info-tech-canon/demand/EvidenceBasis.md

# Demand: Evidence Basis — how a quantity was obtained, and how far a conclusion may be trusted

**Status:** accepted (canon 0.4.0)
**Date:** 2026-08-15
**Source:** resource-control (consumer), domain `financials`
**Owner:** **ITC-GOV** (Evidence), imported by `ITC-CAP` `CapabilityConsumption` (CAP-R10)
**Requested by:** info-tech-canon, in reply to the ITC-CAP restatement, 2026-08-15
**Consumer implementation:** `resource-control/tools/basis.py`, `docs/evidence-basis.md`
**Workplan:** `ITC-WP-0015`
**Catalog:** `infospace/models/governance/evidence-basis.yaml`

---

## Demand signal

ITC-GOV §11.34 defines **Evidence** as "information used to support a claim",
with examples that are all *artifact kinds*: a log extract, a scan result, a
test result, a deployment record. §11.35 defines an **Assertion** as a claim
that something is true, supported by evidence.

Nothing in the canon says **how a quantity was obtained**, and therefore nothing
bounds how far a conclusion computed from it may be trusted.

That is a different question from which artifact supports a claim. An artifact
can be impeccable while the number inside it is a guess. A consumer reading two
numbers of the same shape — `7.35` and `60.00` — cannot tell that one is a
published price and the other is an assumed hour count multiplied by an assumed
rate, and the canon currently gives it no way to say so.

## Why this cannot stay private to one consumer

`resource-control` has implemented and is running this. We are filing rather
than keeping it because the same argument info-tech-canon accepted for
`CapabilityConsumption` applies unchanged: a private vocabulary diverges the
moment a second repository reports evidence-bearing quantities.

It is already wanted in three places on the consumer side alone — capability
consumption rows, monthly actual observations, and threshold evaluations. That
breadth is also why the concept should **not** live in ITC-CAP: it is not about
capabilities. It is about quantities, wherever they occur.

---

## Proposed concepts

### `EvidenceBasis`

A closed, ordered vocabulary naming how a single quantity was obtained.

| Basis | Meaning |
|---|---|
| `invoiced` | Taken from an authoritative financial record |
| `measured` | Directly observed from the authoritative system |
| `quoted` | Stated by a provider or counterparty in a citable source |
| `derived` | Computed from other values by a stated rule |
| `projected` | Interpolated between, or extrapolated beyond, observations |
| `estimated` | Human judgement, neither observed nor computed |
| `assumed` | A modelling constant chosen by the author |
| `unknown` | No value exists |

`invoiced`, `measured`, and `quoted` are **evidenced**: they assert an observed
or contracted fact. Everything below is inference.

**On `invoiced`.** Per info-tech-canon's own instruction, this names a fin-hub
fact rather than originating one. The basis says only *where a value came from*;
it does not define booked-cost, tax, or currency semantics, which stay with the
financial authority under its own exchange contract.

**Strength is a tier, not a total order.** `invoiced` and `measured` are peers.
An invoice is authoritative for a payment and a measurement is authoritative for
a quantity; neither outranks the other outside its own domain. The consumer's
first implementation got this wrong by using a strict list order, and corrected
it — `weakest(["invoiced", "measured"])` must not imply a difference that does
not exist.

### `derived_from` and the propagation rule

> **A derived value is only as strong as the weakest value it was computed from.**

This is the rule that earns the vocabulary its place. Without it, arithmetic
launders assumptions: a two-decimal euro figure reads like a measurement when it
is an assumed hour count times an assumed rate.

A `derived` value carries `derived_from`, the values it was computed from, and
resolves to the weakest of them, recursively. Derivation is not penalised in
itself: GB computed from measured bytes resolves to `measured`.

### `unknown` carries no quantity

A value with basis `unknown` MUST NOT carry a quantity, and MUST name the gap
and the party who owns it. A value with any other basis MUST carry a quantity.
Zero means "measured and none"; unknown means "not measured". Both directions
should be enforced, because the failure mode runs both ways — a missing amount
silently read as zero, and a zero mistaken for an absence.

### `proxy_for`

An orthogonal qualifier: the value is a genuine measurement of a *different*
quantity than the one being reasoned about. It does not weaken the measurement;
it weakens the inference drawn from it. A shared-database allocation driver may
measure bytes precisely while one of those databases has never recorded a
transaction.

### `DecisionGrade`

A rollup over the values a conclusion rests on — an **Assertion** whose strength
is bounded by its weakest supporting value.

| Grade | Condition |
|---|---|
| `evidenced` | Every load-bearing value is invoiced, measured, or quoted |
| `projected` | The conclusion rests on values projected from observations |
| `indicative` | The conclusion is no stronger than an estimated or assumed value |
| `insufficient` | At least one load-bearing value is unknown |

This is the seam to existing ITC-GOV vocabulary: Evidence supports an Assertion;
a DecisionGrade states how far that support actually reaches.

---

## Worked evidence from the consumer

The mechanism was applied to the consumer's own most consequential decision —
the provider selection that produced its backup resource, a €29.14 per month
difference stated to the cent:

```
by_basis         {quoted: 1, projected: 1, assumed: 2}
weakest          assumed
evidenced_ratio  0.25
grade            indicative
```

One of four load-bearing values is evidenced. The two labour figures — the ones
that actually inverted the provider ranking — are each `derived` from an
`assumed` hour count and an `assumed` rate, so both resolve to `assumed`.

The finding did not overturn the decision: the direction is robust under any
plausible rate. It did establish that the *magnitude* is a model output rather
than a measurement, and it identified the cheapest way to strengthen it —
record real operator hours, not refine the arithmetic.

A mechanism whose first act is to qualify its author's own headline number is
behaving correctly.

---

## Proposed placement

**Owner: ITC-GOV.** `EvidenceBasis`, `derived_from` and its propagation rule,
the unknown rule, `proxy_for`, and `DecisionGrade` sit beside Evidence,
Assertion, and Attestation in §11.34–11.36. They describe the epistemic status
of information, which is ITC-GOV's ground.

**Import: ITC-CAP.** `CapabilityConsumption` gains an evidence basis per row.
It already distinguishes unknown from zero, which is the same discipline applied
to presence; basis applies it to the values that *are* present. `CAP-R8` would
be joined by a rule that a consumption row declares its basis.

**Also imports:** ITC-OBS (a measurement is `measured`; a modelled series is
`projected`), and any model carrying quantities that consumers compare.

**Avoid:** defining this only in ITC-CAP. It is not about capabilities, and
scoping it there would force a second, incompatible vocabulary the first time a
non-capability quantity needs grading.

---

## Success criteria

1. A closed basis vocabulary with tiers, not a spurious total order.
2. A stated propagation rule for derived values, computable and testable.
3. Unknown carries no quantity and names its gap and owner; zero is a
   measurement.
4. `proxy_for` expressible and orthogonal to strength.
5. Decision grades defined against existing Assertion semantics.
6. `CapabilityConsumption` imports the basis without ITC-CAP owning it.
7. `resource-control` can drop its private vocabulary in favour of the canon's
   without changing any recorded value.

---

## Non-goals

- Booked-cost, tax, or currency semantics. `invoiced` names a financial fact;
  it does not define one.
- A confidence interval, probability, or statistical error model. Basis is
  categorical provenance, not quantified uncertainty. A value may be `measured`
  and still be a single noisy sample — age and sample size are separate
  qualifiers and the canon should not conflate them with basis.
- A data-quality scoring or maturity scheme for datasets.
- An automated inference of basis. It is declared by the author of the value.