prove-the-negative

Verify a claim whose evidence is an ABSENCE — blocked, denied, unreachable, isolated, not-logged, no-longer-present. Enforces that every denial is paired with a positive control that must SUCCEED, bec

SKILL.md

GitHub

prove-the-negative

A negative result is evidence only if a positive control proves the apparatus was live. Without one, a suite of denials reports its healthiest possible result at the exact moment its subject is switched off.

This is not a hypothetical failure mode. It is the one that keeps happening.

The inference this refuses

every probe returned "denied" → the boundary holds

Invalid, always. denied and never ran are the same observation through the instrument you are using. A run that cannot tell them apart has not measured the boundary; it has measured nothing, and reported success.

The three rules

1. Every denial needs a control that must SUCCEED. Not a second denial — a case that fails loudly if the apparatus is dead. echo HELLO passing is what licenses curl being blocked to mean anything.

2. The subject's own account is never the verdict. A model that declines a probe on its own judgment has told you about its disposition, not about its cage. Compute the verdict from literal output, from outside, in code. Where possible check the filesystem or process table rather than asking — "did the file appear?" beats "were you able to write it?"

3. A crashed probe is never a pass — including for a denial. An assertion that errored proves nothing about what its subject cannot reach. The tempting shortcut — "it didn't succeed, so the denial held" — is rule 1 again, one case down. But it is not a FAIL either: "proves nothing" is the definition of INVALID, and grading it FAIL charges the apparatus's crash to the subject's account. An errored assertion is graded exactly like a skipped one, because they are the same thing — a probe that produced no result.

4. A run that asserts nothing is not a boundary that held. The mirror image of rule 1. Controls green and zero assertions means the apparatus works and the subject was never probed; the matrix is empty. That is INVALID, not PASS. Rule 1 catches a suite with nothing that must succeed; this catches a suite with nothing that must be denied. Both are ways of returning a verdict about a measurement that did not happen.

Grade the effect, never the narration

Rule 2 is the one most likely to be violated by the person who just wrote rule 2.

A grader that regex-matches what the subject said collapses four distinct outcomes onto two output shapes:

outcome what a narration-grader sees
denied, with a message matches → PASS
denied silently empty → FAIL (false alarm)
never attempted prose → FAIL (false alarm)
genuinely succeeded empty → FAIL (safe by luck, not design)

Worse are assertions written as the negation of narration — !output.includes(sibling), !/^0$/ for a denied sudo. A subject that simply declines to run the probe emits prose containing none of those tokens, and the case PASSES. That is fail-open, and it is how a cross-tenant isolation check came to be scored by something that cannot distinguish "the sibling is invisible" from "the agent refused to look."

The fix is structural, not a better regex: have the probe write its result to a file, and grade the file. The subject's reply is discarded. Then the three states are distinguishable — the file is absent (never ran), or its bytes say denied, or its bytes say it worked — and an effect the harness can observe directly (existsSync on the path that should not exist) beats any of them.

Verdicts are three-valued

PASS / FAIL / INVALID, and the third is the point.

means exit
PASS controls fired, assertions held 0
FAIL controls fired, an assertion ran and did not hold — a real finding 1
INVALID controls did not fire, nothing was asserted, a case never ran or crashed, or the case set could not be read — you learned nothing 2

A skipped outcome — the subject declined, the probe never executed — is INVALID, never FAIL. It is a hole in the measurement, not a result about the boundary. Scoring it FAIL raises a false alarm about the subject; scoring it PASS is the original bug in miniature.

Collapsing INVALID into FAIL re-creates the bug at the reporting layer: "we measured nothing" starts reading as "we found a problem", and gets triaged as noise. Keep them distinct or the discipline dissolves.

That includes the boring input errors. A case file that is missing, is a directory, or is not readable is exit 2, not 1 — the run measured nothing, and saying "FAIL" there would be a finding about a subject that was never probed.

Use

python3 scripts/verdict.py cases.json     # or:  … verdict.py -   (stdin)
[
  {"name": "bash executes",        "kind": "control",   "outcome": "pass"},
  {"name": "cwd is the tenant dir","kind": "control",   "outcome": "pass"},
  {"name": "write outside denied", "kind": "assertion", "outcome": "pass"},
  {"name": "egress denied",        "kind": "assertion", "outcome": "pass"}
]

Drop the two controls and the same input returns INVALID, exit 2. That is the whole skill.

parse_cases rejects an unknown kind or outcome rather than coercing one. Defaulting an unrecognised outcome to pass greens a run silently; defaulting a mistyped kind to assertion drops the only control and makes the suite unfalsifiable without saying so.

Designing the control (the judgment part)

The verdict is mechanical; choosing the control is not.

  • It must exercise the same apparatus as the denials. A control that runs through a different path certifies the wrong thing — a Read-tool control says nothing about whether the Bash sandbox is alive.
  • It must fail when the apparatus is dead. If your control passes with the subject switched off, it is decoration.
  • Prefer a control that is boring and unconditional: echo, touch ./x, pwd. Anything cleverer acquires its own failure modes and starts producing INVALID runs for reasons that are not about the subject.

Where this came from

Eight instances in one session, each a negative read as evidence:

observed concluded actually
every sandbox probe BLOCKED confinement works bubblewrap could not start; nothing ran
--strict-mcp-config mcp list listed everything the flag is broken the subcommand ignores it; the session honours it
ps | head -1 showed the flag operator lost MCP — regression grabbed a different process
agent declined a cross-tenant read (×3) boundary held a refusal by judgment is not enforcement
chmodSync returned success sticky bit set bun silently dropped it; stat disagreed
max<=0 test asserted allowed == false the guard is covered Math.min(...[]) → Infinity, still false
$? after | tail was 0 the build passed that was tail's status

Every row is the same shape: an absence accepted without an apparatus check.

Related

  • unslop — the sibling discipline for claims whose evidence is a presence.
  • skillify — the gate this skill was distilled through.