remediation

Files (1)

• SKILL.md

SKILL.md (first lines)

# REMEDIATION — Repairing a Propagated False Claim

**Select → Produce → Accept**

## Why this exists

Detection is not correction, and correction is the expensive half. Two
measurements bound this discipline rather than motivate it hypothetically:

- Authors directly notified that work they had cited was retracted still
  mostly failed to correct it: of 86 authors successfully e-mailed, 51%
  replied, and at one year only **9 of 86** had published a formal
  correction (Avenell, Bolland, Gamble & Grey, *Accountability in Research*,
  DOI 10.1080/08989621.2022.2082290 — grounded in this record at
  `.ledger/log/2026-08-24-the-number-was-narrower.md` [AV2], which corrects
  an earlier rounded restatement; the raw counts are cited here rather than
  a derived percentage for the same reason that entry gives).
- In a 2026 multi-agent pipeline with none of the human explanations
  available — no fatigue, no incentive, no institutional friction — **87 of
  266** caught cases had the gate correctly suppress the injected error and
  the downstream agent still fail to propagate the correction
  (`.ledger/log/2026-08-24-avenell-in-silicon.md` [AS4]).

A discipline that stops at "flag it" reproduces both results. This one does
not stop there. It is a **named, loaded** discipline rather than
documentation because this project has independently and repeatedly
measured that a procedure carried only in memory is a procedure not
consulted — the same failure class the two citations above measure at
larger scale.

> **Verification Dual anchor.** REMEDIATION closes on both paths at once,
> per constraint. Where a deterministic evaluator exists — a re-runnable
> check, a query over the corpus, `git diff` against an anchor — it is used
> and named. Where none can exist — locating a hand-authored sentence
> outside the corpus, judging whether a scope actually covers a dependent
> set — the constraint says so explicitly and names the witness or human
> closer that discharges it instead ([rules.md §2](../../rules.md#2-prime-invariants)).

## Scope

REMEDIATION is not constraint-authoring (`/spec`) and not the general
category model for durable artifacts (`/record`) — it composes both. `/spec`
is loaded because every step below is normative (MUST/SHOULD) and carries
its evaluator, the same discipline `/spec`'s CONNECT step imposes on test
invariants. `/record` is loaded because the artifacts this procedure
produces — a tech-debt entry, a refutation entry, a decision record — are
themselves durable-record categories with their own homes and contracts.

Trigger condition: a claim previously treated as settled — closed
(`proved`/`cited`) or merely asserted in prose, code, or conditioning — is
now known to be false, and something else in the system was built on it.
If nothing was built on it, this is a one-entry correction and this
discipline is unnecessary ceremony (Focus-before-ceremony,
[ambient.md](../../ambient.md#the-focus-level-selector)) — SELECT still
runs, briefly, to confirm that.

## Ontology

- **Claim** — a proposition treated as backing something else. Two homes,
  with different repair economics (PRODUCE, below):
  - **corpus claim** — a `.ledger` entry with an `id`, reachable by
    derivation edges (`because`, `depends`).
  - **out-of-corpus claim** — a sentence in repository documentation, a
    code comment, a type name, a conditioning file, or an `AGENTS.md`
    requirement. No edge tracks what depends on it.
- **Dependent** — an artifact whose correctness rests, transitively, on a
  Claim.
- **Species** — whether a false claim was **born-false** (never true — the
  backing did not hold at its own anchor) or **became-false** (true when
  written, false because the record grew around it). Derived, never
  authored, mirroring the record's existing rule that openness is derived
  rather than stated (`docs/entries.md`, "the pen law").
- **Currency verdict** — a claim of the form "nothing under scope S has
  changed since commit A" — the only form "still clean" is ever entitled to
  take (ACCEPT, below).

## Machinery status — read before citing a command

Everything under **SELECT** below runs against `master` today.
`dependents_of` exists and is verified: an independent `jq` recursive
reverse-walk over the extracted corpus returned the same **44** dependents
of `write-path:W1` that the hand-written view returns
(`.ledger/log/2026-08-24-jq-suffices.md` [JQ1]).

Everything under **PRODUCE**'s born-false/became-false routing and
**ACCEPT**'s currency verdict depends on machinery that, as of this
writing, lives on branch `feat/refutation-and-currency` and is **not yet
merged to master** — verified directly: `ledger/gate/currency.py` and the
`refutes` field are absent from `master`'s tree
(`ledger/contracts/entry.ncl` on master carries no `refutes` field outside
a comment; `ledger/gate/currency.py` does not exist on master). The design
is not speculative — it is an architect ruling
(`.ledger/deposits/expiry-mechanics/architect-seat/2026-08-24-ruling-expiry-mechanics.md`)
plus a compiled model with no `sorryAx`
(`.ledger/log/2026-08-24-refutation-is-one-anchored-edge.md` [RF1]) — but a
walk following this skill **before that branch merges** cannot invoke the
corpus views for species and currency, and must apply the same rule by
hand, as each constraint below states.

## Act 1 — SELECT: which artifacts depend on the false claim

**SEL-1.** For a corpus claim, the walk MUST enumerate its dependent set
via `dependents_of`, never via `grep` — grep answers about text, the corpus
answers about extracted entries, and they disagree
(`.ledger/log/2026-08-24-the-record-was-queryable-all-along.md` [QA3]/[QA5]).
*Evaluator (proved, machine):*
```
python3 ledger/derive/extract_entries.py .ledger -o /tmp/corpus.json
nickel export /tmp/corpus.json --apply-contract ledger/contracts/entries_query.ncl \
  | jq '.dependents_of[] | select(.id=="<claim-id>").dependents'
```
Grounded by `ledger/gate/test_dependents_query.sh` (multi-hop, mixed edge
kinds, cycle, external exclusion, empty-leaf cases) and the independent
`jq` cross-check above.

**SEL-2.** For an out-of-corpus artifact, no edge exists to walk — this is
the hard case named up front, not papered over. The walk MUST establish the
universe before claiming coverage (`git ls-files` for the repository,
`git -C .ledger ls-files log/` for the corpus proper) and then search by

From nrdxp/predicate, MIT.

More from nrdxp/predicate