Jon W. Hansen, FCIPS · Procurement Insights | Hansen Models™ · 15 October 2026
There are three distinct control functions in AI governance, and most current frameworks concentrate on only two. A guardrail limits what an agent may do. A harness governs how it does it. A firm fence governs whether the organization understands what the agent is acting on. The test for a firm fence is simple: can evidence from real outcomes stop an action and change the explanation behind it? If it can’t, there is no fence, whatever the controls say.
The foundation for the fence is built in Phase 0™: the trace, before anything is dispatched, that establishes how the operation actually runs, which outcomes matter, what people are measured on, where the work crosses from one function to another, how exceptions are handled, every shadow operation within the enterprise that never appears on the official map, and who can stop an action. The fence then operates continuously, using real outcomes to test whether that original understanding still holds. Without the Phase 0™ work, there is nothing for the fence to stand on.
Many organizations skipped it. Agents were dispatched first, because the technology was ready and the pressure to deploy was real. Now those organizations are playing catch-up, adding guardrails and harnesses to systems already in flight. Those controls are real engineering progress. But they work inside a frame, and nobody has checked whether the frame is right.
This week, a lab in a completely new environment showed why that order fails, and what the fence does that nothing else can.
What the Isaias Strand Lab did well
On 8 October, before Hurricane Isaias made landfall on the Gulf Coast, Hansen AI Labs™ put a question on the public record: a refinery can come through a hurricane undamaged and still stop shipping. Do you know which connections your fuel supply depends on?
The field was energy, an area in which the lab had no prior expertise. The method was the one developed in 1998 for a Department of National Defence procurement initiative: trace the dependencies before the event, and record what you expect before the outcome is known.
Several things went right.
It asked the dependency question while there was still time to act on it. Weather services were forecasting the storm’s path. The lab asked which connections the storm would cross. That question was available to every buyer, operator and risk officer on the day before landfall.
It named the right strands. On 8 October, the lab identified policy and prices as strands to watch alongside the storm. That is why, when a presidential diesel deal surfaced days later, it was recognized as part of the hurricane story rather than filed as a political headline.
Its structural observations held. Power was the strand everyone could watch, and it recovered first. Recovery estimates ran optimistic: the region’s largest utility cut its own restoration target the day it was due. The strands that actually decide whether fuel moves, such as offshore production, refinery status and port reopenings, were the hardest to see and the slowest to appear in the public record.
Its record was built to be checked. The dependency question and the predictions were committed on 8 October, before landfall, and six of the seven were on the public record that day. When the lab opened on 10 October, the baseline was locked and fingerprinted as those predictions stood on 8 October, so that every later finding and correction could only be appended. Every daily entry was appended, never revised. Every correction was dated and added in the open, with the original wording left visible.
And when the lab got something wrong, the fence caught it. The fence was the combination the method is designed around: named strands that kept the right questions open and brought the evidence forward, and a Human Orchestrator with the authority to act on it. That is the part that matters most for every organization dispatching AI.
What the fence caught
The lab’s price strand recorded diesel rising 4 cents after landfall. The figure was well formed, sourced, attributed and internally consistent. It moved from a session transcript into the lab record, then into a handover note, then into the draft of a post. Every check along the way passed it. The locked baseline passed it, because the baseline was untouched. The append-only rule passed it, because it was appended correctly. Four independent AI models simultaneously reviewed the draft, working in conjunction with a fifth, the anchor model. None challenged the number.
It was wrong. The official federal survey, released a day later, showed diesel down 2 cents after landfall, not up 4. Behind that number was a policy event: on the afternoon of landfall day, hours before the storm came ashore, a presidential announcement of a Russian diesel supply deal knocked more than 4 percent off wholesale diesel in a single session.
Policy was one of the strands the lab identified on 8 October, before landfall, at a time when storm coverage was watching the track, the wind and the power outages. The daily searches had been looking for the forms of policy the predictions expected, such as waivers and emergency orders, so the deal did not enter the daily record on the day it was announced. But because policy was already a named strand, the Human Orchestrator’s first question when the deal surfaced was whether it had moved diesel prices. It had, and that answer changed how the lab read the post-landfall price.
What stopped the wrong number was not an automated check. It was the Human Orchestrator comparing the draft against what the lab had already published, recognizing that both could not be true, and refusing to let the contradiction rest. Evidence stopped an action, and the explanation behind it changed. That is the firm fence test, met. What the lab had not yet built was the automatic contradiction check that would have flagged it first. That is the next layer of the fence, and its job is to relieve the Human Orchestrator, not to replace them.
Notice what happened. Every guardrail and every harness worked exactly as designed, and the error got through anyway. Those controls govern how the record is kept. They are not built to ask whether the number is a correct reading of the world. That is the fence’s job. In this lab, the fence was the named strands and the Human Orchestrator working together.
That is a frame failure, not an execution failure. Execution failures are fast and visible, and guardrails are built to stop them. Frame failures can run for years, because every check along the way confirms that the system is working.
Why Phase 0 cannot be bypassed
“Your ‘Phase 0’ framing is exactly right. Before choosing anything, organisations must confront whether they are structurally capable of absorbing it.” — Phil Fersht, Founder, HFS Research
An organization that starts with guardrails has decided what to protect against before it understands what it is operating. It can make every agent perfectly well behaved inside a wrong understanding of the work.
The lab showed both halves of this. Where the dependencies had been traced, its observations held in a field it had never worked in. Where the instruments had not been traced, where nobody had written down that one price survey is daily and another weekly, that retail is quoted per gallon and crude per barrel, and that policy can arrive as a supply deal as easily as a waiver, the errors would have gone straight through unchallenged.
The work that would have prevented them is Phase 0™ work. It cannot be added later as a control, because it is what tells you which controls you need.
The firm fence rules
These rules are drawn from this 2026 lab, from the 1998 system where the method was first applied in the real world over a seven-year period, and from the errors the lab made (and caught). They are written for any organization deploying AI, including those whose agents are already running.
1. Trace before you dispatch, and if already dispatched, trace now. No agent should be dispatched into an operation until that operation has been traced: how its work is actually timed, what its people are measured on, where the work changes hands, and the shadow operations that never appear on the official map. Where agents are already running, start the trace with the agent whose failure would cost the most.
2. Name the outcome, not the activity, and give it an owner. Every agent serves a stated operational result: the part arrived, the claim was paid correctly, the patient was seen. That result is measured, and a named person owns it. An agent measured on its own activity, such as tickets closed or orders placed, can look successful while the outcome fails. In the 1998 DND operation, a part could ship on time and meet the delivery SLA, yet the service call still failed to close on the back-end confirmation loop, because the wrong part had been ordered or the part had a quality issue. Measured on delivery alone, the order was a success. Measured on the outcome, it was a failure that was recognized and fixed.
3. Close the loop from outcomes back to the explanation. Real results return on a set schedule to whoever governs the agent’s design, not just to a dashboard. If outcomes aren’t measured and fed back, the agent isn’t governed, no matter how many guardrails it has.
4. Give a named person the authority to stop it. At each point where an agent’s action crosses another part of the operation, someone can stop it when experience says the data is wrong. They don’t need approval to stop it, and every stop is logged. The same design specifies who may authorize a restart and what evidence is required before the action resumes.
5. When an outcome materially contradicts the plan, pause and rethink; don’t patch. Not every variance is a contradiction; operations have noise. But when results materially contradict what the deployment was built on, the agent pauses at that point until the explanation is revised. Adding another guardrail around a wrong assumption is the catch-up cycle many organizations are in now.
6. Check for other causes before giving credit or blame. Before attributing a success or failure to the agent, ask what else changed in the same period. Without that check, organizations learn the wrong lesson from their own results. The lab came close. Diesel prices dipped slightly after landfall, and the first explanation offered was that the storm’s crossings had held. But hours before the storm came ashore, a Trump–Russia diesel supply deal had been announced. Because policy was already a named strand, the Human Orchestrator asked whether the deal had moved the price. It had.
7. Monitor every material crossing, and keep the map open. Every crossing the trace identifies is monitored for change, including events from outside the organization, and there is a way for an unexpected cause to enter the record. The lab learned this directly: the strand was named, but the event that mattered arrived in a form nobody had listed. This is Strand Commonality™: parts of an operation that seem unrelated are connected, so a change in one can impact another that was working.
8. Don’t let the sponsor grade the deployment alone. Whoever championed the agent cannot be the only judge of whether it worked. Results are reviewed by someone with no stake in the answer.
9. Keep a dated record that’s added to, never rewritten. Decisions, stops, contradictions and revisions are recorded as they happen, so the organization can see what it knew, and when. An append-only record preserves provenance, not truth, so every correction is linked to the claim it corrects and clearly marked as superseding it.
Three questions for every board
Has management traced the operation this agent works in, including the shadow operations that never appear on the official map?
Who can stop it, what triggers the stop, who can authorize the restart, and has it ever been used?
When results disagreed with the plan, what changed: the explanation, or only the controls?
If management cannot answer all three, there is no fence, whatever the controls say.
These rules rest on a system that ran in a live operation for seven years, and on a lab that tested the same method in a field it had never worked in. They are offered the way every finding in this work is: on the record, open to challenge, and openly revised and transparently updated when the evidence says so.
The genie is out of the bottle, and the agents are already in flight. The fence can still be built. But it has to be built from the ground up, starting with Phase 0™.
Phase 0 Can Be Delayed. It Cannot Be Escaped: The Firm Fence Rules AI Deployments Keep Skipping
Posted on October 15, 2026
0
Jon W. Hansen, FCIPS · Procurement Insights | Hansen Models™ · 15 October 2026
There are three distinct control functions in AI governance, and most current frameworks concentrate on only two. A guardrail limits what an agent may do. A harness governs how it does it. A firm fence governs whether the organization understands what the agent is acting on. The test for a firm fence is simple: can evidence from real outcomes stop an action and change the explanation behind it? If it can’t, there is no fence, whatever the controls say.
The foundation for the fence is built in Phase 0™: the trace, before anything is dispatched, that establishes how the operation actually runs, which outcomes matter, what people are measured on, where the work crosses from one function to another, how exceptions are handled, every shadow operation within the enterprise that never appears on the official map, and who can stop an action. The fence then operates continuously, using real outcomes to test whether that original understanding still holds. Without the Phase 0™ work, there is nothing for the fence to stand on.
Many organizations skipped it. Agents were dispatched first, because the technology was ready and the pressure to deploy was real. Now those organizations are playing catch-up, adding guardrails and harnesses to systems already in flight. Those controls are real engineering progress. But they work inside a frame, and nobody has checked whether the frame is right.
This week, a lab in a completely new environment showed why that order fails, and what the fence does that nothing else can.
What the Isaias Strand Lab did well
On 8 October, before Hurricane Isaias made landfall on the Gulf Coast, Hansen AI Labs™ put a question on the public record: a refinery can come through a hurricane undamaged and still stop shipping. Do you know which connections your fuel supply depends on?
The field was energy, an area in which the lab had no prior expertise. The method was the one developed in 1998 for a Department of National Defence procurement initiative: trace the dependencies before the event, and record what you expect before the outcome is known.
Several things went right.
It asked the dependency question while there was still time to act on it. Weather services were forecasting the storm’s path. The lab asked which connections the storm would cross. That question was available to every buyer, operator and risk officer on the day before landfall.
It named the right strands. On 8 October, the lab identified policy and prices as strands to watch alongside the storm. That is why, when a presidential diesel deal surfaced days later, it was recognized as part of the hurricane story rather than filed as a political headline.
Its structural observations held. Power was the strand everyone could watch, and it recovered first. Recovery estimates ran optimistic: the region’s largest utility cut its own restoration target the day it was due. The strands that actually decide whether fuel moves, such as offshore production, refinery status and port reopenings, were the hardest to see and the slowest to appear in the public record.
Its record was built to be checked. The dependency question and the predictions were committed on 8 October, before landfall, and six of the seven were on the public record that day. When the lab opened on 10 October, the baseline was locked and fingerprinted as those predictions stood on 8 October, so that every later finding and correction could only be appended. Every daily entry was appended, never revised. Every correction was dated and added in the open, with the original wording left visible.
And when the lab got something wrong, the fence caught it. The fence was the combination the method is designed around: named strands that kept the right questions open and brought the evidence forward, and a Human Orchestrator with the authority to act on it. That is the part that matters most for every organization dispatching AI.
What the fence caught
The lab’s price strand recorded diesel rising 4 cents after landfall. The figure was well formed, sourced, attributed and internally consistent. It moved from a session transcript into the lab record, then into a handover note, then into the draft of a post. Every check along the way passed it. The locked baseline passed it, because the baseline was untouched. The append-only rule passed it, because it was appended correctly. Four independent AI models simultaneously reviewed the draft, working in conjunction with a fifth, the anchor model. None challenged the number.
It was wrong. The official federal survey, released a day later, showed diesel down 2 cents after landfall, not up 4. Behind that number was a policy event: on the afternoon of landfall day, hours before the storm came ashore, a presidential announcement of a Russian diesel supply deal knocked more than 4 percent off wholesale diesel in a single session.
Policy was one of the strands the lab identified on 8 October, before landfall, at a time when storm coverage was watching the track, the wind and the power outages. The daily searches had been looking for the forms of policy the predictions expected, such as waivers and emergency orders, so the deal did not enter the daily record on the day it was announced. But because policy was already a named strand, the Human Orchestrator’s first question when the deal surfaced was whether it had moved diesel prices. It had, and that answer changed how the lab read the post-landfall price.
What stopped the wrong number was not an automated check. It was the Human Orchestrator comparing the draft against what the lab had already published, recognizing that both could not be true, and refusing to let the contradiction rest. Evidence stopped an action, and the explanation behind it changed. That is the firm fence test, met. What the lab had not yet built was the automatic contradiction check that would have flagged it first. That is the next layer of the fence, and its job is to relieve the Human Orchestrator, not to replace them.
Notice what happened. Every guardrail and every harness worked exactly as designed, and the error got through anyway. Those controls govern how the record is kept. They are not built to ask whether the number is a correct reading of the world. That is the fence’s job. In this lab, the fence was the named strands and the Human Orchestrator working together.
That is a frame failure, not an execution failure. Execution failures are fast and visible, and guardrails are built to stop them. Frame failures can run for years, because every check along the way confirms that the system is working.
Why Phase 0 cannot be bypassed
An organization that starts with guardrails has decided what to protect against before it understands what it is operating. It can make every agent perfectly well behaved inside a wrong understanding of the work.
The lab showed both halves of this. Where the dependencies had been traced, its observations held in a field it had never worked in. Where the instruments had not been traced, where nobody had written down that one price survey is daily and another weekly, that retail is quoted per gallon and crude per barrel, and that policy can arrive as a supply deal as easily as a waiver, the errors would have gone straight through unchallenged.
The work that would have prevented them is Phase 0™ work. It cannot be added later as a control, because it is what tells you which controls you need.
The firm fence rules
These rules are drawn from this 2026 lab, from the 1998 system where the method was first applied in the real world over a seven-year period, and from the errors the lab made (and caught). They are written for any organization deploying AI, including those whose agents are already running.
1. Trace before you dispatch, and if already dispatched, trace now. No agent should be dispatched into an operation until that operation has been traced: how its work is actually timed, what its people are measured on, where the work changes hands, and the shadow operations that never appear on the official map. Where agents are already running, start the trace with the agent whose failure would cost the most.
2. Name the outcome, not the activity, and give it an owner. Every agent serves a stated operational result: the part arrived, the claim was paid correctly, the patient was seen. That result is measured, and a named person owns it. An agent measured on its own activity, such as tickets closed or orders placed, can look successful while the outcome fails. In the 1998 DND operation, a part could ship on time and meet the delivery SLA, yet the service call still failed to close on the back-end confirmation loop, because the wrong part had been ordered or the part had a quality issue. Measured on delivery alone, the order was a success. Measured on the outcome, it was a failure that was recognized and fixed.
3. Close the loop from outcomes back to the explanation. Real results return on a set schedule to whoever governs the agent’s design, not just to a dashboard. If outcomes aren’t measured and fed back, the agent isn’t governed, no matter how many guardrails it has.
4. Give a named person the authority to stop it. At each point where an agent’s action crosses another part of the operation, someone can stop it when experience says the data is wrong. They don’t need approval to stop it, and every stop is logged. The same design specifies who may authorize a restart and what evidence is required before the action resumes.
5. When an outcome materially contradicts the plan, pause and rethink; don’t patch. Not every variance is a contradiction; operations have noise. But when results materially contradict what the deployment was built on, the agent pauses at that point until the explanation is revised. Adding another guardrail around a wrong assumption is the catch-up cycle many organizations are in now.
6. Check for other causes before giving credit or blame. Before attributing a success or failure to the agent, ask what else changed in the same period. Without that check, organizations learn the wrong lesson from their own results. The lab came close. Diesel prices dipped slightly after landfall, and the first explanation offered was that the storm’s crossings had held. But hours before the storm came ashore, a Trump–Russia diesel supply deal had been announced. Because policy was already a named strand, the Human Orchestrator asked whether the deal had moved the price. It had.
7. Monitor every material crossing, and keep the map open. Every crossing the trace identifies is monitored for change, including events from outside the organization, and there is a way for an unexpected cause to enter the record. The lab learned this directly: the strand was named, but the event that mattered arrived in a form nobody had listed. This is Strand Commonality™: parts of an operation that seem unrelated are connected, so a change in one can impact another that was working.
8. Don’t let the sponsor grade the deployment alone. Whoever championed the agent cannot be the only judge of whether it worked. Results are reviewed by someone with no stake in the answer.
9. Keep a dated record that’s added to, never rewritten. Decisions, stops, contradictions and revisions are recorded as they happen, so the organization can see what it knew, and when. An append-only record preserves provenance, not truth, so every correction is linked to the claim it corrects and clearly marked as superseding it.
Three questions for every board
If management cannot answer all three, there is no fence, whatever the controls say.
These rules rest on a system that ran in a live operation for seven years, and on a lab that tested the same method in a field it had never worked in. They are offered the way every finding in this work is: on the record, open to challenge, and openly revised and transparently updated when the evidence says so.
The genie is out of the bottle, and the agents are already in flight. The fence can still be built. But it has to be built from the ground up, starting with Phase 0™.
The external record:
Related reading:
Truth Is Believing. Accuracy Is Knowing. Outcome Is Proof.™
-30-
Share this:
Like this:
Related