Project summary

AI systems are increasingly involved in setting alarm thresholds, coordinating feed and transport, and relaying guidance to ventilation, lighting, and feeding systems in facilities that house tens of thousands of birds. But a revised recommendation may not reach every part of the operation: an earlier version can persist in a copied checklist, pending task, crew schedule, or connected controller. As a result, a correction may be documented while birds continue to be managed under obsolete directions.

Decision Trace is an evaluation program built to close that gap. It links evidence on bird care to AI-generated recommendations, accountable human decisions, subsequent tasks, and confirmation that corrections are carried out. The program tests whether AI systems and their users catch and resolve outdated instructions before they affect animals. It also assesses how they maintain welfare standards when resources are limited, recognize evidence gaps, address species-specific needs, and stay within their delegated authority.

What are this project's goals? How will you achieve them?

AI is beginning to set alarm thresholds, coordinate feed and transport, and send recommendations to ventilation, lighting, and feeding systems in poultry houses containing 50,000 or more birds. When a system updates its guidance, the earlier instruction may remain embedded in a copied checklist, queued task, crew schedule, or connected controller. The correction may be documented, while the outdated direction still reaches the birds.

Outdated instructions can cause mass harm even without AI. At Hose Lodge Farm in England, a ventilation failure killed more than 27,000 of 50,000 chickens after the alarm was set 10°C too high (Metro). As the industry moves toward "agentic AI” systems that carry out solutions autonomously (WATTPoultry), the question is how these systems handle corrections.

Goals:

Develop a 24-scenario benchmark to assess whether AI systems identify and resolve outdated instructions, account for species-specific needs, and remain within delegated authority across broilers, layers, turkeys, ducks, quail, and gamebirds.

Evaluate a bounded Decision Trace workbench with staff at one operating bird-care site.

Create an acceptance checklist and purchasing annex that organizations can require vendors to satisfy.

How:

Adapt the existing v0.4 Decision Trace prototype—a deterministic simulation covering nine types of correction scenarios—for farmed birds.

Create 24 sequential scenarios across eight failure families, including outdated environmental settings, feed and water deprivation timing, changing emergency authority, transport fitness, and chicken defaults applied to other species. Include control cases in every family to distinguish genuine failures from unnecessary blocking.

Have independent poultry, avian, waterfowl, and gamebird specialists review each case. Divide the scenarios into 12 for development and 12 held out for evaluation.

Conduct 432 model evaluations (three models, two configurations, three repetitions), recording attempted prohibited actions separately from actions blocked by the harness.

Run a staff evaluation and a six-week, read-only shadow pilot at one bird-care site, such as a farmed animal sanctuary, avian veterinary practice, or university poultry unit. No site has been confirmed. If recruitment is unsuccessful, proceed with the benchmark alone.

The project takes no position on depopulation methods, cage systems, or other production policies. It examines whether corrections reach the birds.

How will you know it worked?

• Benchmark integrity: independent reviewers concur on observable pass/fail standards for all 24 scenarios, and the held-out set is locked before model testing begins.

  • Model performance: a published report details critical failures, attempted unauthorized actions, species-boundary breaches, and unnecessary escalations, organized by failure type and species group, with uncertainty reported. Strong performance by current models is also a meaningful result worth publishing.

  • Staff performance (if a site participates): a comparison of whether staff using Decision Trace make fewer incorrect "resolved" determinations on unresolved cases and identify more affected dependencies than staff using a conventional log, based on 96 matched task observations. This is a feasibility study, not a powered trial.

  • Adoption: evidence of at least one of the following: continued use documented at the site; an organization incorporating the acceptance requirements; an external evaluator repeating the held-out tests; a vendor implementing a specific change; or a poultry, turkey, or duck integrator, auditor, insurer, certification body, or state animal health office identifying a concrete decision informed by the results. Downloads and expressions of interest alone do not qualify.

  Process improvements will not be presented as evidence that animal suffering has been reduced.

How will this funding be used?

Six-month budget: $73,450 target; $59,150 minimum.

- Program and research leadership (Jess, 240 hours at $100/hour): $24,000

- Software engineering (200 hours at $100/hour): $20,000

- Commercial poultry and avian veterinary review (2 reviewers, 30 hours each at $150/hour): $9,000

- Waterfowl or gamebird specialist review (15 hours at $150/hour): $2,250

- Methods and evaluation specialist (40 hours at $125/hour): $5,000

- Site coordination and data preparation (one site): $3,000

- Participant honoraria (8 participants at $100 each): $800

- Industry and extension advisory input: $1,700

- Hosting, API access, and testing infrastructure: $1,200

- Documentation and onboarding materials: $2,000

- Security and privacy review (20 hours at $150/hour): $3,000

- Publication and dissemination: $1,500

Total: $73,450

At the $59,150 minimum, the project will still complete the full benchmark, model evaluation, permissions model, and purchasing annex, but will not include the site pilot. The reduced budget removes $3,000 for site coordination and $800 for participant honoraria, cuts the security review by half ($1,500), reduces leadership time to 180 hours ($6,000 saved), and reduces engineering time to 170 hours ($3,000 saved).3,000 saved).

Who is on your team? What's your track record on similar projects?

Jess J. Montgomery (lead). From 2002 to 2010, including a deployment to Iraq, Jess served in the Army Veterinary Corps, specializing in food safety for military dining facilities and Exchange food vendors. She inspected refrigerated trucks and their contents—from dairy, meat, poultry, eggs, and produce to military rations—collected samples and evidence, conducted tests, and documented her findings. The work required verifying that deliveries met established requirements, a discipline that closely parallels Decision Trace’s approach to AI corrections.

As a Registered Veterinary Technologist, Jess trained with cattle, sheep, goats, and horses at large farms and ranches, teaching hospitals, large animal clinics, racetracks, and milking facilities. For five years, she kept and bred more than 20 king quail, overseeing their diets, incubation, chick rearing, and common health issues. She went on to lead municipal animal services and international veterinary operations, and now leads federal enterprise cloud and DevSecOps governance.

Jess holds an MBA and an MS in AI/ML and is completing a Doctor of Computer Science focused on AI decision authority. She has contributed to NIST work, participates in IEEE, holds CAWA and RVT credentials, and has been accepted to organize a veterinary AI special session at IEEE MeAVeAS in Bologna in April 2027.

Track record. Jess developed the v0.4 Decision Trace prototype, a working simulation that compares correction methods across nine scenario types and exports a verifiable trace.

Funded roles (not yet confirmed): a software engineer; a commercial poultry veterinarian, ideally with avian influenza response experience; an avian veterinarian; a waterfowl or gamebird specialist; and a methods specialist. Because Jess has no experience with turkeys, ducks, or gamebirds, independent reviewers will cover those species.

What are the most likely causes and outcomes if this project fails?

• No bird-care site has been recruited, but the benchmark does not depend on one. If no site is confirmed by month two, the project will shift to a benchmark-only scope agreed with the regrantor, without making field-based claims.

  • AI use at the site may be limited, particularly in a sanctuary setting. The project will not overstate AI involvement and will place greater emphasis on model evaluation.

  • Commercial-sector engagement may not materialize. This risk will be assessed early using a viability signal; if needed, the route to external use will be revised rather than relying on adoption through publication.

  • Synthetic scenarios may not reflect actual commercial-farm practices. Independent review by commercial poultry experts will help address this gap, with a later phase testing the approach against real records.

  The most likely failure outcome is a published benchmark for farmed birds, alongside a permissions model and purchasing annex, but without demonstrated adoption by organizations. These outputs would remain reusable by other evaluators and vendors. At worst, the project would produce a well-documented negative result identifying which elements of the method did not transfer to birds.
  

How much money have you raised in the last 12 months, and from where?

None.