The Human^AI and Org^AI calculator: how it works, layer by layer

SState S

S = { metadata, classification, counts, evidence, thresholds, … }

A single object held in memory. It is the only source of truth: every result, on screen, in the report or in the JSON, derives from it.

• No localStorage, no cookies, no database.
• A JSON import rebuilds exactly this object.

Receives: —
Feeds: —

compute()compute()

R = compute(S)   // pure function, no I/O

All computation goes through one pure function: the same state always gives the same results. The logic is delimited by two markers in the code.

• 29 automated tests extract the formulas from the shipped file, between the two markers, so they cannot drift from it.
• A passing test is a statement about the file you have, not about a lab version.

Receives: —
Feeds: 01 Epistemic decoupling

update()update()

update() → rebuilds the outputs only

Display computes nothing. On every change the results are recomputed in full and only the output area is rebuilt.

Receives: —
Feeds: —

HTMLA single HTML file

index.html · 507 KB · 0 dependencies

No framework, no build step, no runtime dependencies. Fonts are embedded, so it also works on an isolated network.

• Zero network requests; a restrictive CSP with connect-src ‘none’ in the locally hosted variant.
• Light, dark and system themes.

Receives: —
Feeds: —

METAReporting metadata

metaComplete ⇔ 9/9 required fields

The framework’s minimum reporting convention, turned into an operating condition: unit, task, error consequence, AI configuration, period, validation method, accountability boundary and assessor, plus the record identifier.

• While anything is missing, results are marked provisional and export is blocked.
• The modules keep computing, so work can continue, but no record can claim more than is known.

Receives: —
Feeds: Gate 2 · Export

CLASClassification test

10 questions × (state, reference) + 5 exclusion cases

The membership test: it establishes whether the situation is a Human^AI case and whether it meets the conditions for Org^AI. Each question has three states: demonstrated, unmet, unknown.

• “Unknown” never counts as demonstrated.
• Excluded cases include an ignored suggestion, automation with no AI component, and several licences used independently.

Receives: —
Feeds: Gate 1 · Classification

n kA kVObserved counts

n, kA, kV ∈ ℕ  ·  0 ≤ kV, kA ≤ n

Whole numbers, not percentages: how many tasks are in the declared set, for how many the AI gave a usable answer, for how many the accountable person produced a verified justification.

• “About 90%” cannot be checked; “126 of 140” can.
• Only counts allow a confidence interval to be computed.

Receives: —
Feeds: 01 Epistemic decoupling

EVIDENCEEvidence

evidence = (reference, type)

For every claim: a reference and its type, for example measurement, permission test, audit. The type matters: a “yes” for security (S) that does not rest on a permission test raises a flag.

Receives: —
Feeds: documented

θₓDeclared thresholds

θₓ per dimension · thresholds for the risk gates

Declared by the user, not by the tool, before the results. If a threshold changes after results have been entered, the change goes into the record.

• A risk gate with no declared threshold stays unmet.
• The tool does not calibrate thresholds.

Receives: —
Feeds: 04 Material contribution profile, 05 Synergy test

JSONJSON import

schema v2 · v1 migrated

Reloads an existing record. Records from versions 1.0.0 and 1.1.0 are migrated, with an explicit note on what could not be recovered.

• It is the end of the re-assessment loop: the same unit, another period, counts redone.

Receives: —
Feeds: —

MODEWorking mode

explore ∈ {true, false}

Explore: values move freely, evidence is not required, results are marked illustrative and export is disabled. Assess: the tool becomes demanding.

• The explicit split prevents results that look like an assessment without being one.

Receives: —
Feeds: requireEvidence, Gate 2 · Export

REQrequireEvidence

requireEvidence = !explore

In assessment mode, every claim must have an evidence reference.

Receives: MODE Working mode
Feeds: documented, documented

DOCdocumented

documented = !requireEvidence
          || ref ≠ ''

A claim is documented if evidence is not required (explore) or if it has a reference filled in.

Receives: Evidence, requireEvidence, requireEvidence
Feeds: Effective state, Effective state

EFFEffective state

effective =
  (state='yes' ∧ documented) ? 'yes'
  : (state='no')             ? 'no'
  :                            'unknown'

All thresholds and chains are computed on the effective state, not the declared one. A “yes” without a reference becomes “unknown”, and “unknown” never counts as demonstrated.

• This is where the “declared · documented” pair in module 03 comes from.

Receives: documented, documented
Feeds: 03 Org^AI threshold, F5 Undocumented compliance

GATE 1Gate 1 · Classification

material contribution unmet ⇒ STOP

If the AI makes no material contribution, the situation is not a Human^AI case for that episode and the measurement modules no longer apply. The record stops here.

• The first question is not “how much” but “what are we assessing”.

Receives: Classification test
Feeds: 01 Epistemic decoupling, 04 Material contribution profile

01Epistemic decoupling

Δepi = A − V
A = kA/n   V = kV/n
CI: Wilson on A and V; the difference by the Newcombe method

The gap between the range of answers AI makes accessible and the accountable actor’s demonstrated capacity to validate them. When access grows faster than validation, the organisation produces claims it cannot support, and then reuses them.

• If the interval includes zero, the verdict is “not demonstrated at this task volume”, not “no gap”.
• 0.22 on 140 tasks → [0.13; 0.31], demonstrated. 0.25 on 12 tasks → [−0.08; 0.53], not demonstrated.
• A positive gap opens the list of seven mitigation measures; whatever is ticked feeds module 02.

Receives: n kA kV Observed counts, Gate 1 · Classification, compute()
Feeds: F9 Below threshold, 02 Propagation exposure, 07 Org^AI profile

02Propagation exposure

Ωprop = max(0, Δepi) · ρ · κ

An unvalidated error is amplified through reuse. The module computes two states on the same scale, now and after the declared controls, and shows the percentage reduction.

• A conceptual ordering indicator, not a probability; without local calibration it predicts nothing.
• It flags separately the case where the target reduces exposure by cutting access instead of raising validation: both lower the figure, but they are entirely different decisions.

Receives: 01 Epistemic decoupling
Feeds: 07 Org^AI profile

03Org^AI threshold

Iorg = 1[Korg ∧ P ∧ R ∧ G ∧ S ∧ T]

Six constitutive conditions, all on the effective state: authorised knowledge, processes, roles and authority, governance, security, traceability. All six, or the threshold is not demonstrated.

• Each criterion has three fields: state, evidence reference and its type.
• The counter becomes a pair: “5/6 declared · 4/6 documented”; the second one counts.
• L, validated learning, is the seventh component and does not enter the threshold: it conditions M5 only.

Receives: Effective state
Feeds: F2 Memory without governance, F8 Unclarified authority, 06 Maturity per domain

04Material contribution profile

Y := ⟨Q, τ, E, C, A, L, Ω⟩
material ⇔ ∃x relevant: |dₓ| ≥ θₓ

The AI contribution assessed on seven dimensions at once: quality, time, error, calibration, autonomy, learning, risk exposure. The result is a profile, never a score.

• The threshold is declared before the results, otherwise classification becomes opportunistic.
• A dimension that cannot be measured credibly does not get a zero; it stays out of the decision.
• The verdict reads, for example: “material on 3 of 5 relevant dimensions”, with the direction of each.

Receives: Gate 1 · Classification, θₓ Declared thresholds
Feeds: F6 Single dimension, 05 Synergy test

05Synergy test

margin = Q* − max(comparators)
synergy ⇔ margin > 0 ∧ every risk gate satisfied

Synergy is not assumed and it is not a maturity level: it is demonstrated on the same task, against every relevant comparator, with every declared risk threshold respected.

• Four risk gates, each with a declared threshold and an observed value; a gate with no threshold blocks synergy just like a failed one.
• The result names the comparator that binds the conclusion, for example “margin +0.06 over Human + AI (0.76)”.

Receives: θₓ Declared thresholds, 04 Material contribution profile
Feeds: F4 Weak comparators

06Maturity per domain

M0 → M1 → M2 → M3 → M4 → M5   (per domain)

The level is assigned to a bounded domain and a process, not to an organisation. A strict chain: a level is reached only if every earlier level is demonstrated.

• M0 Exploratory · M1 Episodic · M2 Managed · M3 Connected · M4 Governed · M5 Cumulative.
• L conditions M5 only. Limits: M3 – connectivity does not demonstrate governance; M4 – learning may remain manual; M5 – adaptation does not demonstrate synergy.
• A domain where AI is prohibited appears with no level, not as M0.

Receives: 03 Org^AI threshold
Feeds: —

07Org^AI profile

9 evaluation dimensions · no aggregation

The threshold says the architecture exists; this module says what it produces: nine organisational evaluation dimensions, each with its value and its evidence.

• Coherence checks: the validation dimension against V from 01, propagation against Ωprop from 02.
• A large difference between them is a finding in itself: it points to a merely formal check.

Receives: 01 Epistemic decoupling, 02 Propagation exposure
Feeds: SCREEN Working screen

F1Licences without architecture

“connected” = yes ∧ ≥2 criteria unknown

A detector predicate. When it fires, the flag shows the concrete values that triggered it and the correction the framework prescribes; it can be annotated, and the annotation goes into the record.

Receives: —
Feeds: —

F2Memory without governance

validated learning = yes ∧ G ≠ yes

A detector predicate. When it fires, the flag shows the concrete values that triggered it and the correction the framework prescribes; it can be annotated, and the annotation goes into the record.

Receives: 03 Org^AI threshold
Feeds: —

F3Nominal governance

G = yes ∧ validation < 50%

A detector predicate. When it fires, the flag shows the concrete values that triggered it and the correction the framework prescribes; it can be annotated, and the annotation goes into the record.

Receives: —
Feeds: —

F4Weak comparators

existing process = weakest benchmark

A detector predicate. When it fires, the flag shows the concrete values that triggered it and the correction the framework prescribes; it can be annotated, and the annotation goes into the record.

Receives: 05 Synergy test
Feeds: —

F5Undocumented compliance

“yes” claims with no evidence reference

A detector predicate. When it fires, the flag shows the concrete values that triggered it and the correction the framework prescribes; it can be annotated, and the annotation goes into the record.

Receives: Effective state
Feeds: SCREEN Working screen

F6Single dimension

only one relevant domain declared

A detector predicate. When it fires, the flag shows the concrete values that triggered it and the correction the framework prescribes; it can be annotated, and the annotation goes into the record.

Receives: 04 Material contribution profile
Feeds: —

F7Irreversible without a test

S = yes ∧ evidence ≠ permission test

A detector predicate. When it fires, the flag shows the concrete values that triggered it and the correction the framework prescribes; it can be annotated, and the annotation goes into the record.

Receives: —
Feeds: —

F8Unclarified authority

role unknown ∧ Org threshold met

A detector predicate. When it fires, the flag shows the concrete values that triggered it and the correction the framework prescribes; it can be annotated, and the annotation goes into the record.

Receives: 03 Org^AI threshold
Feeds: —

F9Below threshold

Δepi ≤ 0.05 ∧ at least 3 signals

A detector predicate. When it fires, the flag shows the concrete values that triggered it and the correction the framework prescribes; it can be annotated, and the annotation goes into the record. Δepi comes from module 01.

Receives: 01 Epistemic decoupling
Feeds: —

F10Threshold without result

Org demonstrated ∧ no relevant dimension

A detector predicate. When it fires, the flag shows the concrete values that triggered it and the correction the framework prescribes; it can be annotated, and the annotation goes into the record.

Receives: —
Feeds: —

GATE 2Gate 2 · Export

canExport = !explore ∧ metaComplete ∧ !notClassified

The record can be exported only in assessment mode, with complete metadata and with the situation classified. Possible states: complete, incomplete, not classified.

Receives: META Reporting metadata, MODE Working mode
Feeds: JSON record

SCREENWorking screen

verdicts · intervals · profiles · flags

Verdicts, confidence intervals, per-dimension profiles and diagnostic flags, recomputed in full on every change. The results strip keeps four figures in view: epistemic gap, propagation exposure, organisational threshold and maturity.

Receives: 07 Org^AI profile, F5 Undocumented compliance
Feeds: —

PRINTPrinted report

same state S → ~17 A4 pages

The document an auditor or client receives: a header with the nine fields, the identifier, status and date, then every module with its values rendered as text. Nothing on screen can be missing from it.

Receives: —
Feeds: —

JSONJSON record

schema v2 · reloadable · stable id

It holds the tool and framework versions with their DOIs, all inputs, the computed values, evidence references, flags with annotations, the identifier, the generation time and the limits statement. It is copied to the clipboard, not downloaded.

• Two records sharing an identifier form a comparable series.

Receives: Gate 2 · Export
Feeds: —

✕What it does not do

does not verify · calibrate · estimate · aggregate · store · certify

The limits stay permanently visible in the interface, not only in the documentation. Results are interpretive and cannot be the sole basis of a consequential decision.

Receives: —
Feeds: —