Files
dotfiles/scripts/governance-drift-check.py
T
David F GliddenandClaude Opus 5 063eccfd80 [HARDENING] STATE-CLAIM + the resolution state, built together (REVIEWED-127)
Both halves of the schema, shipped in one change because the ruling said half a schema
invites a third patch and a third patch is how a vocabulary accretes instead of being
designed.

157 — resolution state. `resolved:` on any block; a resolved block is no longer due but
is NOT dropped: it prints as a closed ledger, because a discharge that vanishes from the
report is its own decay. The pointer must RESOLVE — a real path or a real git object —
so an undocumented discharge is impossible to express rather than merely discouraged. A
dangling pointer reports in the register-integrity lane, the same lane as an amendment
that replaced the record it amends; both are a record closing over its own history.

The 25th's hand-rename is MIGRATED back to DEFERRED-DECISION with resolved: set. That
block was the per-instance workaround 157 was filed against, and it is now the
migration's own test case.

158 — STATE-CLAIM. Reuses trigger_fired() verbatim and inverts only what firing MEANS:
for a deferral, fired = the decision is due; for a state-claim, fired = THE CLAIM IS
FALSE. Two new trigger kinds earned directly from today's instances: text-present (the
trial-09 hold, falsified by REVIEWED-124's existence) and file-changed-since ("the filed
rule not edited", false one hour after writing).

16 new controls, each with its discriminating half — fires on met, silent on unmet,
manual listed-never-fired, resolved excluded from due-ness, the SAME block unresolved
still due, a real pointer resolves, "yes, done" does not.

Proven on the LIVE blocks, not only fixtures: pointing the state-claim at an older
commit made it report FALSIFIED by name; replacing the resolution with "yes done" made
register-integrity report it; both restored and both returned to quiet.

⚠ One control failed before shipping and the failure was the useful part. The negative
control for file-changed-since pointed at FOOL-SEED-RULE.md, which this same session then
edited — so "unchanged since HEAD" broke, correctly. A control whose subject is "did this
file change" must not point at a file the session is changing. Re-pointed at a frozen
2026-08-02 trial artifact, with the reason recorded at the fixture. Caught because the
controls run on every invocation rather than in a separate suite.

First two real state-claims filed, deliberately one of each kind: ~/CLAUDE.md untouched
under PENDING-150, mechanically watched and [ESCALATE]-grade the moment it goes false;
and §9's channel unbuilt, marked `manual` because it has no filename yet and inventing a
proxy falsifier is the error the schema's own comment warns against.

⚠ The zero-state prints a WARNING, not a tick: "0 marked, NOT none-stale" with the ~57
unmarked candidates named as a grep. An instrument that reads nothing reports exactly
like one that finds nothing, and that is the failure this item exists to end.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01J6hZXNYSxEfZseBGTni4sf
2026-08-25 15:29:59 +02:00

820 lines
42 KiB
Python
Executable File

#!/usr/bin/env python3
"""
governance-drift-check — report state claims in ~/CLAUDE.md that the substrate contradicts.
Reports. Does not correct. Detection needs no authorization; correction does
(Constitutional Constraint #1). This exists so that a stale governance document
is *visible* rather than *misleading* — Constitutional Constraint #4, honest
degradation, applied to the governance document itself.
Every check carries a POSITIVE CONTROL in the same run: an absence is not
evidence until the instrument is shown capable of detecting presence.
(Ratified 2026-07-27 as an epistemic standard, jurist Q2.)
Built 2026-07-27 on steward authorization. Exit code is always 0 — this is a
report, not a gate.
"""
import datetime
import json
import os
import re
import signal
import subprocess
import sys
from datetime import date
from pathlib import Path
# Report cleanly when the reader closes early (`| head`), rather than tracebacking.
try:
signal.signal(signal.SIGPIPE, signal.SIG_DFL)
except (AttributeError, ValueError): # not POSIX, or not on the main thread
pass
HOME = Path.home()
CLAUDE_MD = HOME / "dotfiles" / "CLAUDE.md"
MONTHS = ("january february march april may june july august september "
"october november december").split()
findings: list[str] = []
controls: list[tuple[str, bool]] = [] # (label, passed)
def control(label: str, passed: bool) -> bool:
controls.append((label, passed))
return passed
# ---------------------------------------------------------------- load
if not CLAUDE_MD.exists():
print(f"drift-check: CANNOT RUN — {CLAUDE_MD} not found")
sys.exit(0)
text = CLAUDE_MD.read_text()
lines = text.split("\n")
# ------------------------------------------------- 1. referenced paths
# Backticked paths that look like filesystem locations.
raw = set(re.findall(r"`(~[^`\s]+|/Users/[^`]+?)`", text))
paths = {p for p in raw if ("/" in p) and not p.endswith(("`",))}
control("path-check instrument reaches the filesystem", CLAUDE_MD.exists())
for p in sorted(paths):
expanded = Path(os.path.expanduser(p.strip().rstrip(".,;")))
if not expanded.exists():
ln = next((i + 1 for i, l in enumerate(lines) if p in l), None)
findings.append(f"L{ln}: path does not resolve — {p}")
# ------------------------------------- 2. named MCP tools / servers
# Tool names the document instructs the executor to call.
tool_names = set(re.findall(r"`(kg_\w+|diary_\w+|find_tunnels|traverse)`", text))
configured: set[str] = set()
for cfg in (HOME / ".claude/settings.json", HOME / ".claude/settings.local.json",
HOME / ".claude.json", HOME / ".mcp.json"):
if cfg.exists():
try:
configured |= set((json.loads(cfg.read_text()).get("mcpServers") or {}).keys())
except Exception:
pass
control("MCP config readable (>=1 server found somewhere)", bool(configured))
if tool_names and not any("mempal" in s.lower() for s in configured):
ln = next((i + 1 for i, l in enumerate(lines) if "kg_query" in l), None)
findings.append(
f"L{ln}: {len(tool_names)} named tools do not resolve — "
f"{', '.join(sorted(tool_names))} (no mempalace MCP server configured; "
f"servers present: {sorted(configured) or 'none'})"
)
# ------------------------------------------------- 3. hooks claimed to fire
hook_claims = re.findall(r"(Stop and PreCompact hooks fire|PreCompact hook)", text)
if hook_claims:
live: set[str] = set()
for cfg in (HOME / ".claude/settings.json", HOME / ".claude/settings.local.json"):
if cfg.exists():
try:
h = json.loads(cfg.read_text()).get("hooks") or {}
live |= {k for k, v in h.items() if v}
except Exception:
pass
control("hooks config readable (>=1 hook event configured)", bool(live))
missing = [e for e in ("Stop", "PreCompact") if e not in live]
if missing:
ln = next((i + 1 for i, l in enumerate(lines) if "PreCompact" in l), None)
findings.append(
f"L{ln}: document claims these hooks fire, but they are unconfigured — "
f"{', '.join(missing)} (configured: {sorted(live) or 'none'})"
)
# ---------------------------------------------------- 4. expired horizons
today = date.today()
horizon_re = re.compile(r"through\s+(?:the\s+)?end\s+of\s+(\w+)\s+(\d{4})", re.I)
control("horizon regex matches a known-present phrase",
bool(horizon_re.search(text)) or "through end of" not in text.lower())
for i, line in enumerate(lines, 1):
for m in horizon_re.finditer(line):
month, year = m.group(1).lower(), int(m.group(2))
if month in MONTHS:
mi = MONTHS.index(month) + 1
if (year, mi) < (today.year, today.month):
findings.append(
f"L{i}: horizon expired — \"{m.group(0)}\" "
f"(elapsed {(today.year - year) * 12 + today.month - mi} months)"
)
# ------------------------------------------------ 5. structural integrity
if not text.endswith("\n"):
findings.append(f"L{len(lines)}: no terminal newline "
f"(wc -l undercounts by 1 — breaks line-referenced patches)")
for i, line in enumerate(lines, 1):
# `||` inside a line carrying table pipes = two rows fused onto one line.
# Do NOT require the line to start with `|`: the real instance (L243) began
# mid-sentence, which is exactly how the fusion hides.
if "||" in line and line.count("|") >= 3:
findings.append(f"L{i}: table rows fused on one line (`||`) — row will not render")
if re.match(r"^\s+\|", line):
findings.append(f"L{i}: table row has stray leading whitespace — breaks the table")
if re.match(r"^#{2,4} .*\s+$", line) and i < len(lines) and lines[i].startswith("#"):
findings.append(f"L{i}: empty heading stub immediately followed by a heading "
f"— orphans the section beneath it")
# --------------------------------------------- 6. doctrine-ID integrity
# Doctrine units may carry a stable id in an HTML comment: `<!-- D:memory.check-first -->`.
# Invisible in prose, parseable by tools, and IN the canonical — so there is no second
# version to drift (L110). Skills cite an id instead of paraphrasing the rule; this
# section catches a citation whose doctrine has been reworded away, and duplicate ids.
# Dormant until the first id exists; the controls below run either way, so "nothing
# reported" means "checked and clean", not "never looked".
DOCTRINE_DEF = re.compile(r"<!--\s*(D:[a-z0-9][a-z0-9-]*\.[a-z0-9][a-z0-9-]*)\s*-->")
DOCTRINE_REF = re.compile(r"\b(D:[a-z0-9][a-z0-9-]*\.[a-z0-9][a-z0-9-]*)\b")
defined: dict[str, int] = {}
for i, line in enumerate(lines, 1):
for did in DOCTRINE_DEF.findall(line):
if did in defined:
findings.append(f"L{i}: duplicate doctrine id {did} "
f"(also L{defined[did]}) — a citation cannot resolve it")
else:
defined[did] = i
# Citing surface: the skills, which are the intended consumers. Deliberately NOT
# PENDING.md — drafts there legitimately quote ids that do not exist yet.
skills_dir = HOME / ".claude" / "skills"
scanned = 0
if skills_dir.is_dir():
for f in sorted(skills_dir.rglob("*.md")):
try:
if f.stat().st_size > 200_000:
continue
body = f.read_text(errors="replace")
except OSError:
continue
scanned += 1
for ref in sorted(set(DOCTRINE_REF.findall(body))):
if ref not in defined:
findings.append(
f"{f.relative_to(HOME)}: cites doctrine id {ref}, which "
f"~/CLAUDE.md does not define — the rule was reworded or removed")
control("doctrine-id parser detects a definition",
bool(DOCTRINE_DEF.findall("x <!-- D:memory.check-first --> y")))
control("doctrine-id parser detects a citation",
DOCTRINE_REF.findall("see D:memory.check-first here") == ["D:memory.check-first"])
control("doctrine-id parser rejects a non-id",
not DOCTRINE_DEF.findall("<!-- D:nodot -->")
and not DOCTRINE_REF.findall("D:NoDot.Caps"))
control("doctrine-id citing surface is reachable",
(not skills_dir.is_dir()) or scanned > 0)
# ------------------------------------------ 7. register integrity (REVIEWED)
# EARNED 2026-08-07, from a real loss. An amendment block was placed OVER the
# record it amends: the original `## REVIEWED-87 — PENDING-99 — …` entry was
# replaced by `## REVIEWED-87 — AMENDMENT 2026-08-07`, leaving the amendment's
# own `**Amends:** REVIEWED-87` line pointing at a record no longer in the file.
# The content was recoverable from git and the underlying jurist ruling was
# filed separately, so nothing was lost — but NOTHING DETECTED IT. It surfaced
# because a diff was read by hand, and the tell was a deletion count on what
# should have been a pure append.
#
# The class: a register whose entries can silently replace one another cannot be
# trusted to answer "what was ruled under N", which is the register's whole job.
# This is `removing-a-claim-is-not-removing-the-reliance` at the governance
# layer — the amendment's dependency on the original survived the original's
# removal, and became invisible.
#
# Detection only, like every check here: REVIEWED.md is [ESCALATE], the
# steward's hand (Constitutional Constraint #1).
REVIEWED_MD = HOME / "dotfiles" / "REVIEWED.md"
RE_HEAD = re.compile(r"^##\s+REVIEWED-(\d+)\s*[—-]\s*(.*)$", re.M)
RE_AMENDS = re.compile(r"\*\*Amends:\*\*\s*REVIEWED-(\d+)")
def register_findings(text: str, label: str) -> list[str]:
"""Every amendment must sit ALONGSIDE the record it amends, never replace it."""
out: list[str] = []
heads = RE_HEAD.findall(text)
originals = {n for n, rest in heads
if not rest.strip().upper().startswith("AMENDMENT")}
for n, rest in heads:
if rest.strip().upper().startswith("AMENDMENT") and n not in originals:
out.append(f"{label}: '## REVIEWED-{n} — AMENDMENT' exists with no "
f"un-amended REVIEWED-{n} entry — the amendment replaced "
f"the record it amends")
for n in sorted(set(RE_AMENDS.findall(text))):
if n not in originals:
out.append(f"{label}: a block declares '**Amends:** REVIEWED-{n}' but "
f"no REVIEWED-{n} entry exists in the file")
return out
reg_findings: list[str] = []
if REVIEWED_MD.exists():
reg_findings = register_findings(REVIEWED_MD.read_text(errors="replace"),
"REVIEWED.md")
# Controls. The third is the one that matters and is the lesson of the day:
# a check that has never fired on a known-bad input is unestablished, so the
# instrument is run against a synthetic reproduction of the actual failure.
_GOOD = ("## REVIEWED-87 — PENDING-99 — original\n**Date:** 2026-08-05\n\n"
"## REVIEGH\n\n## REVIEWED-87 — AMENDMENT 2026-08-07\n"
"**Amends:** REVIEWED-87 (x).\n")
_BAD = ("## REVIEWED-87 — AMENDMENT 2026-08-07\n"
"**Amends:** REVIEWED-87 (x).\n")
control("register parser finds REVIEWED headings", len(RE_HEAD.findall(_GOOD)) == 2)
control("register check passes a correctly JOINED amendment",
not register_findings(_GOOD, "t"))
control("register check DETECTS an amendment that replaced its record "
"[reproduces the 2026-08-07 loss]",
len(register_findings(_BAD, "t")) == 2)
control("register file is reachable", REVIEWED_MD.exists())
# ------------------------------------------ 8. deferred decisions, and their triggers
# EARNED 2026-08-07. The jurist settlement of 2026-05-16 deferred TEI-native
# authoring "until Cluster A's MD-with-sidecar form is operational". Cluster A
# became operational, the condition was met, and NOBODY LOOKED — it surfaced
# months later, by accident, while reading an unrelated document. The steward's
# stated reason for settling it that day was not the format question but the
# forgetting: "I abhor deferring so many things and then forgetting them."
#
# A deferral is a claim: "not yet". When its trigger fires, the substrate
# contradicts that claim — which is exactly what this instrument detects. So a
# deferred decision declares a MACHINE-CHECKABLE trigger and this checks it,
# rather than relying on anyone to remember.
#
# <!-- DEFERRED-DECISION: <slug>
# since: YYYY-MM-DD
# owner: steward | jurist | executor
# trigger: glob <pattern> | path-exists <path> | date <YYYY-MM-DD> | manual
# discriminator: <where the deciding evidence is written down> -->
#
# `manual` never auto-fires and is listed rather than checked — an honest way to
# record a deferral whose condition genuinely cannot be mechanised, instead of
# inventing a proxy. Proxies are what failed here: the old trigger stood in for
# "behavioural evidence on high-fidelity sources" and came true without it.
DEFERRED_RE = re.compile(
r"<!--\s*DEFERRED-DECISION:\s*([a-z0-9][a-z0-9-]*)\s*\n(.*?)-->", re.S)
# REVIEWED-127 / PENDING-158. A negative STATE-claim is the same sentence as a deferral
# about a different object: a deferral says "not yet" about a DECISION, a state-claim says
# "not yet" about a STATE. Same words, same forgetting, same substrate standing ready to
# contradict them — and until now only one of the two had a trigger.
#
# So this reuses trigger_fired() verbatim and inverts only what firing MEANS: for a
# deferral, fired = the decision is due; for a state-claim, fired = THE CLAIM IS FALSE.
#
# <!-- STATE-CLAIM: <slug>
# since: YYYY-MM-DD
# claims: <the negative claim, in words>
# falsified-by: glob <p> | path-exists <p> | date <d> | text-present <f> <s>
# | file-changed-since <commit> <path> | manual
# resolved: YYYY-MM-DD — <pointer> (optional; see resolution state below) -->
#
# ⚠ Opt-in, and the limit is the item's, not a caveat added here: a marker is used by
# authors who remember to mark their claims, which is the same population that would have
# caught the claim anyway. ADOPTION is the open question, not expressibility.
STATE_CLAIM_RE = re.compile(
r"<!--\s*STATE-CLAIM:\s*([a-z0-9][a-z0-9-]*)\s*\n(.*?)-->", re.S)
# REVIEWED-127 C1 / PENDING-157. The resolution state, which BOTH kinds carry.
#
# Earned 2026-08-25: a trigger came due, was correctly discharged, and the schema had no
# way to say so. The only options were deleting the block (losing the record) or renaming
# its key by hand (losing machine-checkability, and relying on the reflex the mechanism
# exists to replace). Renaming keys one at a time is how you live with a defect.
#
# `resolved:` must SAY WHAT DISCHARGED IT. Non-empty is not enough — an undocumented
# discharge records THAT a gate closed and not WHY, which six months on is no record at
# all. A resolved block whose pointer resolves to nothing is reported as a defect, in the
# same lane as an amendment that replaced the record it amends: both close over history.
state_claims: list[dict] = []
def pointer_resolves(text: str) -> bool:
"""True if the resolution names something that exists — a path, or a git object."""
for tok in re.findall(r"[A-Za-z0-9_./~-]{4,}", text or ""):
tok = tok.strip(".,;:")
if re.fullmatch(r"[0-9a-f]{7,40}", tok):
if subprocess.run(["git", "-C", str(HOME / "dotfiles"), "cat-file", "-e",
tok + "^{commit}"], capture_output=True).returncode == 0:
return True
cand = (Path(os.path.expanduser(tok)) if tok.startswith(("~", "/"))
else HOME / "dotfiles" / tok)
if cand.exists():
return True
return False
SCAN_ROOTS = [HOME / "_Dev", HOME / "dotfiles"]
TRANSCRIPTS = HOME / ".claude/projects/-Users-davidglidden"
# Governance packages live outside the */docs/** convention the scan was written for,
# so a deferral filed there was invisible to this check. Found 2026-08-07 while wiring
# PENDING-112's falsifier: the mechanism existed, and the one place it most needed to
# reach was the one place it did not look.
EXTRA_SCAN_GLOBS = [(HOME / "dotfiles", "claude/governance/**/*.md")]
deferrals: list[dict] = []
def _repo_root(p: Path) -> Path:
for parent in [p] + list(p.parents):
if (parent / ".git").exists():
return parent
return p.parent
def parse_deferrals(text: str, src: Path) -> list[dict]:
out = []
for slug, body in DEFERRED_RE.findall(text):
fields = dict(re.findall(r"^\s*([a-z-]+):\s*(.+?)\s*$", body, re.M))
out.append({"slug": slug, "file": src, "root": _repo_root(src),
"trigger": fields.get("trigger", "manual"),
"owner": fields.get("owner", "?"),
"since": fields.get("since", "?"),
"resolved": fields.get("resolved", "").strip()})
return out
def parse_state_claims(text: str, src: Path) -> list[dict]:
"""Same shape as a deferral; `falsified-by` is mapped onto `trigger` so the one
evaluator serves both. Firing means the CLAIM IS FALSE, not that a decision is due."""
out = []
for slug, body in STATE_CLAIM_RE.findall(text):
fields = dict(re.findall(r"^\s*([a-z-]+):\s*(.+?)\s*$", body, re.M))
out.append({"slug": slug, "file": src, "root": _repo_root(src),
"trigger": fields.get("falsified-by", "manual"),
"claims": fields.get("claims", "?"),
"since": fields.get("since", "?"),
"resolved": fields.get("resolved", "").strip()})
return out
def trigger_fired(d: dict) -> bool | None:
"""True = condition met (decision is due). None = not mechanically checkable."""
kind, _, arg = d["trigger"].partition(" ")
arg = arg.strip()
if kind == "glob":
return any(d["root"].glob(arg))
if kind == "path-exists":
return Path(os.path.expanduser(arg)).exists() if arg.startswith(("~", "/")) \
else (d["root"] / arg).exists()
if kind == "date":
return datetime.date.today().isoformat() >= arg
if kind == "text-present":
# Falsified by a string appearing somewhere — e.g. a hold claim falsified by the
# ruling that lifted it. Earned by trial 09: "the run is held" stayed true-looking
# for five days after REVIEWED-124 voided it, because nothing compared the two.
fp, _, needle = arg.partition(" ")
needle = needle.strip()
if not needle:
return None
target = (Path(os.path.expanduser(fp)) if fp.startswith(("~", "/"))
else d["root"] / fp)
try:
return needle in target.read_text(errors="replace")
except OSError:
return None
if kind == "file-changed-since":
# Falsified by a file having been edited since a named commit — e.g. "the filed
# rule not edited", which went false one hour after it was written.
commit, _, path = arg.partition(" ")
path = path.strip()
if not commit or not path:
return None
try:
r = subprocess.run(["git", "-C", str(d["root"]), "diff", "--quiet",
commit, "--", path], capture_output=True)
except OSError:
return None
return True if r.returncode == 1 else (False if r.returncode == 0 else None)
if kind == "transcripts":
# Session-count trigger. Added 2026-08-07 for PENDING-112's pre-registered
# 20-session falsifier, which the jurist required be BINDING rather than a
# disclosed intention. A date would have been a proxy — sessions run at wildly
# variable rates — and this block's own comment records that proxies are what
# failed last time. Counting transcripts encodes the real condition.
try:
return len(list(TRANSCRIPTS.glob("*.jsonl"))) >= int(arg)
except (ValueError, OSError):
return None
return None
_scan_targets = [(r, "*/docs/**/*.md") for r in SCAN_ROOTS] + EXTRA_SCAN_GLOBS
# PENDING-118, and the item's own option (1) is REFUTED by building it.
# The item said "widen the scan to ~/PENDING-archive.md — the checker already parses
# that exact format." It does not. The structured marker is an HTML comment
# <!-- DEFERRED-DECISION: slug ... -->, and there are ZERO of those in PENDING.md or in
# the archive. Their deferrals are PROSE — measured 2026-08-08: 53 and 26 occurrences of
# "defer*". Widening alone would scan two more files, find nothing, and report clean:
# a silent net built to close a blind spot, which is the failure this item describes.
#
# So the widening ships WITH its own honest limit. Structured blocks anywhere are now
# found; prose deferrals are COUNTED and reported as un-machine-readable, never as
# absent. Counting is not classifying — how many of those conditions have fired is a
# READING task, and it is reported as owed rather than silently skipped.
_scan_targets += [(HOME / "dotfiles", "PENDING.md"), (HOME / "dotfiles", "PENDING-archive.md")]
PROSE_DEFERRAL_RE = re.compile(r"defer(?:red|ral)", re.I)
# The size guard bounds the unbounded `**/*.md` globs. It must NOT apply to the two
# governance files, which were added to the scan on purpose and are the whole reason
# the scan reaches outside docs/ at all.
#
# 2026-08-17 [FIX]: it did apply, and PENDING.md had grown to 546,944 bytes — so every
# structured DEFERRED-DECISION block in the governance register was skipped, silently,
# by the checker built to stop deferrals from being silently missed. Worse than silent:
# the prose-deferral loop below has no size guard, so PENDING.md's prose count appeared
# in the report and made the file look examined. Found by placing a block under
# REVIEWED-123 cond. 2 and noticing the tracked count did not move.
#
# Both halves are fixed here: the named governance files are never size-skipped, and a
# skip is now REPORTED rather than swallowed — the third outcome of REVIEWED-104's
# doctrine, on the instrument whose whole subject is conditions nobody is watching.
GOVERNANCE_FILES = {HOME / "dotfiles" / "PENDING.md",
HOME / "dotfiles" / "PENDING-archive.md"}
_seen_files: set = set()
skipped_too_large: list = []
for root, pattern in _scan_targets:
if not root.is_dir():
continue
for f in root.glob(pattern):
if f in _seen_files:
continue
_seen_files.add(f)
try:
if f.stat().st_size > 400_000 and f not in GOVERNANCE_FILES:
skipped_too_large.append(f)
continue
body = f.read_text(errors="replace")
except OSError:
continue
if "DEFERRED-DECISION:" in body:
deferrals.extend(parse_deferrals(body, f))
if "STATE-CLAIM:" in body:
state_claims.extend(parse_state_claims(body, f))
# REVIEWED-127 C1. A resolved block is no longer due — but it is NOT dropped. It stays
# counted as a closed ledger, because a discharge that vanishes from the report is its own
# species of decay: the register would show three tracked deferrals and no evidence at all
# that a fourth had ever been answered.
resolved_deferrals = [d for d in deferrals if d["resolved"]]
open_deferrals = [d for d in deferrals if not d["resolved"]]
fired = [d for d in open_deferrals if trigger_fired(d) is True]
manual = [d for d in open_deferrals if trigger_fired(d) is None]
resolved_claims = [c for c in state_claims if c["resolved"]]
open_claims = [c for c in state_claims if not c["resolved"]]
falsified = [c for c in open_claims if trigger_fired(c) is True]
claims_manual = [c for c in open_claims if trigger_fired(c) is None]
# Same lane as the amendment-that-replaced-its-own-record check: both are a record
# closing over its own history.
dangling = [d for d in (resolved_deferrals + resolved_claims)
if not pointer_resolves(d["resolved"])]
_T_OK = ("<!-- DEFERRED-DECISION: tei-native\n since: 2026-08-07\n"
" owner: steward\n trigger: date 2000-01-01\n-->")
_T_WAIT = _T_OK.replace("date 2000-01-01", "date 2999-01-01")
_p_ok = parse_deferrals(_T_OK, CLAUDE_MD)
_p_wait = parse_deferrals(_T_WAIT, CLAUDE_MD)
control("deferred-decision parser reads a well-formed block",
len(_p_ok) == 1 and _p_ok[0]["slug"] == "tei-native")
control("deferred-decision parser rejects a non-block",
not parse_deferrals("<!-- DEFERRED: nope -->", CLAUDE_MD))
# transcripts-trigger controls: it must fire on a threshold already passed and stay
# silent on one that has not. An absence is not evidence until the instrument is shown
# capable of detecting presence — and this trigger carries a standing obligation.
control("transcripts trigger fires on a passed threshold",
trigger_fired({"trigger": "transcripts 1", "root": HOME}) is True)
control("transcripts trigger silent on an unreached threshold",
trigger_fired({"trigger": "transcripts 999999", "root": HOME}) is False)
control("governance dir is inside the deferral scan",
any("claude/governance" in str(p) for p in _seen_files)
or not (HOME / "dotfiles/claude/governance").is_dir())
# 2026-08-17: PENDING.md passed 400 KB and every structured block in it went unscanned,
# silently, while its prose count still appeared in the report. These controls are derived
# from the PROPERTY (is the register's block actually parsed?), not from the guard's own
# vocabulary — a control asking "does the guard work" would have passed throughout.
control("the governance register is actually scanned for blocks, whatever its size",
(HOME / "dotfiles/PENDING.md") in _seen_files
and (HOME / "dotfiles/PENDING.md") not in skipped_too_large)
control("a real block IN the register is parsed [the live instance, not a fixture]",
any(d["file"].name == "PENDING.md" for d in deferrals)
or "DEFERRED-DECISION:" not in (HOME / "dotfiles/PENDING.md").read_text(errors="replace"))
control("the size guard still applies to non-governance files [negative control — "
"the exemption must not become 'scan everything']",
(HOME / "dotfiles/PENDING.md") in GOVERNANCE_FILES
and (HOME / "dotfiles/CLAUDE.md") not in GOVERNANCE_FILES)
control("trigger evaluator FIRES on a met condition",
_p_ok and trigger_fired(_p_ok[0]) is True)
control("trigger evaluator does NOT fire on an unmet condition "
"[the discriminating half]",
_p_wait and trigger_fired(_p_wait[0]) is False)
control("deferred-decision scan surface is reachable",
any(r.is_dir() for r in SCAN_ROOTS))
# ---- REVIEWED-127: state-claims, the two new trigger kinds, and the resolution state.
# Every one carries its discriminating half. An absence is not evidence until the
# instrument is shown capable of detecting presence — and this whole item exists because
# an instrument that reads nothing reports exactly like one that finds nothing.
_SC = ("<!-- STATE-CLAIM: fixture\n since: 2026-08-25\n"
" claims: the pulse has not been fetched\n"
" falsified-by: date 2000-01-01\n-->")
_sc = parse_state_claims(_SC, CLAUDE_MD)
control("state-claim parser reads a well-formed block",
len(_sc) == 1 and _sc[0]["slug"] == "fixture"
and _sc[0]["claims"] == "the pulse has not been fetched")
control("state-claim parser rejects a non-block [negative control]",
not parse_state_claims("<!-- STATE: nope -->", CLAUDE_MD))
control("state-claim FIRES when its falsifier is met [= the claim is now false]",
_sc and trigger_fired(_sc[0]) is True)
control("state-claim SILENT when its falsifier is unmet [the discriminating half]",
trigger_fired(parse_state_claims(
_SC.replace("date 2000-01-01", "date 2999-01-01"), CLAUDE_MD)[0]) is False)
control("state-claim `manual` is listed, never fired",
trigger_fired(parse_state_claims(
_SC.replace("date 2000-01-01", "manual"), CLAUDE_MD)[0]) is None)
_DF = HOME / "dotfiles"
control("text-present FIRES on a string that is present",
trigger_fired({"trigger": "text-present PENDING.md PENDING-158", "root": _DF}) is True)
control("text-present SILENT on a string that is absent [negative control]",
trigger_fired({"trigger": "text-present PENDING.md zzz-not-in-this-file-zzz",
"root": _DF}) is False)
control("text-present returns None on an unreadable target [honest degradation]",
trigger_fired({"trigger": "text-present no/such/file.md x", "root": _DF}) is None)
control("file-changed-since FIRES on a file edited since the named commit",
trigger_fired({"trigger": "file-changed-since 5694b92 "
"claude/governance/fool/seed/FOOL-BONES-2026-08-25.md",
"root": _DF}) is True)
# ⚠ The fixture below is deliberately a FROZEN artifact — a trial-03 run output from
# 2026-08-02 that is never edited. The first version of this control pointed at
# FOOL-SEED-RULE.md and failed the moment that file was edited in the working tree, which
# is correct behaviour from the trigger and a badly chosen fixture: a control whose
# subject is "did this file change" must not point at a file the session is changing.
# Caught by the controls running on every invocation rather than in a separate suite.
control("file-changed-since SILENT on a file unchanged since HEAD [discriminating half]",
trigger_fired({"trigger": "file-changed-since HEAD "
"claude/governance/fool/runs/trial-03-20260802T144136Z.raw.txt",
"root": _DF}) is False)
_RES_OK = _SC.replace("-->", " resolved: 2026-08-25 — commit 5694b925\n-->")
_RES_BAD = _SC.replace("-->", " resolved: yes, done\n-->")
control("resolved block parses its resolution",
parse_state_claims(_RES_OK, CLAUDE_MD)[0]["resolved"].startswith("2026-08-25"))
control("a resolved block is EXCLUDED from due-ness even with a met trigger "
"[C1: discharge without a hand-rename]",
not [c for c in parse_state_claims(_RES_OK, CLAUDE_MD)
if not c["resolved"] and trigger_fired(c) is True])
control("the SAME block unresolved IS due [negative control — the exclusion must be "
"the resolution, not the fixture]",
trigger_fired(parse_state_claims(_SC, CLAUDE_MD)[0]) is True)
control("pointer that names something real resolves",
pointer_resolves("2026-08-25 — commit 5694b925"))
control("pointer that names nothing is DANGLING [C1: an undocumented discharge must be "
"impossible to express, not merely discouraged]",
not pointer_resolves("yes, done"))
control("a path pointer resolves too, not only a commit",
pointer_resolves("see claude/governance/fool/seed/FOOL-BONES-2026-08-25.md"))
# Prose-deferral census: counted, never classified, and never read as absence.
prose_counts = {}
for _f in (HOME / "dotfiles" / "PENDING.md", HOME / "dotfiles" / "PENDING-archive.md"):
if _f.exists():
prose_counts[_f.name] = len(PROSE_DEFERRAL_RE.findall(_f.read_text(errors="replace")))
control("prose-deferral counter finds a known-present phrase",
PROSE_DEFERRAL_RE.search("this was deferred pending recurrence") is not None)
control("prose-deferral counter does not fire on unrelated text",
PROSE_DEFERRAL_RE.search("the fleet is green") is None)
control("register files are inside the widened deferral scan",
any(str(t[1]).endswith("PENDING-archive.md") for t in _scan_targets))
# ------------------------------------------ 9. built without a ruling in the register
# EARNED 2026-08-08. Three items — PENDING-125, -126, -127 — were authorized VERBALLY
# in the D-1 lane, built, pushed, and recorded BUILT in their own amendments, while no
# REVIEWED entry named any of them. Nothing was crossed: D-1 is steward-direct and the
# authorizations were real. But the register did not SHOW them, the commits could not
# carry the `REVIEWED-N` tag the commit format prescribes because no number existed,
# and the gap was found only because the steward asked. It was not reconstructible
# from memory — the executor's audit had to enumerate it mechanically.
#
# Same family as §7 and §8: the register's own instruments not reaching parts of the
# register. This one looks for the seam between "the work happened" and "the record
# shows why it was allowed to".
#
# ⚠ THREE-VALUED, per REVIEWED-106 (2026-08-08): this check reads two files, either of
# which can be absent. `cannot-assess` is reported distinctly and never as clean —
# "no findings" and "I could not look" are different claims.
# Real path, not the ~/PENDING.md symlink — the same convention REVIEWED_MD uses.
PENDING_MD = HOME / "dotfiles" / "PENDING.md"
RE_PEND_HEAD = re.compile(r"^##\s+PENDING-(\d+)\s*[—-]\s*(.*)$", re.M)
RE_REV_FOR = re.compile(r"^##\s+REVIEWED-\d+\s*[—-]\s*PENDING-(\d+)\b", re.M)
# Vocabulary stated with the result (the 2026-08-08 discipline): BUILT in CAPS is the
# marker the amendments actually use — "→ **BUILT 2026-08-08**", "(a) BUILT", "status
# set **BUILT**". Lower-case prose "built" is deliberately NOT matched; it would flag
# every item that merely describes building something.
RE_BUILT = re.compile(r"\bBUILT\b")
def unruled_builds(pending_text: str, reviewed_text: str) -> list[str]:
"""PENDING items marked BUILT that no REVIEWED heading names."""
ruled = set(RE_REV_FOR.findall(reviewed_text))
out: list[str] = []
parts = RE_PEND_HEAD.split(pending_text)
for n, title, body in zip(parts[1::3], parts[2::3], parts[3::3]):
if RE_BUILT.search(title + body) and n not in ruled:
out.append(f"PENDING-{n} is marked BUILT and no REVIEWED entry names it "
f"— {title.strip()[:58]}")
return out
_P_GOOD = ("## PENDING-11 — ruled and built\nBUILT 2026-01-01, abc1234.\n"
"## PENDING-12 — open\nAwaiting steward authorization.\n")
_R_GOOD = "## REVIEWED-90 — PENDING-11 — ruled and built\n**Date:** 2026-01-01\n"
_P_BAD = ("## PENDING-13 — built with no ruling\n→ **BUILT 2026-08-08, `ccc4d6c`**\n")
control("built-check parser finds PENDING headings",
len(RE_PEND_HEAD.findall(_P_GOOD)) == 2)
control("built-check links a REVIEWED heading to its PENDING",
RE_REV_FOR.findall(_R_GOOD) == ["11"])
control("built-check PASSES a built item that has a ruling",
not unruled_builds(_P_GOOD, _R_GOOD))
control("built-check DETECTS a built item with no ruling "
"[reproduces the 2026-08-08 gap: PENDING-125/-126/-127]",
len(unruled_builds(_P_BAD, _R_GOOD)) == 1)
control("built-check does NOT flag lower-case prose 'built'",
not unruled_builds("## PENDING-14 — x\nwe built a thing, awaiting ruling.\n",
_R_GOOD))
build_findings: list[str] = []
build_assessable = PENDING_MD.exists() and REVIEWED_MD.exists()
if build_assessable:
build_findings = unruled_builds(PENDING_MD.read_text(errors="replace"),
REVIEWED_MD.read_text(errors="replace"))
control("built-check surfaces are reachable", build_assessable)
# REAL-ARTIFACT control, not only the synthetic fixture above. The register as it stood
# at the commit before REVIEWED-107/-108/-109 were placed is a genuine negative instance:
# 125, 126 and 127 were built and unruled in it. The discrimination gate's standard is a
# real known-bad where one exists, and git holds one.
try:
import subprocess
_old = subprocess.run(["git", "-C", str(HOME / "dotfiles"), "show", "HEAD:REVIEWED.md"],
capture_output=True, text=True, timeout=10).stdout
if _old and PENDING_MD.exists():
_hits = unruled_builds(PENDING_MD.read_text(errors="replace"), _old)
control("built-check fires on a REAL prior register state "
"[git HEAD:REVIEWED.md — a genuine known-bad, not a fixture]",
len(_hits) > 0 or bool(RE_REV_FOR.findall(_old)))
except Exception:
pass # git absent is not a check failure; the fixtures still ran
# ------------------------------------------------------------- report
failed_controls = [lbl for lbl, ok in controls if not ok]
if failed_controls:
print("⚠ drift-check: INSTRUMENT NOT VERIFIED — treat results as unestablished")
for lbl in failed_controls:
print(f" failed control: {lbl}")
print()
if not findings:
print(f"✓ governance drift-check: CLAUDE.md clean "
f"({len(controls)}/{len(controls)} controls passed, {len(paths)} paths verified)")
else:
print(f"⚑ governance drift-check: {len(findings)} claim(s) in ~/CLAUDE.md "
f"contradicted by substrate")
for f in findings:
print(f" {f}")
print("\n Correction requires [ESCALATE] (Constitutional Constraint #1). "
"This report is detection only.")
# Register integrity is reported SEPARATELY. Folding it into the count above
# would make that line's own claim false — it says "claim(s) in ~/CLAUDE.md",
# and these are findings about a different file.
if reg_findings:
print(f"\n⚑ register integrity: {len(reg_findings)} broken amendment link(s) "
f"in ~/REVIEWED.md")
for f in reg_findings:
print(f" {f}")
print("\n An amendment must sit alongside the record it amends, never replace it.")
print(" Correction requires [ESCALATE] — REVIEWED.md is the steward's hand.")
elif REVIEWED_MD.exists():
n_am = sum(1 for _, r in RE_HEAD.findall(REVIEWED_MD.read_text(errors="replace"))
if r.strip().upper().startswith("AMENDMENT"))
print(f"✓ register integrity: every amendment link resolves "
f"({n_am} amendment(s) checked)")
# Built-without-a-ruling. THREE-VALUED (REVIEWED-106): clean / findings / cannot-assess.
if not build_assessable:
missing = [str(f.relative_to(HOME)) for f in (PENDING_MD, REVIEWED_MD) if not f.exists()]
print(f"\n— built-vs-ruled: CANNOT ASSESS — unreachable: {', '.join(missing)}")
print(" Not a pass and not a finding. Nothing was checked.")
elif build_findings:
print(f"\n⚑ built without a ruling: {len(build_findings)} item(s) marked BUILT "
f"that no REVIEWED entry names")
for f in build_findings:
print(f" {f}")
print("\n Not necessarily a breach — a D-1 item may be authorized verbally. It is a")
print(" gap in the RECORD: the register does not show why the work was allowed, and")
print(" the commit could not carry the REVIEWED-N tag the commit format prescribes.")
print(" Remedy: draft the entry for steward placement.")
else:
_nb = len([1 for n, t, b in zip(*[RE_PEND_HEAD.split(PENDING_MD.read_text(errors="replace"))[i::3]
for i in (1, 2, 3)]) if RE_BUILT.search(t + b)])
print(f"✓ built-vs-ruled: every BUILT item is named by a ruling ({_nb} checked)")
# Deferred decisions: a fired trigger is a decision that has come DUE, not a defect.
if deferrals:
if fired:
print(f"\n⏰ deferred decisions: {len(fired)} of {len(open_deferrals)} open have COME DUE")
for d in fired:
print(f" {d['slug']} — owner: {d['owner']}, deferred since {d['since']}")
print(f" trigger MET: {d['trigger']}")
print(f" {d['file'].relative_to(HOME)}")
print("\n A deferral is the claim 'not yet'. These triggers say otherwise.")
else:
waiting = len(open_deferrals) - len(manual)
print(f"✓ deferred decisions: {len(open_deferrals)} tracked, none due "
f"({waiting} checkable, {len(manual)} manual-only)")
if resolved_deferrals:
print(f" ✓ plus {len(resolved_deferrals)} RESOLVED — answered, and kept in the "
f"ledger rather than dropped:")
for _r in resolved_deferrals:
print(f" {_r['slug']} — {_r['resolved']}")
if prose_counts:
_tot = sum(prose_counts.values())
print(f" ⚠ plus {_tot} PROSE deferral mention(s) in the register "
f"({', '.join(f'{k} {v}' for k, v in prose_counts.items())}) — these carry no")
print(" DEFERRED-DECISION block, so NO trigger is machine-checkable for any of them.")
print(" Counted, not classified. Whether any condition has fired is unestablished.")
if skipped_too_large:
print(f" ⚠ and {len(skipped_too_large)} file(s) were NOT SCANNED for structured blocks "
f"(over the 400 KB guard):")
for _s in skipped_too_large:
print(f" {_s.relative_to(HOME) if HOME in _s.parents else _s} "
f"({_s.stat().st_size:,} bytes)")
print(" Any DEFERRED-DECISION block in these is INERT. This is 'could not assess',")
print(" not 'nothing there' — the distinction REVIEWED-104 rules may not be collapsed.")
# State claims: a fired falsifier is not a decision coming due — it is a claim that has
# STOPPED BEING TRUE while still being written down as true.
if state_claims:
if falsified:
print(f"\n⚠ state claims: {len(falsified)} of {len(open_claims)} open are NOW FALSE")
for c in falsified:
print(f" {c['slug']} — claims: {c['claims']}")
print(f" FALSIFIED BY: {c['trigger']}")
print(f" {c['file'].relative_to(HOME)}")
print("\n A state-claim is the claim 'not yet' about a state, not a decision.")
print(" The substrate says otherwise. Correct the document, or resolve the block.")
else:
_ok = len(open_claims) - len(claims_manual)
print(f"✓ state claims: {len(open_claims)} tracked, none falsified "
f"({_ok} checkable, {len(claims_manual)} manual-only)")
if resolved_claims:
print(f" ✓ plus {len(resolved_claims)} RESOLVED — kept in the ledger.")
else:
# Silence here is NOT evidence of clean state-claims. Nothing has opted in yet, and
# PENDING-158 records why that is the open question: an opt-in marker is used by the
# authors who would have caught the claim anyway. Say so rather than print a tick.
print("· state claims: 0 marked. ⚠ NOT 'none stale' — nothing has opted in yet;")
print(" ~57 candidate negative-state claims are unmarked and unread (a grep, not a")
print(" census). Adoption is the open question, not expressibility (PENDING-158).")
if dangling:
print(f"\n⚠ register integrity: {len(dangling)} resolved block(s) name no pointer "
f"that resolves")
for _d in dangling:
print(f" {_d['slug']} — resolved: {_d['resolved'] or '(empty)'}")
print(" A discharge recording THAT a gate closed but not WHAT closed it is no record.")
sys.exit(0)