HomeParameters
The parameter file
Every weighting, threshold and task score the audit uses, in one machine-readable file. Published so any verdict can be checked, disputed and cited without asking.
Download
The canonical file is v4.2.json.
It is licensed CC BY 4.0, so you may reuse, redistribute and build
on it commercially, with attribution.
Suggested citation: AI Work Audit automation readiness parameters, v4.2 (2026). https://aiworkaudit.net/parameters/
| Field | Value | Meaning |
|---|---|---|
| Framework version | 4.2 | Bumped whenever a weighting or a task gate changes. |
| Released | 2026-09-09 | First publication of this version. |
| Last recalibrated | 2026-09-09 | Date of the most recent re-score. |
| Recalibration trigger | GPT-6 Astra (2026-09-03) | The frontier release that prompted it. |
| Tasks | 33 | Across 11 industry categories. |
| Capability-gated | 19 | Expected to move with model capability. |
| Liability-gated | 14 | Structural. A better model does not move these. |
| Licence | CC BY 4.0 | Free to reuse with attribution. |
What is in it
- The readiness formula and its risk and judgment multipliers
- The output model, including the leverage curve and iso-output staffing
- The three cost scenarios, oversight premiums and spot-check rates
- The full oversight contract: sampling bases, ramp multipliers, error ceilings, sample-adequacy rule and the four delegation levels
- All 33 tasks with their repeatability, judgment, risk, gate and primary failure mode
Where these numbers came from
A published formula invites the question of where its constants came from. Most of them here are hand-calibrated judgment with no empirical basis, and the table says so per constant rather than leaving you to guess. Two are genuinely sourced, one follows from a statistical result, and the rest are arguable.
| Constant | Value | Status | Basis |
|---|---|---|---|
| MODELS.*.input / output | per published rate card | observed | Provider list prices, re-checked at each recalibration. The only externally sourced numbers in the framework. |
| oversight_contract.min_sample | ceil(3 / error_ceiling) | derived | The rule of three: observing zero events in n independent trials gives an approximate 95% upper bound of 3/n on the true rate. Not a choice. |
| WORKING_DAYS / STANDARD_DAY_HOURS | 22 / 8 | convention | Standard working month and day. Changing them rescales cost and output together, so cost per unit is unaffected. |
| RISK_FACTOR | low 1.00 · medium 0.85 · high 0.55 · critical 0.00 | judgment | The most load-bearing constant in the framework. Hand-calibrated so that high-risk work cannot reach an automate verdict on economics alone, and critical work cannot score at all. No empirical basis. If you disagree with one number in this file, it should probably be this one. |
| JUDGMENT_FACTOR | low 1.00 · medium 0.85 · high 0.60 | judgment | Calibrated alongside RISK_FACTOR so that judgment-heavy work carries a penalty comparable to one risk tier. No empirical basis. |
| LEVERAGE_MIN / LEVERAGE_MAX | 0.15 / 1.60 | judgment | How much extra output one supervised person gets from AI, as a multiple of their unaided rate, scaling with repeatability. The upper bound is the single least defensible number here and the biggest driver of results at low volume. Treat 1.60 as an optimistic ceiling, not a measurement. |
| OVERSIGHT_PREMIUM | low 0.20 · medium 0.35 · high 0.60 | judgment | The salary uplift on retained staff for reviewing AI output and handling exceptions. Cross-checked against the sampling schedule by the coverage test, which flags when the premium cannot fund the review hours. |
| Scenario B spot-check | max(0, review hours needed – hours the premium funds) | derived | Follows from the sampling schedule and the oversight premium rather than being chosen. Replaced the AUTO_OVERSIGHT table in v4.2, which was scaled by pre-automation headcount and double-charged against the premium. |
| shape / economics split | 55% / 45% | judgment | Introduced in v4.1 so volume could not be ignored. Sensitivity testing shows it moves no verdicts at mid volume and one at low volume. The v4.1 changelog overstated this and was corrected in v4.2. |
| CPU_GAIN_TARGET | the cost-per-unit improvement treated as full marks | judgment | Sets where the economics half of the score saturates. Low impact on verdicts at either scenario tested. |
| FIXED_COVER_TARGET | 6 | judgment | Manual spend should cover platform fee plus amortised setup this many times before the economics score is full. Matters only at low volume, where it moves two verdicts. |
| sampling base rates / ramp multipliers | 0.02–1.00 · 5× then 2.5× then steady | judgment | Starting sampling rate by risk tier, and the step-down schedule. The shape is conventional audit practice; the specific numbers are not measured. |
| error ceilings | low 5% · medium 3% · high 1% · critical 0% | judgment | Stop-the-line thresholds by risk tier. These drive min_sample, which is derived, so an unjustified ceiling propagates into a real statistical requirement. Worth challenging. |
| task repeatability / judgment / risk / gate | per task, 33 tasks | judgment | Assigned per task from practitioner experience, re-reviewed at each recalibration. These are the inputs most worth disputing for your own business, because your version of a task may not behave like the reference version. |
Which constants actually change the answer
Knowing a number is judgment matters less than knowing whether it moves anything. Each constant below was perturbed and all 33 tasks re-scored. The reference case is 6 people at 6h a day; results at low volume differ and are noted underneath.
| Constant | Change tested | Verdicts moved | Max score shift |
|---|---|---|---|
| RISK_FACTOR.medium | 0.85 to 0.70 | 7 of 33 | 14 pts |
| RISK_FACTOR.high | 0.55 to 0.40 | 4 of 33 | 11 pts |
| LEVERAGE_MAX | halved (1.60 to 0.80) | 2 of 33 | 10 pts |
| JUDGMENT_FACTOR.high | 0.60 to 0.45 | 0 of 33 | 5 pts |
| OVERSIGHT_PREMIUM.high | 0.60 to 0.90 | 0 of 33 | 3 pts |
| LEVERAGE_MAX | raised 50% (1.60 to 2.40) | 0 of 33 | 2 pts |
| CPU_GAIN_TARGET | halved | 0 of 33 | 2 pts |
| LEVERAGE_MIN | set to 0 | 0 of 33 | 1 pts |
| SHAPE_WEIGHT | shape 55% to 45% | 0 of 33 | 1 pts |
| SHAPE_WEIGHT | shape 55% to 65% | 0 of 33 | 0 pts |
| FIXED_COVER_TARGET | doubled (6 to 12) | 0 of 33 | 0 pts |
Read it this way. Risk weighting is the load-bearing choice at normal volume — softening the medium tier alone moves seven verdicts. At low volume the leverage ceiling takes over and moves nine. Those two are where scepticism is best spent. The 55/45 shape-to-economics split, which the v4.1 changelog presents as the headline change, moves no verdicts at mid volume and one at low volume; that entry overstates its effect and will be corrected.
Why it is public
A verdict you cannot inspect is an opinion with a number attached. Publishing the parameters means anyone can reproduce a result, disagree with a specific weighting rather than the conclusion, and cite the version they argued with. Every change is recorded in the changelog.