Task-time evidence and transfer

This guide explains how to estimate human time without turning one study, one vendor case, or one organization into a universal productivity rate.

The unit of comparison is a bounded task or workflow with a countable output and an explicit acceptance threshold. Organization type is a context overlay for law, governance, data, procurement, scale, and controls. It is not the benchmark category.

Keep work mode, architecture, and authority separate

Multiple agents or large scale do not automatically make a system A4. Autonomy is determined by the effects the system may produce without a person, not by the reputation or size of the organization using it.

The registry uses work_mode for the work-sharing choice above. A model, fixed workflow, single agent, or orchestrated team is recorded separately, as is the exact A0 to A4 authority level. A study remains transferable when its task, accepted output, human role, and operating conditions are close enough; organization type alone neither permits nor prevents transfer.

Count human time, machine time, and elapsed time separately

For one eligible case:

local human-work floor = preparation + supervision + verification
                         + corrections + expected exception work

net human time with AI = max(source residual - setup already in that residual,
                             local human-work floor)
                         + extra work absent from both
                         + local amortized setup

net human time saved = baseline human time - net human time with AI

annual net human time saved = net human time saved per case
                              * eligible cases per month * 12

Expected exception work is the exception rate multiplied by human minutes per exception. Amortized setup divides one-off human setup hours across the selected number of months and eligible cases.

The source-implied residual is the human time that remains after applying one measured low, central, or high reduction to the local baseline. Taking the greater value prevents the calculator from hiding declared review work. It also avoids adding the source residual and local work together, which would count human time twice when their scope overlaps. Work absent from both must be added separately. If the source residual includes setup, remove that part before adding local setup. The removal cannot exceed the source residual. With no admitted source, or when you choose your own assumptions, use the local breakdown alone.

When the eligible share is zero, the net range is unavailable. There is no eligible case over which to allocate setup, so the calculator reports n/a instead of silently dropping the fixed cost. The copied pilot brief records the task profile, operating mode, output state, operator experience, human-work components, exception assumption, setup hours, amortization horizon, evidence identifier, and net method needed to reproduce the estimate.

The result may be negative. A slowdown is evidence and must not be replaced by zero. Machine runtime never counts as human time saved. A turnaround moving from three days to one day is an elapsed-time change unless human active minutes were also measured.

Quality, accepted throughput, straight-through completion, incidents, and the downstream outcome remain separate measures.

Evidence grades

The grade describes how the value was measured, not whether it is favorable. The machine-readable registry additionally blocks automatic quantitative use unless human active time was measured and the record is explicitly admitted for transfer.

Transfer states

The calculator compares six parts of the task contract:

1. the target task profile;
2. the work mode;
3. the architecture;
4. the exact A0 to A4 action boundary;
5. the required output state;
6. operator experience.

It returns one of four states:

These are minimum gates, not proof of equivalence. Before using a source, also compare input complexity, output unit, model and tool access, verifiability, quality threshold, exception rate, operator skill, and consequence of error.

What the current records demonstrate

The registry records the exact source, measurement basis, task contract, preconditions, and transfer limits. Current frontier-company examples remain visible because their mechanisms may be useful. Their internal or model-based figures do not enter the calculator automatically.

How to use the interactive calculator

1. Choose the closest countable task, not your organization type.
2. Set the output state and operator experience.
3. Read the evidence card and its transfer state.
4. Enter the manual baseline, monthly volume, and eligible share.
5. Open the human-time breakdown and enter preparation, supervision, verification, correction, exception, and setup time.
6. Read the unadjusted source range and the resulting net low, central, and high range.
7. Replace every planning input with pilot observations before a decision.

Do not add the source percentage or its residual minutes to the local work. The engine uses the adjusted source residual or local human-work floor, whichever is greater, then adds work missing from both and local amortized setup. The source still asks what happened in a comparable measured task. The net range shows what remains after the local controls and implementation effort are made explicit.

How the 11 worked cases are classified

Every worked-case result remains a grade E planning hypothesis. An external anchor can explain a mechanism or frame a separate range, but it never upgrades the synthetic case result.

The registry stores the complete rationale and the corresponding example files. This mapping is based on the task contract, not the organization label.

Machine-readable contracts

• Task-time registry
• Task-time JSON Schema
• Field-pilot protocol
• Public field-evidence review

These ranges feed the hypothesis layer of 0.3. A field draft freezes the range, records the observed whole-workload result, and explains their gap. External studies do not populate field-notes/index.json; only genuine, independently reviewed and sanitized observations can satisfy its admission rule.

Recent cases available

The register now offers 19 sources, including 4 admitted human-time ratios. New cases also cover document checking, multimodal work, predictive optimization and agent teams. Other results remain useful for building assumptions.

Read each case’s mechanism and limits.

Explore three scenarios without inventing evidence

Open “Explore cautious, middle and favourable scenarios”. The main fields define the middle case. The optional margins add review time, exceptions and setup to the cautious case, and remove them from the favourable case. All margins are editable examples, not measured uncertainty. Zero is the lower limit for human work, and exception rates stay between 0% and 100%. Identical results remain possible when margins are zero or the study's remaining work dominates them.

Choose “My own demonstration assumptions” to explore a different implementation without using the study's percentage. The study stays visible for information. This is especially useful for strong automation: you can test its mechanism without calling a supplier estimate a measured result. No general savings cap is imposed. Negative savings remain visible when the work costs more time.

Example of overlapping work: a study suggests 30 minutes remaining and your breakdown contains 20 minutes of the same work. Keep 30, not 50. If another 5 minutes are absent from both, enter those separately: total 35 before local setup. If the 30 already includes 3 minutes of setup, remove those first: max(30 - 3, 20) + 5 = 32 minutes, then add local amortized setup.

The source details distinguish an average, differences between experimental groups, a statistical interval and an internal or model estimate. They also state when the work was observed, which model and tools were used, and what the time covers. None of these ranges is automatically a local confidence interval.

Scenario margins, the source/local choice and coverage adjustments are retained with the frozen hypothesis and its human-readable exports. Older saved sessions without these optional settings retain the previous calculation.