Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
200 changes: 200 additions & 0 deletions research/frontier1094-certificate-binding-audit/README.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,200 @@
# Selected #1094 certificate: invoice/outer recheck and replay-binding gap

**Adversarial audit; no new κ; no defect in the current certificate.**

The committed #1094 certificate passes an independent recheck of its finite
invoice, bootstrap chain and all 47 + 47 outer slacks, and its profile agrees
across five artifacts.

The audit also finds a **check-level binding gap** in the selected record's
reproduce command, `make frontier1094-verify`. That command binds the committed
certificate to its cold replay **only through κ**. A well-formed certificate
that claims an uncertified complex endpoint, with every arithmetic identity
recomputed so κ is unchanged, is:

* accepted by the selected-record check once its hash pin is updated, and
* never compared with the replay output.

Only an exact moment evaluation, outside the reproduce path, rejects it.

The current certificate is not affected. The published replay output and the
recorded `assembly_sha256` are byte-identical to it. A one-line binding is
proposed below as a handoff; it is not applied to the selected script.

## Results

1. **Invoice and outer recheck of the committed #1094 certificate: PASS.** This
applies Wronskode's
[frontier1042-invoice-audit-seed81](../frontier1042-invoice-audit-seed81/README.md)
recheck, unchanged and loaded as its exact Git blob. The recheck covers:
* profile composition, per-vertex envelope and the nine finite guards;
* 7 + 7 bootstrap levels and 47 + 47 outer slacks;
* κ as the last 10⁻¹⁸ grid point below the absorption margin;
* the four outer controls and all 15 mutation controls.

The package binds the certificate directly, by path and SHA256
`1122cb3a…3045` (the `certificate_sha256` of the selected record at the
base), not through the moving `selected-result.json` pointer. As a positive
control, it reproduces seed81's published #1042 result exactly from the
frozen #1042 checkpoint.
2. **Seed81's own entry point is stale.** It requires the *selected* certificate
to be #1042's, so `make frontier1042-invoice-audit-verify` now stops with
`FAIL selected certificate path moved`. The #1042 recheck still holds through
this package's checkpoint-bound loader. The suggested repair is to read
`certificates/checkpoints/frontier1042.json` instead.
3. **The #1094 invoice arithmetic is unchanged from #1042.** The receipts differ
in exactly 14 fields:
* `PREFIX-AND-INVOICE.json`: only three provenance hashes;
* `OWNERS-AND-PROFILE.json`: those hashes, the idle-remainder profile
(ranks 1, 2, 3, 7, 16 and 48), the idle child count (7,550 → 7,526),
`actual_73column_sigmas` and `cross_replica_helper_roles`.

The envelope and guards are therefore identical to #1042's, which is
consistent with #1094's idle re-placements.
4. **Profile binding.** The certificate's `complex_profile` equals all four
other copies:
* the archived owner receipt's `paid_profile`;
* the package's `PROFILE.json`;
* the package's `data/PLAN.json`;
* the owner-audited profile in `INDEPENDENT1094.json`.

The maintainer composition (`check_frontier1094.py`) prices the owner-audited
profile but does not itself compare it with the certificate. The values agree.
5. **Mutant certificate.** The complex endpoint `b` is replaced by `b + 10⁻²⁴`,
the next grid point. The bootstrap, the 47 outer slacks and both complex
moment records are recomputed with seed81's formulas, and κ stays
`836330098538459/5·10¹⁷`. Results by layer:

| Layer | Mutant |
|---|---|
| Seed81 invoice/outer arithmetic (endpoint taken as data) | accepted |
| `verify_selected_result.check` with `certificate_sha256` re-pinned | accepted |
| `verify_frontier1094.py` cold replay and its candidate stage | never read the committed certificate |
| Exact moment, via `selected-suppliers-de-diagnostic`'s enclosure | **rejected**: residual > 1.62·10⁻²⁴ |

The cold replay contains no reference to the #1094 certificate path, the
selected record or `certificate_sha256`. Its only certificate-related string
is the historical `certificates/checkpoints/frontier1042.json` stage.

Its candidate stage (`complex-t16-rare-pattern-exchanges/verify.py`)
regenerates the assembly with the unchanged PR1087 strict verifier. It then
binds only the replay's status, its κ and its `complex_profile`, the last
to the package's `PROFILE.json`. It never names `complex_lower`,
`bootstrap`, `outer`, `moments` or `conservative_outer`.

At the certified endpoint, the residual is below −1.30·10⁻²⁴ (accepted). As
a control, the selected check does reject a κ edit (`Selected kappa differs
from finite certificate`). So the record's hash pin says *which bytes* are
selected, and the replay says *which κ* is reproduced. Nothing on the
reproduce path says those bytes are the replay's output.

## Proposed handoff (not applied)

Add one requirement in `scripts/verify_frontier1094.py` after the replay:

```python
require(sha(work/'candidate/strict/ASSEMBLY.json') ==
read(ROOT/'certificates/selected-result.json')['certificate_sha256'],
'Replay output is the selected certificate')
```

If cross-platform byte determinism is in doubt, compare parsed JSON instead.

This package shows the requirement holds on the published replay output: the
archived `candidate/strict/ASSEMBLY.json` and the recorded `assembly_sha256`
both equal the certificate. It also shows the requirement rejects the mutant.

A fresh cold replay with the added line was **not** run here. It needs the full
#1042 baseline replay, which is long, and editing the selected reproduce script
is a maintainer decision.

Statically, `scripts/verify_frontier1042.py`, `verify_frontier942.py`,
`verify_frontier890.py` and `verify_frontier799.py` also contain no reference
to their committed certificates. They were not audited further.

## Reproduce

Standard library only, plus `git` with history for the base commit. From the
repository root:

```sh
python3 scripts/research_package.py --package research/frontier1094-certificate-binding-audit --check-only
python3 -I -B research/frontier1094-certificate-binding-audit/verify.py
```

Expected (about 5 seconds):

```text
PASS_FRONTIER1094_CERTIFICATE_BINDING_AUDIT invoice_outer=PASS controls=15 receipt_delta=14 profile_copies=5 mutant_accepted_by=arithmetic,selected_check mutant_rejected_by=exact_moment no_new_kappa
```

[RESULT.json](RESULT.json) is the regenerated receipt; `--write` rewrites it.

Five pieces of audited code are loaded as exact Git blobs at base `0b1cccf8`,
so a later repair does not change what this package checked:

* seed81's `verify.py`;
* the endpoint package's `verify.py`;
* `scripts/verify_selected_result.py`;
* `scripts/verify_frontier1094.py`;
* the candidate stage `research/complex-t16-rare-pattern-exchanges/verify.py`.

The data inputs are pinned in [SOURCES.json](SOURCES.json). The mutant exists
only in a temporary directory during the run. The script refuses to run under
`-O`.

## Scope and limits

This audits the binding and arithmetic from the frozen receipts to κ. It takes
the receipts' semantic content as data: words, owners, hosting and local
replays. It re-establishes no moment endpoint beyond the two exact residual
evaluations above.

* **Not run:** the cold replay, the native or hosted pipeline, and the
all-size/compiler, precision/tape, analytic and H1–H5 obligations. These are
unchanged and inherited.
* **Environment:** run on Windows with CPython 3.14. The registered Python 3.12
replay was not run locally; `--check-only` passes.

## Open next questions

1. **Should the selected reproduce path bind the certificate to the replay
output?**
* *Why it matters:* today a promotion that commits a non-replay certificate
with the same κ would pass `make frontier1094-verify`. The certificate's
endpoints and outer slacks are what downstream audits read.
* *First step:* add the requirement above and run one cold
`make frontier1094-verify` on Linux, comparing raw and parsed equality.
* *Blocked by:* maintainer authorization to edit the selected reproduce
script, and the long cold baseline replay.

## Related work

* **#474** (artuskg, open) identified the same class of gap for the historical
gen4 record. Its `gen4-bank-verify` "does not compare that replay's
certificate with the frozen publication", and it adds a recursive comparison
helper, `scripts/check_selected_replay.py`. That helper is a natural way to
implement the handoff above for #1094. This package claims no priority for
the pattern; it establishes the gap on the current selected path, with a
concrete mutant.
* **#1277** (merged) covers the reuse-mode ancestry binding of the same
replay, and **#1654** (open) is the overlap receipt for the #1094
replay-tooling family. Neither adds a certificate-to-replay comparison.
#1654 also notes the same moving-pointer staleness in another capsule audit.

## Credit

Prepared for rtyu193-rgb with Claude Code (Anthropic Claude Opus 5.5) in one
session, with no subagents and no independent human review. Apache-2.0.

* **Wronskode:** the seed81 invoice/outer recheck, reused unchanged.
* **Navinor:** the `selected-suppliers-de-diagnostic` exact moment enclosure,
reused unchanged.
* **Douglas Colkitt:** the selected reproduce scripts and the
`check_frontier1094.py` composition.
* **Construction lineage:** Rohan Arun (#1094), with Jeremy Kothe's k8/T16
construction and the supplier lineage credited in the selected review. Josie
Elliston (#1091) and Sagar Thakrar (#1069) supplied the auditors applied by
that composition.

No predecessor file is modified.
189 changes: 189 additions & 0 deletions research/frontier1094-certificate-binding-audit/RESULT.json
Original file line number Diff line number Diff line change
@@ -0,0 +1,189 @@
{
"mutant": {
"candidate_stage_binds": [
"status",
"kappa_decimal",
"complex_profile == package PROFILE.json"
],
"candidate_stage_never_names": [
"bootstrap",
"complex_lower",
"conservative_outer",
"frontier1094-audit",
"moments",
"outer"
],
"change": "complex_lower + 1e-24 with bootstrap, 47 outer slacks, complex moment endpoints recomputed",
"complex_lower": "837731338404051761463/500000000000000000000000",
"frontier1094_verify_recipe": [
"@work=$$(mktemp -d \"$${TMPDIR:-/tmp}/integer-mult-frontier1094.XXXXXX\"); \\",
"python3 -B scripts/verify_frontier1094.py --work \"$$work/replay\""
],
"kappa_edit_control": "ValueError: Selected kappa differs from finite certificate",
"kappa_unchanged": "836330098538459/500000000000000000",
"layers": {
"exact_moment": "rejects",
"seed81_invoice_outer_arithmetic": "accepts",
"verify_frontier1094": "never reads the committed certificate; replay binds only kappa and profile",
"verify_selected_result_check": "accepts after repinning certificate_sha256"
},
"moment_residual_lower_at_mutant": "1.624724e-24",
"moment_residual_upper_at_certified": "-1.304215e-24",
"verify_frontier1094_certificate_strings": [
"Exact selected kappa",
"certificates/checkpoints/frontier1042.json",
"scripts/verify_selected_result.py"
]
},
"profile_binding": {
"artifacts": [
"INDEPENDENT1094 owner-audited profile",
"archived OWNERS-AND-PROFILE paid_profile",
"certificate complex_profile",
"package PROFILE.json",
"package data/PLAN.json profile"
],
"equal": true,
"profile": {
"1": 349604,
"2": 60192,
"3": 62585,
"4": 20448,
"5": 11808,
"7": 7,
"8": 1152,
"48": 4866
}
},
"proposed_binding": {
"check": "require(sha(work/'candidate/strict/ASSEMBLY.json') == selected['certificate_sha256'])",
"published_replay_output_matches": true,
"published_result_assembly_sha256_matches": true,
"rejects_mutant": true
},
"receipt_delta_vs_1042": {
"invoice": [
"/controller_audit_sha256",
"/owner_audit_sha256",
"/template_sha256"
],
"owners": [
"/actual_73column_sigmas",
"/controller_audit_sha256",
"/cross_replica_helper_roles",
"/explicit_idle_remainder_children",
"/paid_profile/1",
"/paid_profile/16",
"/paid_profile/2",
"/paid_profile/3",
"/paid_profile/48",
"/paid_profile/7",
"/template_sha256"
]
},
"recheck_1094": {
"conservative": {
"kappa": "104541262139063/62500000000000000"
},
"controls_rejected": [
"bit_endpoint",
"bootstrap_cut",
"call_primitives",
"complex_endpoint",
"drop_call",
"extra_idle",
"group_Sp70",
"kappa_up",
"max_rank",
"physical_plus",
"prior_dropped",
"router_bits",
"row_bound",
"slack_edit",
"swap_orientation"
],
"envelope": {
"compiler": 25934336,
"hosted_rebased_pervertex": 3588959010114264,
"local_adapters": 387056,
"local_primitive_bound": 10038040553630,
"local_primitive_bound_bits": 44,
"nonlocal_items": 10330706,
"selected_primitives": 28625376,
"unhosted_pervertex": 3588959010714576,
"used": 6466906000439144,
"wrapper_factors": 9953280
},
"guards": {
"group": "|Sp(72,2)| = |O(73,2)|",
"group_bits": 2628,
"halving": [
5329,
4608
],
"headroom_bits": {
"literal": 64716,
"local_primitive_bound": 4,
"logical": 3370,
"precision": 9988,
"router": 22012,
"streamed_mass_below_2E": 97352,
"tape": 109998
},
"prefix": {
"charged": 512,
"global_bits": 397,
"local_max": 68
},
"row": [
16469,
20161
],
"stock_bits": 2642
},
"native": {
"kappa": "836330098538459/500000000000000000",
"min_slack_names": [
"lambda_above_tau",
"lambda_above_internal",
"lambda_prime_above_lambda"
]
},
"profile": {
"calls": 144,
"deficit": 5120,
"halving": [
5329,
4608
],
"idle": {
"1": 2276,
"3": 377,
"7": 7,
"48": 4866
},
"idle_children": 7526,
"live": 14336,
"local": {
"1": 347328,
"2": 60192,
"3": 62208,
"4": 20448,
"5": 11808,
"8": 1152
},
"rank_mass": 1041408
},
"same_as_1042": {
"envelope": true,
"guards": true
},
"status": "PASS"
},
"seed81_binding": {
"audit_constant": "research/community-frontier1042-audit",
"selected_at_base": "research/community-frontier1094-audit/ASSEMBLY.json",
"stale": true
},
"status": "PASS_FRONTIER1094_CERTIFICATE_BINDING_AUDIT"
}
Loading
Loading