ltl-accumulator-verified/KNOWN-GAPS.md
mrwulf 172a1d0653 Round 5 (housekeeping): doc-consistency welded into the button; both round-4 approvals recorded
Round-4 verdicts: Claude reviewer — nothing blocks the freeze, no
remaining findings; GPT-5.6 — approve after minor documentation fixes,
attestation scoped to the mechanized model. This round is those fixes;
no Lean surface changed.

- 218/59 → 222/61 everywhere, and STRUCTURALLY: check.sh Phase 3c
  asserts the audit counts (STATEMENT-MAP + README vs allowlist/CONES)
  and the four fidelity pins (STATEMENT-MAP vs run_fidelity.py
  constants) on every run — stale-count drift is a red button now
  (R4-1, third recurrence of the class).
- Gap 14 reworded to evidence-vs-inference (the invariant "is assumed",
  not "transfers"), witnesses cited (paper §5.3/§5.4; pacta
  sthstore.py/logclient.py — outside the fidelity target). New gap 15:
  deployment refinement invariant unmechanized (GPT's principal
  finding, split out because it carries the deployed-soundness claim).
- Runbook: A1 marked done (both approvals on SD); B2 gains the REQUIRED
  scoped attestation wording (GPT §11) as a gate condition — entry 13
  cannot claim "deployed verifier formally verified".
- run_bare.sh fail-closes on Lean version AND commit (rejection path
  tested with a fake toolchain: FATAL, exit 1).
- Harness: "consistency baseline family" line (GPT §8); gap 14 says
  "fixed offsets n−1/n+1/n+7" (R4-5).
- RESPONSE round 5, incl. refutation of GPT §7 (the target tarball
  demonstrably contains MANIFEST.sha256 + TARGET-PROVENANCE.md; the
  round-5 kit also ships both unpacked as a courtesy).

check.sh exit 0 ATTESTATION GREEN (Phases 0-4 incl. new 3c); selftest
exit 0, 9/9 + control. Live LTL untouched (12 leaves, bcd15f9d…).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-15 09:40:20 +02:00

128 lines
8.3 KiB
Markdown
Raw Blame History

This file contains ambiguous Unicode characters

This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.

# Known gaps and scope boundaries (honest ledger)
Deliberate, documented, and none silent. Reviewers should verify this
list is COMPLETE, not merely that the items are acceptable.
1. **SHA-256 is opaque** — the single boundary axiom (`LTLAcc.sha256`),
by design identical to the paper's posture: soundness theorems
construct collisions, never assume collision resistance.
2. **No consistency-completeness theorem** (honest ConsRec acceptance).
Matches the paper (its Theorem 1 is inclusion-only); honest
consistency behavior is covered by the fidelity harness's honest
cases (within the 230,016-case consistency agreement with the
deployed verifier).
3. **Lemma 2 is mechanized as specializations, not as one general
theorem.** The paper's Lemma 2 is a single statement quantified over
an abstract hash-fold `F` and a connected subtree `S`. The corpus has
no hash-fold datatype/predicate; it proves the needed instances
directly — whole-tree (`extractMTH_correct`), ConsRec
(`consRecBinding`), inclusion (`extractIncl_correct`), and the width
fact (`hnode_preimage_inj`). These suffice for Theorems 23. Two
consequences: (a) the abstract lemma itself is not a mechanized
object; (b) the path-instance receipt-uniqueness for `Root` (removed
with the vacuous `root_binding`) is not restored — optional, unused.
Any paper claim that "Lemma 2 is mechanized" must read "its
specializations sufficient for Theorems 23 are mechanized."
4. **Signature layer abstract** — Ed25519 EUF-CMA, the poison/evidence
retention state, and transferability of fork evidence (paper Prop
1(2)) are not modeled; `fork_distinct` is the Merkle-layer share only.
5. **Transliteration bridge**`fidelity/lean_defs.py` mirrors the Lean
definitions by quoted-source inspection (the Lean defs are
noncomputable over the opaque hash, so the bridge cannot be #eval'd
closed). Same inspection bridge the paper's own harness uses.
6. **Proposition 2 (verdict integrity) out of scope** — per paper §10's
mechanization list (iv). It is a property of the consumer tooling's
construction, enforced and regression-tested in the pacta repo.
7. **Multi-step pin monotonicity** — mechanized per-step
(`pinAccept_monotone`); the paper's multi-step chain is its
reflexive-transitive iterate, not separately mechanized.
8. **Process history** (candor): three cone pins were guessed (not read)
during S5.3S6 and the audit's failure went unnoticed until S7
because green was claimed from tailed output rather than the exit
code. No theorem was affected (kernel-checked throughout); pins were
corrected, the audit surface defined, and the standing rule is now:
exit code + ALL GREEN, cones read from #print axioms only.
9. **Asymptotic cost not mechanized.** Paper Theorems 2 and 3 assert the
extractors run in `O(n)` / `O(n₁)` hash evaluations. The mechanization
proves functional correctness of the named extractors only — no cost
semantics, recurrence, or computability-after-hash-instantiation. (The
extractors are `noncomputable` over the opaque `sha256`.)
10. **Pin-store initialization from the empty pin not modeled**, and
`pin_prefix_correct` assumes `0 < n`. Trust-on-first-use / the size-0
initial state is a separate operation; the theorems cover transitions
from a positive-size pin. (Related to gap 7's per-step scoping.)
11. **acceptIncl now named (was review F1).** The consumer's inclusion
acceptance `m < n ∧ Root … = some r` is now the Lean object
`acceptIncl`, with `acceptIncl_complete`/`acceptIncl_sound` routing
completeness/soundness through it, and the fidelity harness exercises
the out-of-range families (`m ≥ n`). `Root` alone still accepts
out-of-range `m`; that is by design (it is the reconstruction, not the
accept predicate).
12. **Audit-gate lineage** (candor; was round-2 GPT H1 / Claude NEW-1,
both round-1 "fail-closed" claims were overclaims). The round-2
coverage gate enumerated declarations with a source regex and was
evadable (attributes, indentation, private/protected, `instance`,
nested-namespace basename collisions). Round 3 replaced it with an
environment-derived inventory (`Proofs/Inventory.lean` +
`inventory-allowlist.txt`, fully-qualified names, no filtering) and
`selftest_audit.sh`, which runs the published evasion table plus a
namespace collision, an axiom smuggle, a stale-entry case, and two
unmanifested-module cases against the exact production gate.
Residual honesty: the inventory sees what the compiled environment
contains; it cannot see source that is never compiled (which the
dead-file checks cover) or defeat a hostile Lean toolchain.
13. **Review-kit fidelity target was not self-contained in round 2**
(GPT H2: missing load-time imports made `run_fidelity.py` unrunnable
from the kit). Round 3 ships the complete stdlib-only import closure
of `pacta.transparency`, content-addressed against pacta commit
`3d81d53`, plus the clean-extraction transcript with exit code.
14. **Deployed `verify_consistency` accepts strictly more than the
mechanized `ConsRec` on lied-size inputs** (round-3 Claude addendum
F1*, reproduced by the operator against deployed pacta). Witness:
for the honest proof P between sizes 2→3, `verify_consistency(1, 3,
R2, R3, P)` returns True — a semantically false claim ("R2 is the
root of a size-1 prefix") — while `ConsRec` rejects; 3,405 such
divergences exist for n < 60, ALL one-sided (the mechanized model
never accepts anything the deployed verifier rejects; inclusion
shows zero divergences under identical abuse). Mechanism: when the
claimed old size is a power of two, the deployed RFC 9162 iterative
algorithm seeds the walk with the old root and uses the sizes only
as bit-navigation state, so several size claims navigate one proof
identically. Consequences: (a) fidelity between the two consistency
verifiers is agreement over the pinned case families, NOT
extensional equality the harness's lied-size family pins the
boundary (73,573 cases: lied old size exhaustive for n < 60, lied
new size at fixed offsets n1/n+1/n+7; 3,867 expected divergences,
direction asserted one-sided per case); (b) Theorem 3 /
`acceptCons_sound` cover the MECHANIZED accept set. The exhibited
divergence is outside the intended pin-store input invariant;
applying the mechanized soundness result to the deployed flow
therefore additionally ASSUMES that the deployed state machine
always binds each root to its authentic size and exposes no
alternate invocation path an invariant that is NOT mechanized in
this corpus (gap 15). Where the invariant's witnesses live: paper
§5.3 (the signed head binds `(n₁, r₁)` together under one
signature) and §5.4 (the pin `(n₀, r₀)` comes from the consumer's
own store, never from the peer), implemented in the pacta repo at
`src/pacta/sthstore.py` and `src/pacta/logclient.py` code OUTSIDE
the supplied fidelity target (review R4-3/GPT-4). No exploitability
against that flow is claimed or ruled out here; assessing it
requires the signature/STH layer (gap 4). No pacta code change is
made (deployed behavior matches upstream RFC 9162 implementations).
15. **Deployment refinement invariant unmechanized** (round-4 GPT, its
principal finding split out from gap 14 because it carries the
deployed-soundness claim). The corpus proves soundness of the
mechanized `acceptCons`; it does NOT prove the refinement
`AuthenticPair(n₀,r₀,D₀) ∧ AuthenticPair(n₁,r₁,D₁) ∧
verify_consistency(…) → acceptCons(…)`. The operational invariant
(authentic-size/root binding via the signed-head + pin-store flow)
is relied upon but unverified, and the consumer flow implementing it
is not in the supplied fidelity target. Consequently any attestation
of this corpus must be scoped to the MECHANIZED model: "the deployed
consistency verifier is formally verified" is NOT a claim this
corpus supports. Closure paths (roadmap, operator decision, post
paper-freeze): (A) make the deployed verifier adopt
ConsRec-equivalent acceptance; or (B) mechanize the signed-head +
pin-store state machine and prove the refinement; or (C) keep the
boundary and this scoped claim permanently.