Human Gates and Number Grades
Human gates and number grades are the two governing devices of a planning agent team: a grade on every number that says how it is known, and a gate at every transition where a plan would change, at which a named person signs. The page's claim is that the two devices are what make "agents propose, humans sign" a mechanism rather than a slogan, and that both must be enforced in the files the team reads and writes, because a rule enforced in policy is followed until the day it is inconvenient. This page gives the number grades, reconciles them with the wiki's claim grades, lists the five gates and who holds each, and states the enforcement rule.
Number grades
| Grade | Meaning | What it takes to hold it | Example |
|---|---|---|---|
| [M] Measured | Read from a versioned ledger, definition cited | The ledger version and the definition record | Voice handled on 3 March: 4,210 [M], demand ledger 2026-03-04.v1, definition HND-01 |
| [C] Computed | Produced by a stated formula from graded inputs; inherits the weakest input grade | The formula and every input's grade | Hours miss: 92 percent handle time [C] from [M] inputs |
| [E] Estimated | A range and the assumption behind it | The range, the assumption, the source of the assumption | Shrinkage 30 percent [E], range 28 to 32, from the prior year's supply ledger |
| [A] Asserted | One source, no producible artifact; never load-bearing alone | The source and the date | Pre-migration handle time 412 seconds [A], "carried from the legacy plan" |
Three rules make the scale useful. A computed figure is never better than its weakest input: an FTE figure built on [M] volume and [E] shrinkage is [E]. Nothing carried across a channel, platform or cohort change is [M] in the new state, however well it was measured in the old one; it is [A] until the new state measures it, which is the lesson Asynchronous-to-Synchronous Channel Conversion draws for carried forecasts and the reason the series relabels a carried assumption rather than deleting it. And an [A] figure may inform a proposal but may not by itself carry a plan; the evaluator blocks a plan version whose requirement rests on an [A] input with no [E] range beside it.
Reconciling number grades with claim grades
The wiki grades claims with four words, on Data Synthesis Before Decision and Executive Issue Register: Established, Inferred, Asserted, Open. This series grades numbers with four brackets. They are different objects and the series keeps both, because a claim and the numbers inside it are known differently: "the partner cohort's handle time is structurally higher" is a claim at Inferred; the 470 seconds inside it is [M].
| Number grade | Nearest claim grade | Relationship |
|---|---|---|
| [M] | Established | A measured number on a cited definition is what makes a claim Established |
| [C] | Established or Inferred | Established if every input is [M]; Inferred if any input is [E] or the formula embodies a stated assumption |
| [E] | Inferred | An estimate is a claim that follows from stated reasoning and a range |
| [A] | Asserted | The same word; a number and a claim from one source with no artifact |
| (a blank with an owner) | Open | A number the team does not have is written as blank with who obtains it, not as zero and not as a guess |
The five gates
| Gate | Transition | Who signs | What the signer must produce | Where it is enforced |
|---|---|---|---|---|
| Planner gate | A reforecast version becomes the forecast the platform of record schedules against | The planner who owns the book | Confirmation or amendment of each changed assumption in the register | The forecast adapter refuses an unsigned version |
| Register review | A register row's finding takes its grade; a card is issued; where the answer reveals a risk a leader must act on, the Q row's owner proposes an XR row citing the Q row as its source; the executive register's owner opens it and sets its severity; the Q row's card becomes the XR row's evidence at grade Asserted until an artifact raises it | The planner, or the leader who asked, for the grade; the executive register's owner for the XR row and its severity | The grade, and the "what would change this" line | The reporter cannot publish a card without a grade and that line |
| Plan signature | A capacity plan version becomes the plan of record; moves and requisitions are approved | The planning lead; node owners for their moves | A signature on the version; acceptance or decline with price on each move | The planning-platform adapter refuses an unsigned plan |
| Publication gate | A schedule or a shift bid is published to workers | The scheduler | A signature carrying the reason for any tolerance breach in the check record | The platform's publication step reads the signature block |
| Action gate | A real-time action is executed on the platform | A real-time analyst inside guardrails; a manager outside them | A decision on each proposal, logged with outcome | The platform executes only from a signed proposal row |
The "must produce" column is the page's most important one. Human supervision of reliable automation decays into assent unless the supervisor has to produce something, an amended figure, a stated reason, a decision per row, rather than merely approve; the human-factors literature has documented the effect, under the names automation bias and complacency, across two decades of studies.[1][2] A gate that asks for a click is not a gate.
Where the gates sit follows the ordinary rule for placing automation authority by consequence: the automation may go as far as the stage before the one whose error is costliest, and no further, until its reliability is measured.[3] The five gates are the five points in a planning function where the next step changes what the floor does or what the business commits to. The evaluator agent's pass or block is not a gate; it is a check before the gate, and a person may overrule it with a logged reason.
When a gate is released. The action gate is released for a class of action when a catch rate has been published for it, not at a maturity level; the evidence, not the level, is what releases the gate. Level 3: The Automation Layer and Real-Time Schedule Adjustment now state the same condition, resolved on 17 September 2026: released per class of action on a published catch rate, first met by routine micro-moves at Level 3, extended to broader classes at Level 4 and run continuously at Level 5. The other pages of the series point at this sentence rather than naming a level.
Enforced in the format, not the policy
A gate lives in a file format when the downstream step cannot proceed without a field only a person can fill. The forecast version has a signature block: signer, timestamp, version signed. The adapter reads the block before writing the platform's file and writes nothing if it is empty. The same block is on the plan, the schedule, the card and the action proposal. No configuration flag skips the check, and the check's refusals are logged.
This is stricter than a policy that says "planners must approve reforecasts," and the strictness is the point. Policies are followed until the morning the planner is late and the batch job is due; the worked example on Integration Agents shows what the format rule does on that morning. The rule also makes the gate auditable: every signed version names its signer, and every refused crossing is a row. Where a function later wants to lift a gate, for a process whose catch rate has been published per The Agent Overseer, it changes the format, a versioned and reviewed act, rather than instructing people to stop checking.
The same principle governs the grades. A ledger cell without a grade is rejected at load. A computed figure whose formula is not stated is rejected by the evaluator. Neither depends on anyone remembering.
Worked example
Friday 27 March 2026, the series example's plan signature. The capacity planner's decision pack carries the phase 2 plan at about 310 FTE [E]. The planning lead opens the register and finds the [E] rests on shrinkage at 30 percent [E], range 28 to 32, and on handle time at 452 seconds [M]; the comparison scenario at 412 seconds is labeled [A], carried. The pack's "would change this" line names a partner-cohort learning curve by 20 March; the lead checks the register row Q-004, whose settling test ran on 20 March and found no curve, tag structural at Inferred. The lead signs the 310 FTE plan, writes "shrinkage range accepted; revisit at phase 3" in the signature block's reason field, and declines one of the clearing moves, an email-to-voice transfer of 22 FTE, with the recorded price "training lead time exceeds phase 2 date; requisition instead." The adapter exports the signed plan that evening. The declined move and its price are in the plans ledger, and in May, when someone asks why phase 2 was hired rather than moved, the answer is one row.
What would change this
The grade scale and the five gates are drawn from a prototype and rest on the human-factors literature's findings about supervision. The number of gates is a design choice; a function that measured its register review as never changing a grade could fold it into the planner gate. The "must produce" rule is the load-bearing one and would be relaxed only on evidence that a lighter gate caught as much. The two-scale decision would be revisited if a function found its planners confusing the scales in practice.
How this connects
The grades are element 6 of The Agent Team Model and are carried by every cell of Living Ledgers. Each gate is exercised on its own page: the planner gate on The Short-Term Forecasting Loop with an Agent Team, the register review on The Question Register and Knowledge Base, the plan signature on Long-Term Planning Agents and the Plan of Record, the publication gate on Scheduling Agents, the action gate on Real-Time Agents; and Integration Agents enforce the signatures at the boundary. The claim grades belong to Data Synthesis Before Decision. The gates on this page govern what an already-handed-over process may change; the gate that decides whether a process is handed over at all is The Agentic Handover Gate, and the two are not the same.
Maturity Model Position
Grades on numbers and signatures on plans are Level 3 practice on the WFM Labs Maturity Model™ and can be adopted in a spreadsheet function before any agent exists. Gates enforced in file formats, with refusals logged and a catch-rate program that earns their relaxation, is the Level 4 form. Gate-free execution inside guardrails is released per class of action on a published catch rate, as the release condition above records and the wiki's level pages now state; this page places the release on evidence, not on a level.
See Also
- AI Agent Teams for Workforce Management — the series hub
- The Agent Team Model — where the grades and gates sit in the pattern
- Living Ledgers — the cells the grades are carried in
- Integration Agents — the adapters that enforce the signatures
- The Question Register and Knowledge Base — the register review and the card
- Data Synthesis Before Decision — the four claim grades
- The Agent Overseer — the catch-rate program that earns a gate's relaxation
- The Agentic Handover Gate — the different gate that admits a process to agents
References
- ↑ Skitka, L. J., Mosier, K. L., & Burdick, M. (1999). "Does automation bias decision-making?". International Journal of Human-Computer Studies 51(5), 991–1006. doi:10.1006/ijhc.1999.0252.
- ↑ Parasuraman, R., & Riley, V. (1997). "Humans and Automation: Use, Misuse, Disuse, Abuse". Human Factors 39(2), 230–253. doi:10.1518/001872097778543886.
- ↑ Parasuraman, R., Sheridan, T. B., & Wickens, C. D. (2000). "A model for types and levels of human interaction with automation". IEEE Transactions on Systems, Man, and Cybernetics — Part A 30(3), 286–297. doi:10.1109/3468.844354.
