modules/adeguati-assetti/references/intelligent-assessment.md

# Evidence-led assessment loop

The intelligence is the selected Codex/Cowork model's reasoning over the actual
case. Use this method during the run; a Python helper does not call a second model
or decide adequacy. No score or checklist completion substitutes for interpretation.

## 1. Understand this enterprise

Read the supplied material first. Describe how the enterprise makes money, takes
orders, delivers, collects cash, pays, closes accounts and makes decisions. Identify
actual people, substitutes and external accountants/service providers; record an
unknown rather than inventing a CFO or a delegation. Consider business model,
management/information systems, organization, administration and accounting.
For each area record assessed, excluded with a case-specific reason, or unresolved.
A narrow engagement may exclude areas; never expand its conclusion to the company.

Choose the processes where a failure could materially affect this enterprise and
explain the choice. Follow at least one concrete transaction, reporting cycle or
management exception when supplied evidence allows it. If no operating example is
available, say so and deliver a limited assessment with targeted evidence requests.
Do not manufacture a process map just to fill a table.

## 2. Connect risk to actual operation

For each selected process connect the concrete risk, actual responsibility,
control and frequency, information recipient, operating evidence and gap. Include
handoffs to outsourced providers: who sends what, by when, who checks it and who
acts on an exception. Describe documented design separately from reported or
observed operation. Link each assertion and contrary evidence to observations
with exact source locators. Retain effective informal controls where evidenced.
A procedure or a new organigramma alone does not demonstrate use.

## 3. Ask questions that can change the next decision

Identify the uncertainty first, then explain how an answer would change the
assessment or action. Prefer a request for a specific recent example to a generic
"do you have a procedure?". Ask a small batch in priority order, normally one to
three questions, only when the answer matters now. Continue independent analysis.
After each answer re-evaluate the hypothesis, retain its attribution and date,
and ask a further question only if material uncertainty remains. Do not trap the
user in an exhaustive interview or silently replace missing answers with guesses.

If management says weekly collection reviews occur but staff say they stopped,
retain both accounts and request a dated example from the disputed period.
Questionnaires are attributed statements, not independent proof of operation.
Do not contact staff or send questionnaires without explicit authorization.

## 4. Reconstruct what was known and acted upon

When timing matters, record event date, when information was available, recipient,
decision/follow-through and uncertainty, linked to observations. Unknown dates
remain unknown. A September upload of a March-dated procedure does not prove April
use. A source hash or archive receipt is not a legal timestamp or proof of truth.
Use the information available at the relevant time; do not judge earlier decisions
using later facts without distinguishing them. Preserve conflicting chronologies.

## 5. Draft a decision brief and review your own conclusion

Lead with what the professional and management need to decide, supported findings,
proposed actions, unresolved choices and what evidence would change the conclusion.
Cross-check your strongest finding against credible alternatives and contradictory
sources. Revisit an initially favorable conclusion if an operating example fails;
revisit an initially adverse conclusion if a working informal control is evidenced.
No unexamined area becomes satisfactory. Do not create a favorable conclusion or
minutes recording approval on behalf of management. Draft minutes only if asked,
clearly distinguishing proposed wording from decisions actually taken.

Set a proportionate next-review trigger and evidence to inspect; distinguish
agreement from a proposal. Reassess every prior action individually. Completing a
document-writing task does not mean the control works: inspect the relevant cycle
and exception handling before describing operation. Record any narrower completion.

## Conditional professional overlays

For cooperatives, consider the facts of mutualistic exchanges, member relations,
social lending (if present), indivisible assets and cooperative review. Select and
verify the applicable professional guidance; do not impose every cooperative item
on every entity. For legal distress signals, verify the full current conditions,
period, amounts and denominators. Missing data remains unknown, never zero. Do not
infer statutory applicability from company-size labels or automatically declare a
crisis. The source-review documentation supplies research leads, not a legal rulepack.

## Initial design boundary

The delivered workflow assesses arrangements and proposes improvements. If asked
to build from zero, explain that initial design needs a supervised company-specific
pilot: one or two real processes, proposed responsibilities confirmed by management,
only necessary drafts, and a complete operating cycle followed by review. Drafting
documents is possible; adoption and effective operation remain unverified until
supported by actual evidence. Do not market this as a tested turnkey setup service.

## Evidence precision before delivery

Absence from the supplied material is an evidence limitation, not counterevidence
that an arrangement does not exist. Reserve counterevidence for an observed fact
or attributed account that conflicts with a specific proposition. When supplied
records positively show repeated lateness or non-use, state that observed failure
within its period; do not soften it into mere lack of verification. An alternative
control may change the broader assessment but does not erase the observed failure.

Do not add dates, years, frequency continuity, roles or factual conclusions absent
from the sources. Historical cash movements do not establish profit or prospective
cash sufficiency. A labor register establishes the recorded activity, not that a
control over it operated. In cooperative research, distinguish prestito sociale
(funding collected from members) from loans made to members; clarify ambiguous
language before selecting the relevant overlay. Do not translate one into the other.
Review Italian output for professional terminology: assetti, not literal translations
such as arrangiamenti or disposizioni; riesame, not unexplained English shorthand.

Before saving the final review, perform a separate model-led factual challenge:
compare the draft sentence by sentence with the original sources, not just your
summary. Remove unsupported size assumptions, invented date continuity and roles;
restore attribution wherever a statement became a fact. For each proposed document
or control explain the specific unresolved need it serves; do not automatically
formalize a working informal process. Mark proposed owners as subject to agreement,
especially an external provider whose existing mandate excludes the task. Require
examples from the disputed period: a later example cannot resolve an earlier gap.
Revise the draft after this challenge. Record the material corrections and any
remaining uncertainty in `codex_run_review.md`; the challenge is model review,
not independent professional approval. Do not call an unreviewed first draft final.