is required for that contribution.
Residual issue, not a release gate: the domain’s forty-five-minute clinic describes three formats and four roles. Treat it as an optional facilitated extension. Its proposed duration is not evidence that four busy collaborators will finish three formats in that time. The short paired route is now independently usable, so that ambitious clinic no longer blocks first contact.
Biology newcomer: “A batch-confounded table should not become a clinical claim”
The revised chapter accurately separates observed contrast, exact additive-model coefficient and causation. Yet a novice encountering the human path had to locate the actual four readings elsewhere. The facilitator’s amendment anchor said to issue a “new, narrower record” after crossed evidence, without saying narrower than what. In fact, the newly supported model scope can be broader than the old arithmetic-only scope while remaining much narrower than causal validation.
Repair made: the starter presents the four values and a ready sentence retaining treatment/batch ambiguity. The facilitator now explicitly limits the amended record to the stated additive-model calculation. The linked chapter remains the source for the interaction counterexample and causal countermodel.
What this reader can now do: compute control mean 10 and treated mean 12, state that batch changes with treatment, and decline an unsupported causal upgrade. No real assay, patient information, replication experiment or statistical software is required. The novice is not asked to settle the identification problem by confidence or consensus.
Residual issue: the outreach draft proposes testing whether communication is easy but labels the encounter a methods discussion. That is a reasonable proposed teaching purpose, not an automatic exemption from human-research review if the organizer later collects systematic participant evidence. The facilitator guide already requires a human-led decision before such a study. Preserve that distinction when turning draft invitations into a real event.
Economics/social-science newcomer: “Why am I filling nine fields to point out one denominator?”
The chapter offers several worthwhile entry points, but a first reader could think they must master transport, selection, routing and the incentive model before contributing. The nine-field contribution card looked mandatory even when the observation was simply that a completed-review statistic omitted unresolved work. It also required a destination that has deliberately not yet been selected for public release.
Repair made: the starter isolates two weighted sums and their limitation. The minimum contribution is explicitly just the example/version, what was tried, and what confused or failed. The longer card is optional. While no feedback route is announced, readers may retain their note locally or discuss it with a colleague; they need not find a personal address or post sensitive material publicly.
What this reader can now do: compute 1.6 and 0.4 and explain why their different weights change the modeled population. A skeptical reader can instead point to one missing denominator, without deriving the audit equilibrium or writing code. The richer model suite is available to deepen the contribution, not an entrance exam.
Residual issue: equal group attention and equal per-offer coverage remain different normative objectives. The chapter and ecology manuscript disclose this. Do not turn an onboarding completion or a favorable toy metric into an earned scientific standing or a claim of fair opportunity.
Law newcomer: “Whose deadline is this, and am I being asked for legal advice?”
The law case carefully distinguishes fictional timing, federal evidence rules, conditional statutory procedure and unsupported operational authority. The human path nevertheless offered no concrete task small enough to begin without reading those legal references. A reader might assume they were being invited to assess a real complaint.
Repair made: the starter uses only two explicitly invented jobs, both available now. Its first sentence states that B then A meets those deadlines while excluding actual legal obligations. The facilitator links the domain case and permits fictional roles and private responses. The first contribution can be completed without client facts or professional authority.
What this reader can now do: explain why total available effort is not a deadline guarantee, or point out that citation agreement does not confer decision power. Nobody needs to offer a legal opinion about an actual person.
Small publication edit suggested to the integrator: the law invitation says “what information a affected party must receive”; change “a” to “an.” This is an editorial typo, not a substantive release blocker.
Cross-domain repairs and before/after record
| Before | After in the onboarding documents | Purpose |
|---|---|---|
| Blank card before any example | Four linked microtasks and one fully worked card | Make the first action executable by a novice |
| Page-and-colleague opening could imply a partner requirement | Explicit solo first contact, pair needed only later | Avoid an unnecessary participation dependency |
| Four implied roles within a paired exercise | Two-person fictional-role mapping; optional facilitator | Avoid silently requiring a team or uncounted labor |
| Nine-field submission seemed mandatory | Three-item minimum; fuller card optional; local-note fallback | Permit a small contribution before public intake exists |
| Facilitator always gives prose first | Teaching order distinguished from counterbalanced comparisons | Avoid confounding format with learning order |
| Facilitator “rewards” an observation | Acknowledgment explicitly separated from rank/endorsement | Avoid recreating a status ladder through the exercise |
| Biology amendment called merely “narrower” | Exact stated additive-model scope named | Prevent an accidental causal interpretation |
No tests of human efficacy were added or claimed. The changes are reversible prose repairs and direct links, not a new mandatory rule table.
Publication and dissemination judgment
The public-seed recommendation remains appropriate. A static introduction, versioned source packet and a clearly bounded feedback invitation do not require a fully functioning appeals institution. An actual release still needs an accountable publisher, rights disposition, accurate AI contribution disclosure, the final exact bytes and a destination. None of these is supplied by an agent’s positive review. An optional maintainer routine should not become a prerequisite for someone reading or adapting a synthetic example.
I independently retrieved three primary dissemination pages during this pass:
- FORCE11 Upstream guidelines support an original open-research perspective, not copying a product-style launch announcement. Its editorial workflow and post-publication discussion require real author attention. No acceptance or eligibility ruling was obtained.
- RDA Reproducibility Interest Group describes an audience concerned with data/code reproducibility and bridges to other efforts. Its historical activity listings do not establish an available current speaking slot. The packet correctly treats a methods discussion as a possibility to investigate, not an endorsement or scheduled event.
- Zenodo upload guidance distinguishes reserving a DOI from registering it on publication. A draft metadata record is not a publicly released artifact or peer review.
The previous independently verified OSF AI-content exclusion remains decisive for that route. No outreach was sent, accounts created, venue acceptance implied or archive published during this pass. Human-owned publication plus small domain discussions is a reasonable first sequence. Subsequent scholarly submissions should be chosen for the actual contribution and their then-current policies.
Verdict: the repaired human path now supports a concrete first action for each domain without coding, institutional status, a new collaborator or a large form. This is a reasoned design assessment, not measured usability. Publish the curated prototype with those limits and invite evidence that this assessment is wrong.