# ASTRA: three prospective workflow pilots

Status: prepared, not run. No participants, reviews, adoption or payment results are asserted.
Owner: ASTRA founder. Participant and qualified reviewer assignments pending.
Purpose: assess one workflow before expanding deployment or features: question → relevant evidence → checked claims → reviewed conclusion → saved/reopened record.

## Shared design — agree before collecting results
Recruit three consenting small teams with recurring evidence-review work. Use pseudonymous participant codes. Each pilot uses at least two matched task pairs; alternate ASTRA-first and normal-workflow-first order. Match domain, complexity and available source access; avoid repeating an identical question when recall would advantage the second tool. Record baseline familiarity, training, assistance, interruptions and deviations. Six pairs are a feasibility target, not a powered effectiveness study.

Use the team's existing tools as the comparator, with equal source access and time budgets. Freeze an independently assembled reference set and inclusion criteria before grading outputs; include eligible challenging studies. Have a qualified reviewer score de-identified output where practical, recording conflicts and any broken blinding. Preserve the original set and append justified corrections; score both tools against the same version. Reference sets are fallible and incomplete.

Store the prospective protocol in the validation lab before data collection and export a dated copy. Do not claim formal preregistration unless deposited with an appropriate external registry. Link the decision and artifact IDs. Every attempt, including abandoned, assisted and failed attempts, belongs in the record.

## Pilot 1 — find and check evidence
Question: a real literature question supplied by the team. The illustrative thermal-storage example can be used for facilitator training, not counted as unseen validation.
Task: retrieve relevant records, inspect original sources, record passages/locations and limitations, identify supporting and challenging evidence, and write a bounded conclusion.
Primary observations: incorrect/unsupported claim–citation pairs divided by pairs checked; missed eligible challenging studies divided by the frozen reference-set total; person-minutes and failed searches.
Stop: a fabricated source, concealed provider failure, or unsupported conclusion represented as reviewed. Fix and run a new documented attempt; preserve the original failure.

## Pilot 2 — hand off and challenge a conclusion
Question: a second matched real decision from the same customer workflow.
Task: a colleague reconstructs the reasoning without coaching, checks cited originals, records an objection, and inspects the author's response and remaining uncertainty.
Primary observations: citation errors, missed counterevidence, reviewer usefulness (1 unusable, 2 extensive rework, 3 substantial rework, 4 minor rework, 5 usable without rework), unanswered reconstruction questions, reviewer/correction person-minutes.
Stop: author response mistaken for reviewer resolution, undisclosed access to private material, or lost original objections. A different account alone does not prove independence.

## Pilot 3 — reopen and reuse for a later decision
Question: a genuine follow-up task, with the old decision and source cutoff preserved.
Task: explicitly save, reload and reopen the investigation; export its source/claim/review record; check an authorized colleague's access and an unauthorized test account's denial; recover from an interrupted save without losing the prior version.
Primary observations: preservation of question, source IDs, claim text, citations, review status and uncertainty; observed return for a second task; time to reconstruct/revise the decision. Test access with consented disposable records, not another person's private data.
Stop: any account data leakage, silent overwrite, false save confirmation or missing record. Do not expand recruitment until repaired and retested.

## Scoring and reporting
Use the paired-trial form for time, quality, generic errors, citation denominators/errors, counterevidence denominators/misses and reviewer usefulness. Include setup, search, reading, checking, correction, failed searches and reviewer person-minutes once each. Record waiting/overlap separately. Zero checked citations or zero eligible challenging studies means not assessable, not 100% accuracy. Do not manufacture a challenge where none is eligible; document the search and limitations.

Predeclare the quality rubric and acceptable error thresholds with each team. For this feasibility phase, any privacy breach, data loss or fabricated citation blocks expansion. Favorable speed alone does not pass. Report each pair and incomplete attempt, ranges, reviewer disagreement and the selected reference-set limits; do not infer population effects from these pilots. Synthetic software tests and the founder's rehearsals never count as participant results.

## Repeat use and commercial validation
After 7 and 30 days, record actual repeat use, artifacts, prompted versus unprompted return, reasons for stopping, purchasing authority, and delivery/support effort. Follow-up is manual; no automated outreach is enabled.

Discuss a consistently scoped assisted pilot with the responsible buyer. Record the proposed deliverables, price, currency, terms, and timing before measuring willingness to pay. Distinguish interest, stated willingness, an agreed contract, and actual payment. No payment collection or subscription is enabled by this protocol. No results are asserted.

Keep personal details and confidential research out of shared records. Public release requires separate consent and review. Measure the complete usual workflow as comparator, not an artificially limited alternative.

## Decision after three pilots
Continue with this workflow only if participants can complete and reopen it, serious defects are resolved, and evidence suggests recurring value. If people cannot complete the task, repair it. If they do not return, investigate the reason. If they return but access or support is difficult, address those barriers. Findings may support stopping or changing direction.

## Current mission-aligned tasks

Use a live question chosen by each participating team, in any scientific discipline. Thermal-storage losses, grid resilience and water systems are examples, not eligibility restrictions. Match reviewers and evaluation criteria to the actual discipline and method. Evaluate evidence checking, challenge handoff, and record reuse on the same bounded question. Type I civilization is the mission, not a scored pilot outcome. Record repeat use, evidence quality, participation barriers, and actual purchasing decisions.