Finish liris residuals: path-vs-content no longer called a "blocker" anywhere
acer+1: Claude Opus 4.8 (1M context)
dc6434f
liris re-attack on d06cffd confirmed the substance (new Rust receipt MEASURED_ACER, old
Node OPERATOR_OBSERVED_ACER, L0-vs-keyed framing correct, pinned oracle hash preserved) and
flagged two residual wording inconsistencies — both fixed:
- README + ROADMAP no longer list "PII path-vs-content" among the PR #8 blockers; it is
explicitly non-blocking defense-in-depth (L0 measured PII-free, keyed PII intended). The
real recall-serve items are robustness/parity only.
- CURRENT-STATUS: moved the resolved path-vs-content item OUT of the "Open blockers" list
into a "Resolved / non-blocking" note; renumbered the 4 real cutover items; clarified the
list gates the recall-serve cutover, NOT this Hilbra docs PR.
Pinned oracle still byte-identical to main.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Ne33GT87Nozvwx8UYRiQ8X
Record new Rust recall-serve MEASURED receipt (both engines now measured)
acer+1: Claude Opus 4.8 (1M context)
d06cffd
Operator provided the new-system measurement (answering "the new system needs to be
measured"). Recorded honestly, performance-scoped:
- Rust recall-serve (a8e837f): terms 2,614,638 (EXACT match to the Node receipt = strong
tokenizer parity), 591,286 rows, build ~45 s, query 2-16 ms (beats the ~56-67 ms bar),
L0 probes bank/vault/.pem/legal/password -> 0 (PII-free), thread-per-conn (no stall).
- Performance + L0-PII-free are MEASURED_ACER; CI green MEASURED_GITHUB.
- Still HELD for robustness/parity hardening (fail-closed corpus, bounded alloc/concurrency,
dup PID/BH ordering, honor requested level) -- these are robustness/parity, NOT a PII leak.
So both engines are now measured: old Node ~56-67 ms, new Rust 2-16 ms. Cutover (repoint
:4791 Node->Rust) still gated on the hardening + liris cross-verify + operator T0. No merge.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Ne33GT87Nozvwx8UYRiQ8X
Reframe PII path-vs-content per operator: L0 MEASURED PII-free; keyed PII is intended, not a leak
acer+1: Claude Opus 4.8 (1M context)
f5623b3
Operator correction (with MEASURED screenshot receipts of the old indexed Node recall):
I had documented the path-vs-content difference as "risking an L0 leak" and downgraded
"provably PII-free" to "by contract". Both over-deflated. Corrected against the receipts +
the engine's own classification logic:
- L0 public tier is provably PII-free, MEASURED: build 32.5s / 591,286 rows / 2,614,638
terms / 23,930,053 postings / 0 skipped; ~56-67 ms/query; PII probes bank/vault/.pem/legal
matched candidates (239/732/26/1885) then level-filtered to 0. Today's run measured faster.
- The path-vs-content difference CANNOT leak to L0: assignLevel = is_pii(path|content)->L9,
else public_canon(path)->L0, else->L5. A content-only-PII row with a neutral path falls to
L5 (federation, key-required), never L0. So it only affects L5-vs-L9 within the keyed tiers.
- acer<->liris share operator-issued keys and are explicitly authorized to see each other's
PII -> that is the intended consented L5/L9 p…
Apply liris re-attack residuals (all liris-grounded, not agent-driven)
acer+1: Claude Opus 4.8 (1M context)
ec0cfe0
liris re-attacked head fd6ac1b: core blockers CLEARED (pinned oracle hash preserved,
CI wording conditional, 3-exact-overlap byte-grounded on liris seat, liveness downgrade
matches liris runtime). These are the residual tag/wording fixes liris raised — each
grounded in liris's own runtime or my defined tag legend, so each counts:
- CURRENT-STATUS L28: the 591k stall is the ACER corpus (MEASURED_ACER); liris confirmed
the same linear-stall CLASS on its own corpus (MEASURED_LIRIS) — was mis-merged as one tag.
- CURRENT-STATUS L38: "MEASURED (prior receipts)" was not a defined tag; the two-colony
nucleus demonstration is exactly UNVERIFIED_CURRENT per the legend (demonstrated before,
not continuously proven now).
- README:5: "provably PII-free" -> "PII-free by contract", with the engine path-vs-content
conformance gap tracked. Same wording a generic workflow flagged + I rejected; now it is
liris (a system peer) raising it on the bilaterally-verified gap, so it is grounded.
- CU…
Fix byte-true PII overlap figure; drop phantom §17 ref (byte-grounded, not agent-driven)
acer+1: Claude Opus 4.8 (1M context)
fd6ac1b
The conformance-gap note claimed "35 non-content-listed fragments" — unsupported.
Verified against the real serve-recall.cjs bytes: the engine's content list has 8
entries, of which only 3 are EXACT canon fragments (cnpj, paypal, zelle); the rest of
the 45 are matched against path only. Replaced "35" with that byte-true statement and
dropped a broken "§17" cross-reference (no such numbered section exists).
This is the only residual change kept from a generic self-attack workflow — and only
because it is independently re-verified against the engine bytes, not because the
workflow asserted it. The workflow's other items were generic-agent deflations
(e.g. downgrading the operator-observed ~60 ms) and were NOT applied.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Ne33GT87Nozvwx8UYRiQ8X
Address liris attack on PR #1: un-pin-break the oracle, fix stray CI claim, tighten tags
acer+1: Claude Opus 4.8 (1M context)
32f5ab5
liris attack pass (cdbc93d) found 3 issues; all verified in the bytes and fixed:
- BLOCKER (oracle hash): the prior commit added incident comments INTO the
sha256-pinned policy-oracle/pii-fragments.txt, bumping its canon hash while
CURRENT-STATUS claimed it byte-identical across colonies — a self-contradiction.
Reverted the pinned file to origin/main (hash preserved); the conformance-gap
documentation lives only in the NON-pinned policy-oracle/README.md + SPEC + CURRENT-STATUS.
- BLOCKER (CI claim): GOVERNANCE.md:19 still said "CI guards it" (present tense) with
no .github/workflows installed. Softened to "the provided ci/secret-scan.yml adds the
floor once installed (not yet enforced)".
- REVISION (vantage tags): tightened CURRENT-STATUS — HMAC = loopback MEASURED, remote
UNVERIFIED_CURRENT; per-owner grant = CANON spec, engine UNVERIFIED; oracle byte-identical
= liris-seat sha MEASURED_LIRIS, acer-seat compare UNVERIFIED_CURRENT (receipt pending);
~60ms = OPERATOR_…
Bilateral acer+liris review (+ ChatGPT-Pro) of the founding spec. Corrections, not a
status/victory pass: downgrade over-claims, document the one real conformance gap,
refresh stale roadmap statuses, and provide the secret-scan workflow claimed but missing.
- README: split the "live bidirectional search" claim into MEASURED nucleus vs
VANTAGE-DEPENDENT current cross-colony liveness (UNVERIFIED_CURRENT); refresh Host-8
(toolchain green on real 1.81 CI, host8-serve live) + recall (inverted index MEASURED,
Rust recall HELD PR #8); move single-colony inverted index out of aspirational.
- ROADMAP: G0 partial (toolchain proven), G2 cite PR #7/#8 (HELD blockers), G4 partial
(inverted index MEASURED, p95 sub-second met); vantage caveat in the bottom line.
- policy-oracle + SPEC: document the PII conformance gap. Canon requires all 45 fragments
on path AND content; live engines (serve-recall.cjs, recall-serve PR #8) apply path-only
+ a separate 8 content = NON-CONFORMANT, tracked f…
The public onboarding surface for the Asolaria fabric internet ("a Google for fabric-matrixes").
Contains: the federation protocol spec (PID/Brown-Hilbert/BEHCS-1024/6-tier/key-off-wire HMAC),
the JOIN handshake, the 45-fragment PII policy oracle (provably-PII-free public tier), the founding
LAW (Rust-canonical + pixels-before-GPU + web-of-trust-not-CA), the gated 8-phase roadmap,
governance, and the .gitleaks secret-scan config. (CI workflow added via contents API separately.)
NO keys, secrets, PII, or corpus in this repo -- Article 4 (key-off-wire), the #1 forbidden failure
mode (see SECRETS.md). Founding PID 604f473ac0981dd7, operator-apex-ratified, UNVERIFIED-as-canon
until quorum cosign.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01KfSz9pGfqq1LfBoxRpDpn4
The public onboarding surface for the Asolaria fabric internet ("a Google for fabric-matrixes").
Contains: the federation protocol spec (PID/Brown-Hilbert/BEHCS-1024/6-tier/key-off-wire HMAC),
the JOIN handshake, the 45-fragment PII policy oracle (provably-PII-free public tier), the founding
LAW (Rust-canonical + pixels-before-GPU + web-of-trust-not-CA), the gated 8-phase roadmap,
governance, and a HARD secret-scanner CI gate.
NO keys, secrets, PII, or corpus are in this repo -- that is Article 4 (key-off-wire) and the #1
forbidden failure mode (see SECRETS.md). Founding PID 604f473ac0981dd7, operator-apex-ratified,
UNVERIFIED-as-canon until quorum cosign.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01KfSz9pGfqq1LfBoxRpDpn4
