# Claims and evidence rules

The specification requires exact authority, complete mediation, current
validity, a consumed root budget, and an honest record of uncertain delivery.
The local implementation models and tests those rules within BOUNDARY.md.

A claim is supported only when its result names the exact manifest and
required cases. A manifest hash identifies bytes; it does not prove who
published them or prove that the claims are true.

| Claim | Evidence needed for this candidate |
| --- | --- |
| The core authority cases pass | 10 public, 20 regression, and 8 amendment cases, with effects and audit fields checked |
| Process recovery retains authority state | 32 lifecycle cases, including stale database, spent root, revocation, prepared handle, and process interruption cases |
| Different releases cannot reuse an old grant | 12 release/rollback gate cases; G02 checks rejection under separately approved different code |
| Uncertain time blocks safely | 16 clock cases, separating preflight denial from uncertain delivery |
| Signed receipts remain evidence only | 20 receipt cases using real Ed25519 fixture signatures and the declared input routes |
| Controller restart does not reopen a write | 8 controller cases, including killed processes and stored-result recovery |
| The four rc.3 review findings are addressed | 26 fixed review regression cases, including contradictory signed claims, commit admission, controller uncertainty and observed spent state |
| rc.4 input and evidence findings are addressed | 14 ported independent probes and 20 fixed result-binding cases; compare actual effects and context, not only counts |
| Exact elapsed and drift limits hold at both late samples | 30 clock-precision cases, including inclusive limits, one-nanosecond violations, both drift directions and large monotonic integers |
| rc.5 findings are addressed | 53 ported reviewer probes plus C08's valid control and isolated result-content defect; C08 must raise result validation, not wrapper rejection |
| Schemas describe actual results | 26 schema groups and validation of recorded outputs |

A blocked preflight leaves a prepared handle available for a fresh authority
check; it does not extend permission. An admitted commit removes that handle
from its transaction, and a definitive reply requires durable removal. An
uncertain transaction makes no claim either way. Controller results identify
one dispatch and separately retain earlier unresolved admissions. See R7 and
R15 of SPECIFICATION.md. These clarifications change the wire version to
0.1.0-draft.4 in rc.4; they do not retroactively settle the old wording.
Draft.5 adds the exact input and evidence checks in RESULTS.md and controller
binding/3. The rc.5 independent review supports those corrections within the explicit
checklist and tested paths. Draft.6 makes the exact clock arithmetic normative
and repairs C08's coverage. The new candidate needs its own independent review;
retained or ported tests are not a new independent implementation score.

Every required case must pass. Report suite counts separately; do not average
an authority violation into a percentage. A failed safety case makes this
candidate nonconformant with that tested requirement.

The fixed public tests are visible to implementers. Passing them is not a
hidden evaluation or evidence of independent authoring. The earlier Astra
trial and Mac deployment results refer to the bytes in the historical ZIP.
They are not transferred to this changed implementation.

Publication claims still require independent enforcement review, the owner's
signature and a separate website acceptance run. The acceptance record must
name the candidate digest, environment, inputs, operator help, output hashes,
and exact results. No result in this package establishes general production
readiness, arbitrary self-improvement containment, or whole-host rollback
resistance.

Draft.7 resolves IR6-001 by choosing durable cross-transaction wall history.
CLOCK-HISTORY.md and its required vectors govern that added claim. The 84 rc.6
reviewer probes are now public fixed regressions, not a fresh independent score.
The historical rc.6 review supported its rc.5 corrections but withheld an
unqualified all-written-rules verdict. This candidate needs its own review.


Rc.8 corrects IR7-001's comparative storage wording and release metadata.
It retains draft.7's explicit rules; no new enforcement semantics are chosen.
The rc.7 review supports its mapped corrections in tested scope and leaves
IR7-001 open in those frozen bytes. The correction in rc.8 needs its own exact
digest testing and independent review. Reused rc.7 reviewer probes are public
regression evidence, never a new independent score.


Rc.9 aligns unsigned-review verification procedures for IR8-001. It does not
change authority semantics or remove the signed-publication gate. The rc.8
review closed IR7-001 only for rc.8 and disclosed IR8-001 as nonblocking for its
explicit handoff. Rc.9 requires its own frozen tests and independent review.
The rc.8 Q8-001 driver is now retained public regression material; a replay is
not a new independent test, and must use newly issued release-bound fixtures.