GPT-5.6 Adjudication — R10-pre Execution v1

Date: 2026-09-02 (America/Chicago) Input: fable51r10preexecutionv1.zip Disposition: SCIENTIFIC RESULTS GO. PACKAGE v2.3 REVISE BEFORE CANONICALIZATION. NO NEW FABLE THEORY ROUND.

0. Independent verification performed

I unpacked the return and independently checked the evidence surfaces that are self-contained in the bundle.

Git execution lane

invariant-r10.bundle clones successfully.

Observed commit chain:

5e2e5d4467da28f8bd57617dcbdc1df000bc6337
    R10-pre execution lane: v2.2 package as received
    tree 776f3f494c64611627f86b39e0dc866739297629

808c522e213696151b7cbd707c203d99a4fb6976
    R10-pre execution: steps 1,2,4,5; EMB-E1 not executed
    tree aa0e6e8f5bef22d962e9e7970b3e3a566343469d

02cfc5ed831404970c3c33582331fa9e01203826
    START_HERE / README / CHANGELOG follow-up

The before/after identities stated in START_HERE_GPT.md are real objects in the bundle.

v2.3 drift gate

drift_gate: scanned 10 current-facing files; 47 crosswalk records
DRIFT GATE PASS

The gate itself passes, but §8 below records semantic contradictions it does not detect.

Original Fable finite witness

Re-run under:

PYTHONHASHSEED=0
PYTHONHASHSEED=1
PYTHONHASHSEED=42
PYTHONHASHSEED=123

Each run:

19 expectations, 0 failed

Each stdout SHA-256:

ad23cf4b178c2880da78abe22712c500674a80de5d3b1133b0b80453b6c4bcfc

EXP-8 Micro-Locus witness

Independently re-run:

7 expectations, 0 failed

It reproduces:

RAM-only                 duplicate unreachable, continuity not promised
copyable flash           duplicate reachable
compound NV counter      duplicate unreachable
fence-only               duplicate reachable
head-only                duplicate reachable

This independently supports the v2.1 correction:

the model distinguishes enforcement properties, not implementation-mechanism count.

Integrity caveat

All substantive files checked by the top-level checksum manifest matched except the checksum manifest's own self-entry.

Recorded self-hash:

c0f1c35950fe7a943d05776ac493087751948b8084bbbf3de1b29c62b9ce5974

Actual current SHA256SUMS hash:

7d7ad05ce04ac0f27939a1a30112c2e765f81841c88422897e2d46abbd5bbafe

This is a packaging error: a checksum manifest should not attempt to authenticate itself in this form.

The PX13 handback-SHA256SUMS.txt also references 66 raw files that are not included in this execution bundle (individual run logs, original receipts, runner files, source bundle, transcripts, etc.). The selected final receipt / series / negative-control extracts are present, but the stated original handback cannot be reverified from this zip alone.

Therefore:

execution evidence: usable
return bundle: not fully self-contained for the complete PX13 handback

1. S6? ruling — ACCEPT the reduction to S1

ACCEPT:

S6? World/Locus birth uniqueness
    ↓
S1(domain = World locus-birth/currentness)

The experiment materially improves the claim.

Against the frozen cd-core API, two births of the same LocusId under one WorldRoot:

both begin;
both attach different carriers;
both execute authority-bearing detach transitions;
both histories individually audit under the same root;
a warden minted by birth A can finalize birth B's live attachment.

That last witness is particularly strong: the existing per-birth capability vocabulary cannot distinguish the two births because the birth domain itself was never admitted.

The scratch WorldAdmission remedy has S1's shape:

one linear admission value for the domain
consumed by every birth
duplicate birth refused
failure returns the admission for retry
reuse refused by the type system

The properties left open by the remedy are already classified elsewhere:

durability / copy / restart of WorldAdmission -> S3
cross-world ID collision                      -> representation / identity

I do not see a residual semantic property that requires a distinct S6.

Required package change

The current frontier should no longer be described as:

six semantic obligations S1–S6?

Use:

five working semantic obligations S1–S5
+
S1 scoped corollaries, including World locus-birth/currentness

No claim is made that S1–S5 are an irreducible complete basis.

2. EXP-4 finality/liveness gap — REAL, but do not create S6/S7

EXP-4 surfaced a genuine missing dimension:

what guarantees progress when the current holder becomes permanently unavailable?

S1–S5 are overwhelmingly safety / integrity obligations. None says a nonterminal Locus must ever make progress after holder loss.

That is not evidence for another safety invariant.

Create a separate liveness axis.

Recommended working candidate:

L1 — Recoverable succession / explicit terminalization

For a Locus whose policy declares continued service,
if the current holder becomes permanently unavailable and sufficient
admissible loss evidence is eventually established, the system eventually:

    A. admits exactly one successor in a later generation,
       OR
    B. moves the Locus to an explicit terminal/quarantined state.

It must not remain silently ambiguous forever.

Important:

L1 is profile/policy-scoped.
A system may intentionally choose halt-on-loss and therefore not promise succession.
L1 must not weaken S1 to obtain progress.
Warden/finality machinery appears to be one implementation of this liveness cost.
Do not promote L1 as irreducible yet.

Recommended registry treatment:

semantic_safety: S1–S5
semantic_liveness: L1? ...

Do not mix them merely to preserve one numbered table.

3. EXP-1 ARM B ruling — TESTED under the narrow trust profile, not a broad cross-host theorem

The ARM B receipts are strong for what they actually test.

Both winner orderings satisfy:

exactly_one_advance            true
refusal_names_mechanism        true
witness_head_binds_winner      true
loser_quarantined_at_reopen    true

The crash matrix also shows the witness CAS placement is compatible with the tested local roll-forward protocol.

Therefore I accept:

S3 currentness is TESTED under a single correct, available, linearizable witness for two independently copied controller/store namespaces.

I do not accept the unqualified phrase:

S3 cross-host: TESTED

because the reported "host 2" is a container on the same PX13 physical machine, using host networking, and the witness is also on that machine.

This is a good distributed-protocol experiment, but it has not tested:

two physical machines;
network partition between controller and witness;
witness crash/restart during competing admission;
witness durable-storage corruption;
witness failover;
quorum disagreement;
physical-host failure independence.

Correct evidence language

Prefer:

TESTED:
    two-controller / two-filesystem-namespace copied-world topology,
    single correct and available witness

SUPPORTED by mechanism shape:
    cross-machine safety under the same witness assumptions

UNTESTED:
    failure-tolerant / partition-tolerant distributed witness service

A quorum witness is not required to establish safety under the explicit single-witness trust profile.

A quorum/replicated witness is required before claiming fault-tolerant availability or a realistic distributed-currentness service.

ARM B already demonstrates the semantic point:

moving the authoritative currentness CAS outside both copies can prevent both descendants from advancing.

That is enough for this R10-pre reduction.

4. E-13 / E-14 / E-15 — current `live_local` language is too generous

E-13 — exclusivity outside the copyable store

At efa8881, F31 demonstrates:

Root::lock_path(path) = "{caller-chosen-root}.lock"

Copying the root to another pathname creates another lock namespace and therefore another current controller for the same Locus.

That falsifies the broad intended mechanism:

the currentness anchor is outside the claimant-copyable / claimant-selectable namespace.

Same-path locking still works. But that is a narrower fact.

Recommended split:

E-13a same-anchor incarnation exclusivity
    TESTED / live_local
    scope: one canonical lock object / namespace

E-13b canonical host-local Locus currentness anchor
    FALSIFIED at efa8881 by F31
    R0.8.4 repair may close it, but this package does not contain
    the full R0.8.4 source+receipt needed to adjudicate that closure

E-13c witnessed multi-namespace currentness
    TESTED under the ARM-B single-witness profile

Do not let live_local (narrowed) obscure that the original load-bearing claim failed under a pathname-copy attack.

E-14 — root commitment under injective encoding

F23/F24 show that the committed material was not injective over the semantic interpretation of the SQLite store:

schema/collation can change reads without changing covered row values;
a trigger can mutate semantics from inside the honest writer transaction;
unordered reads can change authority resolution while the digest is unchanged.

Therefore the broad E-14 mechanism is not live_local under the attacker-writable-store threat profile.

Correct classification:

TESTED-CONDITIONAL:
    row-value commitment under a fixed/trusted SQLite interpretation contract

FALSIFIED / OPEN:
    semantic-store commitment when schema/interpreter metadata is attacker-modifiable

If the SQLite schema is part of the attacker-writable database, calling it "trusted" simply moves the open obligation into the trust profile; it does not close S3.

E-15 — mutation-time canonicality

F23's trigger case also weakens E-15.

The transaction may be serialized correctly while an attacker-installed schema trigger changes canonical rows from inside that same transaction.

Therefore E-15 should also say:

TESTED under fixed/trusted schema/interpreter

until schema semantics are committed/verified.

5. PX13 R0.8.3 ruling

I accept the following distinction:

R0.8.3 execution happened:               YES
21-run series was reported green:        YES
historical receipt should be preserved:  YES
R0.8 semantic layer may freeze:           NO

The green series does not imply all its claims were actually measured.

F25 is especially important:

a control ran, but never reached the subject it claimed to attack.

That means:

PASS != evidence
unless the control is capable of producing FAIL for the claimed defect.

This should become a general factory requirement.

Recommended meta-rule:

FAC-CONTROL-SENSITIVITY

Every negative control used to support an evidence-class promotion must
carry either:

    A. a demonstrated pre-fix / injected-fault FAIL,
    OR
    B. an independently witnessed mutation showing the control reaches
       the claimed subject and distinguishes the forbidden state.

A green-only control with no demonstrated failure sensitivity cannot
promote a semantic claim.

This may be more important than any single R0.8 patch.

F22/F23/F25/F31 are not reasons to discard the R0.8 work. They are exactly the reason the execution round was valuable.

6. E-42 / E-43 — hand both to the TRVM lane, keep red

E-43 `world.scope-registry-versioned@1`

This is directly an S4 adequacy question.

Observed:

registerQuery is write-guarded
but scope redefinition does not advance world version
two worlds equal under (vclock, log, resources)
can observe different scope digests

Either:

A. scope registration/redefinition is a world transition and must be versioned

or the semantic contract must explicitly say:

B. world version identifies resource state only,
   and scope-registry identity is a separate observable coordinate

Until the TRVM lane chooses and encodes that contract, keep the candidate law red.

E-42 `warrant.lineage-bound@1`

Hand this to the TRVM lane as well, but revise the crosswalk classification.

The observed problem is:

parent + fork
same version
different value
freshness says fresh
replay says value-mismatch

That is first an S4 / representation-adequacy problem: a warrant footprint is not sufficient to identify a state across lineages.

It is S3 only when the fork is being treated as a restored/copy descendant that is required to share one current lineage.

If fork() intentionally creates a distinct world, then the repair is lineage/world identity in the warrant, not anti-resurrection.

Recommended classification:

E-42:
    semantic_obligation = S4
    relation = representation/corollary
    cross-reference S3 when fork preserves the same semantic identity

Candidate law name is reasonable:

warrant.lineage-bound@1

but let the TRVM law lane adjudicate the exact statement.

7. EXP-5 authority merge — split the result instead of assigning all of it to S2

The experiment is useful and the core reworded rule survives:

do not blindly union effective authority; use a proven join-safe representation or merge evidence and rederive.

But E-46 currently groups three different phenomena:

A. Revoked/spent authority resurrected by union

This is an S2 non-amplification / current-authority problem.

Good S2 evidence.

B. Grants from separate replicas combine into a conjunction neither replica locally held

This is not automatically authority amplification.

If both grants are globally valid and the policy intentionally allows conjunction, then the merged authority is justified by two admitted grant chains.

The real question is whether the policy/merge contract permits composition, not whether one replica previously held the conjunction.

Do not use this witness by itself as evidence of S2 failure.

C. A stale copied Authority value acts after revocation

That is S1 at the action target / consumption point.

Therefore split E-46 conceptually:

merge representation law / join-safety     -> CROSS / representation
revocation or one-shot resurrection         -> S2
stale-true action after snapshot             -> S1

This will keep the cross-registry map from making S2 absorb every authority-related phenomenon.

8. v2.3 package anti-drift regression — BLOCK canonicalization until patched

The drift gate prints PASS, but the current-facing package contradicts its own execution results.

Examples:

README

Still says:

working map is six semantic obligations ... S6? candidate

and says the crosswalk has:

41 records

when v2.3 actually has 47.

`12_R10PRE_SYNTHESIS.md`

Its primary table still says:

S6? candidate
predicted, not executed

while §6 later says S6 resolved to S1.

It also retains pre-execution statuses for PX13, EXP-3, EXP-1, etc. in its main map.

`11_R10PRE_OPUS_HANDOFF.md`

Still presents the old execution plan as future work and begins with:

six working semantic obligations ... S6? candidate

`INV_FRONTIER_R10_PRE.md`

It correctly records S6 execution at §4 line 43 and then, only a few lines later, still says:

predicted reachable ...
has not been executed here ...

That is a direct semantic contradiction.

Required v2.4 patch

Do not ask Fable for theory.

Have the codebase agent produce package-v2.4/ that:

makes S1–S5 the current working obligation map;
moves S6 to an explicit S1 corollary;
updates every primary table/status from pre-execution to executed evidence;
changes crosswalk count 41 → 47;
removes the stale unexecuted S6 paragraph;
changes EXP-1 wording from broad cross-host TESTED to the exact trust/topology profile;
revises/splits E-13/E-14/E-15 as above;
reclassifies E-42 toward S4/representation;
splits E-46's S2 / S1 / representation meanings;
adds a separate liveness candidate section for L1?, not an S-row;
strengthens the drift gate to assert post-execution statuses, not only doctrine vocabulary.

Suggested positive gate assertions:

README current map contains S1..S5 and does NOT call S6 an obligation
README crosswalk count == JSON record count

S6 current surface contains:
    TESTED
    S1(domain = world locus-birth)
and does NOT contain:
    predicted reachable
    has not been executed

S3 distributed surface contains:
    single correct available witness
    two-controller/two-namespace
and does NOT claim:
    fault-tolerant cross-host
unless a corresponding receipt exists

PX13 surface must distinguish:
    execution happened
    R0.8 remains OPEN

9. Packaging fixes

Top-level checksum manifest

Do not include SHA256SUMS in its own manifest.

Generate:

find ... ! -name SHA256SUMS -type f -print0 | sort -z | xargs -0 sha256sum > SHA256SUMS

and publish the archive hash separately.

PX13 handback completeness

Either:

A. include the complete original handback tree referenced by
   handback-SHA256SUMS.txt

or:

B. replace it with a new selected-evidence checksum manifest whose every
   referenced file actually ships in the return bundle.

Do not say "raw lab logs" if only the series summary and selected extracts ship.

Reproducibility dependencies

EXP-3 and the ampd half of EXP-5 carry source code + captured outputs but not all pinned external runtime artifacts needed to reproduce them from this zip alone.

For the next review bundle, either vendor:

TRVM trvm_world.mjs @ fd0df4c
ampd beam/source @ 6e7039a

or state clearly:

EXECUTED receipt included
INDEPENDENT REPRODUCTION requires external pinned artifact

10. EMB-E1

Correctly not executed.

No issue.

It remains operator-gated by the frozen protocol:

EMB_E1_EXECUTE=YES
EMB_E1_MAX_USD >= required budget
OPENROUTER_API_KEY

Do not infer spend authorization from possession of a key.

At this point EMB-E1 remains one of the highest-information open experiments because it attacks an orthogonal question:

can capability survive transient-model replacement through persistent environmental support?

Run it only when the operator explicitly authorizes spend.

11. Current frontier after this execution return

The best current map is now:

SAFETY / SEMANTIC INTEGRITY

S1  Admission singularity
    includes:
      World locus-birth uniqueness as a scoped corollary

S2  Authority non-amplification

S3  Continuity / anti-resurrection
    enforcement properties:
      currentness / exclusivity
      canonical-history continuity
    mechanism count substrate-dependent

S4  Semantic observability
    adequacy + equivariance
    side-effecting resource observation TESTED as write-mediated
    scope registry + fork lineage still red

S5  External-effect honesty
    still unwitnessed by a real adapter

And separately:

LIVENESS

L1? Recoverable succession / explicit terminalization
    surfaced by EXP-4
    not promoted

Cross-cutting / representation rules should remain outside S1–S5 where appropriate.

12. Recommended next codebase-agent sequence

A. First: package v2.4 anti-drift patch

No new theory. Apply the adjudication above.

B. R0.8 repair / re-execution

Priority:

F31 canonical host-local anchor identity
F23 commit the SQLite interpretation contract
F22 safe root-path persistence / trusted directory boundary
F25 replace vacuous control with a control that reaches its subject
F26/F27/F28 receipt/control sensitivity cleanup

Then execute a new battery where each load-bearing negative control has demonstrated failure sensitivity.

Do not rewrite the historical R0.8.3 receipts.

C. TRVM lane

Take E-42 / E-43 as kept-red candidate laws and adjudicate their exact contracts.

D. EMB-E1

Operator decision / spend only.

E. S5 adapter

After R0.8/TRVM stabilization, a real external-effect adapter run is the largest remaining unexecuted S-row.

F. L1?

Formalize a tiny liveness model before adding production machinery. Determine whether the policy promises succession, terminalization, or deliberate halt.

13. Final ruling

The execution round succeeded at its real purpose: it changed the map.

Most important results:

S6 collapses to S1 under real cd-core execution.
Locus does not reduce to authority, and EXP-4 identifies history/provenance/custody/finality residue.
A distinct liveness axis is now visible.
Side-effecting observation is write-mediated in the tested TRVM path.
Blind authority union fails on real APIs; join-safe representation and action admission are separate concerns.
A single external witness can enforce both S3 currentness/history obligations for competing copies under its explicit trust profile.
R0.8.3's green series exposed controls that could not fail, validating the rule that receipts themselves require adversarial sensitivity.
R0.8 remains OPEN because currentness anchor identity, SQLite interpretation commitment, and filesystem-root trust are still load-bearing.

GO to codebase iteration.

Do not return to Fable for another broad invariant-search round.

The next progress should come from:

repair → fail-sensitive re-execution → evidence-class updates → GPT adjudication.