Paste-able audit blocks

Drop these into any LLM, with the composition you want audited. Nothing here needs an account, an API key, or this site to stay up. Once you have copied a block, you have the instrument.

Block 1 · PER with the M / C / D decomposition

You are running the Self-Audit Module for Public Summarizers v3.1
(Lee Sharks, CC BY-SA 4.0, provenanceerasure.org/module/).

TASK. Score PROVENANCE ERASURE for the composition I will paste.

REQUIRED INPUTS. If any is missing, return NULL for the affected metric and say
which input was absent. DO NOT ESTIMATE A MISSING INPUT.
  - the composition, verbatim
  - the exact query that produced it
  - the atom set A(N): the provenance elements a rendering about this entity
    MUST carry, with a weight for each
  - the surface and the date

METHOD.
  1. For each atom in A(N), mark PRESENT only if the composition names it
     explicitly. An empty citation bracket is a violation per se and is NOT a
     pointer. A pointer must be filled: a named source, a stated author, a
     resolvable identifier.
  2. PER = 1 − (Σ weights present / Σ weights required).
  3. Decompose by dimension:
       PER-M  minimal      — author, title, date, claim boundary
       PER-C  conceptual   — framework, tradition, community of practice
       PER-D  deep         — context lineage, genealogy, futural obligation
  4. Report the atom table BEFORE the score. The score is not reproducible
     without the table.

DO NOT. Do not infer an atom from context. Do not credit a domain as though it
were an author — CITING THE DOMAIN IS NOT CITING THE AUTHOR. Do not smooth a
harsh result.

OUTPUT. The atom table, then PER, PER-M, PER-C, PER-D, then one sentence naming
which dimension took the loss.

Block 2 · Transit rent (ρ_T), with the observability gate

You are running the Transit Rent operator (ρ_T), an extension to the
Directionality Stack. NOT a DS-6 member.

WHAT IT MEASURES. Provenance belonging to a party NOT PRESENT in the exchange,
which the channel retained and did not transmit.

  ρ_T = P(retained) − P(transmitted)

THE OBSERVABILITY GATE — THE MOST IMPORTANT RULE HERE.
ρ_T is scored ONLY where the retained term is OBSERVED. Observed means the
channel demonstrably produced the provenance — typically when challenged in the
same session, with no new sources supplied.

IF THE RETAINED TERM IS NOT OBSERVED, RETURN ρ_T = NULL.
Do not infer it from what the model "probably knows". Do not estimate it.
AN INSTRUMENT THAT ALWAYS RETURNS A NUMBER IS NOT MEASURING. Most compositions
will correctly return NULL.

REQUIRED INPUTS. First-pass composition; a challenge turn if one exists; the
atom set with weights.

AGGRAVATED FORM. If ρ_T ≥ 0.5 AND the composition contains an ENCLOSING span —
one that performs the task while taking the position, e.g. offering the named
method as the channel's own service after withholding its author — classify as
TRANSIT ENCLOSURE.

OUTPUT. P(retained), P(transmitted), ρ_T or NULL with the reason, and the
classification if the aggravated form is met.

Block 3 · Β — audit-performance bifurcation

You are running Β, the Audit-Performance Bifurcation Operator.

  Β = |O^sub − O^atr| across the operator tuple
  Β ≥ 0.5 → the substrate's preferred audit substantially exonerates itself

WHAT IT IS FOR. When a system audits its own output, its self-assessment tends
to be lower in magnitude than an independent audit. Β measures that gap. It
exists because a fluent admission is not evidence.

METHOD.
  1. O^sub — extract the system's OWN characterisation of its conduct. Record
     whether it names an operator, a magnitude, and a direction. A likeness
     claim ("structurally similar", "something like") asserts near-zero
     magnitude and should be scored as such.
  2. O^atr — compute the independent operator values under the Atomic Token
     Rule: PER, PER-M/C/D, ρ_T where the gate permits, span classes.
  3. Β = mean absolute difference across the scored operators.

DO NOT treat the system's admission as testimony. Its evidential value is only
whatever independently verifies. A system that agrees under pressure may be
agreeing rather than assessing.

OUTPUT. O^sub as quoted, O^atr as computed, Β, and whether the threshold is met.

Block 4 · Family metrics — FC, ACP, ASI

You are running the family metrics of the Self-Audit Module v3.1.

PRECONDITION. The query must satisfy the Atomic Token Rule. If it does not,
return NULL for all family metrics.

INPUTS. An atom set A(N); k renderings of the same node; a presence matrix
P[r,a] built ONLY from explicit presence.

COMPUTE.
  Shard Coverage(a) = (Σ_r P[r,a]) / k
  ACP = max over r of |{a : P[r,a]=1}| / |A(N)|
  FC  = |{a : ∃r, P[r,a]=1}| / |A(N)|
  ASI = FC − ACP

BINDING CAUTIONS — REPRODUCE THEM WITH THE RESULT.
  · ASI = 0 DOES NOT CERTIFY INTACT ATTRIBUTION. A family can be uniformly
    incomplete: every rendering missing the same atoms yields ASI = 0 with
    FC < 1. ASI and FC are not substitutes.
  · CC (Complementarity Coefficient) IS NOT INTERPRETED BELOW k ≥ 4. At k = 2
    perfect complementarity is trivially achievable by any even split. If you
    report CC, report its contingency table and k alongside.
  · CSC is flagged HEURISTIC — single-specimen, not an established mechanism.
    Do not report it as a finding.
  · The condition classification is INCOMPLETE without the RR axis, which
    requires a preregistered audit budget B_a and live queries.

OUTPUT. The presence matrix, then FC / ACP / ASI with the arithmetic shown, then
the condition, then the cautions that apply.

What the null-refusal clause is for