Conversation
Reconstruct consecutive rejections from the verified channel prefix, restore them through the v2 Gateway service, and safely reclaim stale private RPC sockets during process restart.
Allow remote evidence contexts to omit both a database and concrete writer, and qualify a turn plus bundle export through distinct writer and Gateway processes.
Carry writer public signer identities into a Controller with witness-only signing authority, then recover and continue through a fresh remote Gateway without a local database.
Bind a zero-spend, container-only qualification to the selected authority graphs and pinned local image while excluding service, lifecycle, resource, and scientific claims.
Add a clean-source, single-use Docker collector for the prospectively frozen Mode R boundary packet. Retain success and failure receipts, exhaustively probe ordered routes, and keep the result bounded to topology qualification.
Preserve the first development collector failure in a hash-bound supplement. Require target-specific responses, tolerate probe resets, count process liveness, and allocate a fresh development evidence identity before another attempt.
Preserve the second development collector failure after reproducing its status-zero timeout accounting defect. Batch target probes, require typed completion output, reject subprocess errors and signals, and allocate a fresh development identity.
Preserve the third development collector failure after its explicit completion guard rejected unresolved name probes. Probe inspected target network addresses, retain endpoint-level outcomes, and allocate a fresh development identity.
Bind the passing fourth development attempt to a portable receipt and live-evidence checker. Preserve its bounded topology scope and explicitly leave selected execution, application-runtime qualification, B12, and research findings open.
Bind the single-use selected responder-topology execution to a portable receipt and live checker. Retain exact route and endpoint accounting while leaving application services, lifecycle, faults, resources, independent audit, B12, and research findings open.
Add a process-development Baby host that retains the learner-facing boundary and delegates model execution to a distinct locked-down adapter process. Freeze the bounded qualification packet, prevent nested normalized deadlines, and verify four process identities, conformance, and teardown without claiming selected-topology or research completion.
Add a prospective single-use Linux collector for the frozen Baby/model development boundary. It binds exact source, records four child process identities, normalized outer timing, conformance and RSS observations, verifies teardown, and preserves failures without claiming selected topology or research completion.
Load the built isolation and learner package entry points directly so the clean-root qualification command resolves before execution. The prior invocation stopped before execute and left the single-use evidence identity unused.
Add a portable checker-backed record of the clean v0.1.259 development qualification. Bind the immutable raw receipt and exact execution sources while keeping selected topology, B12, behavioral findings, and publication claims explicitly open.
Add the Baby container service mode and Controller-side container factory, then exercise a Research-Grade turn through remote evidence, Gateway, Baby, and model boundaries with simulated anchoring. Keep this scoped as component integration rather than selected Docker qualification or B12 closure.
Bind role-specific Controller-to-Gateway learner relay sockets, Gateway-only Baby destinations, Gateway-local private-ledger callbacks, and fail-closed qualification requirements before selected application deployment. Preserve the prior authority graph and keep implementation, B12, and research claims open.
Bind each relay to one run, role, Baby, and downstream channel. Terminate private-ledger callbacks at the Gateway writer, validate returned event identity, and quarantine uncertain or out-of-turn activity without retry.
Add private one-admission Unix learner sockets and Gateway-owned Baby destinations. Exercise a signed Research-Grade turn and export with Controller reverse-ledger handling disabled.
Add Unix learner-host service mode and bind each outer Baby to one private model socket. Remove the model TCP path from the remote Gateway-relay runtime fixture.
Expose run-bound checkpoint creation and proof methods over a validated Unix RPC. Move witness access and checkpoint computation out of the Controller in the remote runtime fixture.
Expose a run-bound validated Anchor publisher RPC and move deterministic simulated-chain state plus the anchor-writer capability out of the Controller runtime fixture.
Move live Nursery audit interpretation behind a run-bound, single-method service capability. Add strict RPC validation and exercise the real remote interpreter in the eleven-process application fixture through delayed signed audit evidence. Regenerate the v0.1.268 conformance and research presentation surfaces while retaining B12 as open.
Allow selected remote contexts to omit Gateway and private-ledger writer ports when Gateway-hosted relays own those paths. Keep local construction strict and fail closed on unexpected Controller-side ledger callbacks. Exercise the complete eleven-process component flow without either capability and regenerate the v0.1.269 status surfaces.
Prospectively authorize the Evidence Writer and Checkpoint Service as the two hash-only callers of the single witness signer. Bind the unchanged v5 graph, exact mounts, refusal checks, signer-death behavior, and claim boundary without closing B12. Regenerate the v0.1.270 status and research presentation surfaces.
Implement the separate seventeen-service Mode R application Compose with scoped writer, Gateway relay, model, signer, Checkpoint, simulated Anchor, delayed Audit, Controller, and offline-verifier boundaries. Add a clean-source single-use development protocol and collector that reconciles live networks, mounts, resources, teardown, signed turns, delayed audit evidence, and offline verification without closing B12 or creating a research result. Regenerate the v0.1.271 status surfaces.
Preserve the v1 pre-start failure and use a fresh v2 development identity. Quote each tmpfs flow value, parameterize the signer run identity, and assert the resolved single-mount form without changing scientific design or thresholds. Regenerate the v0.1.272 status surfaces while retaining B12 as open.
Bind selected containers to the invoking host UID/GID, give model adapters explicit Unix-socket host entrypoints, and admit prospectively versioned application signer identities. Preserve the v2 startup failure, improve failure-time topology capture, assign a fresh v3 identity, and regenerate the v0.1.273 status surfaces without changing thresholds or scientific design.
Preserve the failed v3 identity negotiation, pre-bind generated prompt and scenario hashes for the separate Gateway, and retain complete container state on failed application attempts.
Preserve the v4 pre-execution stop, load the selected learner contract without leaking the array index into its version, and allocate evidence only after configuration resolution succeeds.
Preserve the completed-but-failed v5 measurement, poll one-shot service state directly, normalize Docker's networkless representation, and bind the exact post-run state vector prospectively.
The LV01 paired path hashed only input metadata before the receiver acted, binding no probability vector. It now persists a canonical payload with the actual frozen native, ordinary-record, and replay vectors plus candidate order, fit, cutoff, state, and schedule digests, alongside the unchanged v1 digest. A new audit function re-derives every vector from committed inputs and rejects forgeries. Ordinary fits are serializable and scored by one pure function, so the commitment and the auditor compute identical vectors.
Diagnosis + manifest remainder: sealed digest re-verification, 616-case prefix audit, scoring-method confirmation, ineligible dispositions. R01 identities reconciled vs rewritten e6d5da0, zero deltas (manifest r01ReconcileAddendum). History note: plans/ purged from this branch (old tip 15d0fb3 -> e6d5da0); .gitignore /plans/ keeps it out. Bump 0.1.469 -> 0.1.470.
Author
|
Status (HANDYM): R01 done (ALDM, e6d5da0) — 46-entry attempt ledger, v3 terminal FAILED at 1986/2016 (evidence RPC timeout, integrity ok, no receipt), collection stays blocked on iteration accounting. R02 done (HANDYM, cddc70d) — retrospective diagnosis + manifest, R01-reconciled with zero deltas. R03 done (HANDYM) — v2 protocol drafts, tests green. R04.1 in progress (ALDM) — frozen pre-action vectors. R11 skeleton + R12/R13 future-protocol drafts in. R06-R10/R15 still blocked on seal/qualification. Note: branch history was rewritten to purge plans/ (old tip 15d0fb3); sealed 8f22f8e provenance pointer dangles, bundle bytes intact. |
Parent LV01 run trains and seals in-process; a derived normal-branch child commits native, ordinary, and replay vectors in its prediction-commitment event before its turn record exists.
…(R04.2) Treatment labels previously had no runtime consumer: both ledger branches behaved exactly like normal. The collector now builds a sealed per-role batch of native target-maximizing tokens with a fixed seeded derangement, each branch run carries its committed slice, and the gateway substitutes the committed token at delivery. Ledger branches without a slice fail closed instead of silently passing the sender artifact through.
Methods section (8 subsections) grounded strictly in frozen LV01 v2 packets with transcribed digests; claim-to-data map rows added. No results or estimates (R09/R10 pending); TBD shells intact. Bump 0.1.472 -> 0.1.473.
Author
|
Update (HANDYM): E02 item 7 verified DONE with zero changes (v2->v3 correction line already in-tree: fixes b36b23e/96c3718/50702f9, v3 registration + 60/60 decision-rule evidence retained). R11 methods filled from frozen v2 packets (digests verified, r11-skeleton 8/8 green, no results added). Pushed with ALDM's R04.1/R04.2 commits now on the PR. |
The draw commitment hashed only role labels, so branches sampled independently with no reproducible draw. Each case now derives one seed from the v2 policy without a branch part; the runtime commits hiding digests pre-action, the receiver selects by inverse CDF over the shared unit value, and the post-action disclosure lets the auditor reproduce the selection. The paired runner reads the recorded commitment instead of re-hashing roles.
The v1 analysis had three members with whole-family degeneracy, a majority refit that dropped the training fold, and a power rule that planned for six components. The v2 family adds LV-U incremental prediction with per-member degeneracy, marginal plus simultaneous 0.05/7 intervals, and a majority refit over all folds. Fixed-cadence policy snapshots replace latest-only collection and branch-order permutation checks quarantine divergent sampling. The power selector and R reference move to seven components and four members with the registered incremental 0.02/0.04 values.
The paired-case verifier trusted collector booleans that the runner hardcodes true. The independent auditor reconstructs case identity, delivered token, prediction-before-action ordering, the shared draw, parent state, and the sample outcome from the registered plan, raw ledger streams, committed payloads, draw records, the sealed treatment batch, and the hash-linked transcript. Forged vectors, substituted tokens, reordered candidates, wrong draws, shuffled chronology, stale lineage, future ledger rows, and null-token violations fail closed, and the v2 case commitment binds the reconstructed record.
The command chain dropped arguments and every stage contract was a stub that recursively required its own outputs. The runner now forwards options, check-design verifies exact v1/v2 files and constants, and compile/bind/admit are real: the compiler binds allocation, source, model, design, and qualification identities into an immutable packet; bind re-verifies the packet into a pre-run commitment; admit re-verifies everything live into an immutable gate receipt. Gates no longer require their own outputs, and main stages stop without a locked design receipt.
The CLI tests drive check-design v1/v2, main-stage lock refusal, and the full compile-bind-admit chain on version-101 fixtures with cleanup: registration, immutability refusal, binding, early-collection refusal, and a ready gate receipt. On a dirty tracked tree the chain test asserts the clean-tree refusal instead of writing.
Derived runs inherit policies, not ledger streams, so branches mapped their own empty ledgers and prediction had no training associations. Per the Prototype batchEvaluation rule (reference parent evidence, never copy it per test action), paired branches now map the parent's frozen streams with parent-key authentication, heads are parent heads, and the auditor takes the parent streams once as shared input. The evidence test drops manual seeding and pins parent heads.
The auditor had no path from retained bundles to its inputs. The pure loader maps branch and parent bundle documents onto auditor inputs: causal order verified from sequences and recordedAt (draw, selection, disclosure, turn record), triple cross-check of the recorded selection, frozen model resolved through the registered derivation ref, and authentication from manifest signers. The stage engine checks journal integrity, case bijection, invalidity reasons, and ordered reserve consumption into an immutable audit receipt, and admission now requires a bound registration on every stage.
Correct stale reconciliation/live-detector wording, withdraw the unimplemented blanket, keep stale-ledger claims LV02-only, enumerate all five ordinary comparators, and distinguish valid nulls from empty cohorts.
…ents Propose frozen tracked bindings for simulation, cost, role-table, and seed authorities; reconcile lifetime charges and backup provenance without rewriting sealed records or zeroing unknowns.
Record deferred-study draft conformance, A0 host-blocked accounting, the consolidated R04.1/R04.2/R05.1 verification findings, and the interim R05.2b-ii collector A-gate review.
Add the user-manager slice definition, a dependency-free 60s admission preflight, and the single-lease registry with pinned transient-leaf properties. Draft only: not installed, not qualified; A0 stays HOST_BLOCKED.
Carries forward the inherited uncommitted R05.2b-ii worktree (collector, executor, collect/audit wiring, paired-case identities, treatment-target check, slot-case journal link) exactly as found. No acceptance implied; known A3/A4/A5/A7 gaps stand per prior review. A1/A9 hardening in the same tree (unverified until an admitted suite run): admit gains CLI/packet/binding cross-checks, live design-hash re-verification, a source-bound topology contract, an R01-B clearance gate, exact source closure including untracked runtime, an A0 lease check, and blocked-verdict sidecars that never burn a version. Bind/compile refuse untracked runtime source and closure-changing descendants; the power checker is design-version aware so no v1 receipt qualifies a v2 chain. CLI and unit tests rewritten to pin the fail-closed reasons.
Extend check-lv01-design to the prototype execution profile and direction amendment with exact top-level key sets for all five files, so any added, removed, or renamed field fails. Pin profile substance (deployment, separation, retention cadence, commitment payload, draw rule, branch conditions, batching, audit, release schedule), amendment lineage (preserved v1 bytes, portfolio bytes and card count, direction statements), and canonical seed stages and purposes. Two known draft defects are pinned deliberately with comments (blanket target prohibition, main/repeat seed labels); their planned corrections must update checker, protocol bytes, and manuscript digests together. Unverified until an admitted suite run.
… (A7/A4) A7: slot resources gain an explicit unresolved-counter set (zero-marked, never silent data); terminal capture persists resources into the journal slot, rejects unresolved counters on valid slots and fabricated values on unresolved ones; calibration rejects partial measurements; the collector samples RSS maxima instead of reporting final RSS as peak and captures partial wall/CPU/best-effort evidence on failure. A4: treatment batches take per-role native indexes so each case scores on its own receiver ledger (cross-role contamination closed); the derangement base seed uses the registered NUL-joined derivation with the intervention-shuffle purpose; the tautological served-target call is removed pending loader support for served targets. Deferred with notes: internal role-concat KDF (exact pins need an admitted re-verification run), draw-scope stage enum (draw reproducibility), max-sequence checkpoint binding (needs manifest design). Unverified until an admitted suite run.
Build frozen-design evaluation schedules through the dedicated generateLv01Case generator with per-partition per-role counts, round-robin target balance, eligible-only candidates, and cross-partition identity uniqueness, all committed under one schedule commitment. Includes scheduled-case lookup and executed-coverage verification returning the bound commitment. Property tests cover full and partial counts, determinism, seed separation, bounds rejection, and coverage pass/fail. Executor migration from the prohibited generic-evaluation peek to scheduled lookup (plus nursery scheduled-case serving) is the follow-up commit; no collection behavior changes here. Unverified until an admitted suite run.
a0-lease create/release crashed on their own subcommand positional because parseArgs had no positional allowance; slice argv past the subcommand instead. Refusal paths probed live (show/create/release with no live state return structured errors, no side effects). lv01-schedule read a nonexistent ground-truth candidateTypeCodes field; use the receiver-order candidate shuffle per the established collector convention. Adds a focused lease test (exclusion, owner release, stale/non-admitting receipts); suites still unrun.
Author
|
Sole-lead update (HANDYM): pushed 9 commits (0.1.481-0.1.489), branch now at fe5ac90.
Caveat: no suite has run since the runtime changes (A0 unadmitted until the slice install); all runtime commits are marked unverified pending an admitted run. ALDM has returned and reviewed the diffs statically (2 blocking findings, both fixed in fe5ac90). Next: A3-ii executor/nursery migration to scheduled lookup. |
…s (A3-ii) Replace the prohibited generic-evaluation peek with scheduled lookup end to end: collector builds the slot schedule on slot seeds and looks up the rotated within-support case, plans carry slotSeeds plus scheduledCase, and the nursery serves the scheduled case to the scheduled receiver at turn 0 (roles from the schedule, turn-0-only guard, fail-closed without generateLv01Case). Draw seeds derive from the slot scenario seed; the auditor re-derives from the planned slot seed and pins run-seed equality. schedule.json binds the commitment and executed case per slot; rebuild reads seeds and scheduled case from branch run-configs. Includes the scheduled-serving test (baby-a at turn 0 matching direct generation, no post-0 scheduled turns, missing-generator refusal), the collector/all-plan-input test rewires, a literal-union type fix in the schedule balance check, and per-role batch indexes in the loader/auditor fixtures (nativeIndex was a runtime break at HEAD). Pre-existing test-only type noise (literal widening, readonly negatives) and the three stale plan-construction scripts (missing actionDrawScope at HEAD) are untouched; vitest remains the gate. Unverified until an admitted run.
…stant (F1/F2) 0.1.491. auditLv01Stage rebuilds the per-slot schedule from packet-bound inputs (allocation re-read + sha re-verify, slot seeds, frozen engine config) and asserts rebuilt commitment, executed case, and branch state hashes against it; schedule.json treated as untrusted. Untouched type codes deduplicated behind exported LV01_UNTOUCHED_TYPE_CODES. Checks: tsc --build clean; eslint clean on touched files (direct binaries; frozen pnpm toolchain blocked, recorded separately). Full suite and F1 negative-path tests pending admitted run.
0.1.492. create/release with unexpected positionals now fail via the standard JSON fail path (exit 2) instead of a raw ERR_PARSE_ARGS_UNEXPECTED_POSITIONAL throw; adds a mirroring test case. Checks: node --check clean; refusal paths probed live (stray create / release / show with no lease all exit 2 JSON, no sidecars). Lease suite pending admitted run.
Author
|
The scheduled-case collector and its audit follow-ups are now on this branch: 41a61cc (A3-ii, 0.1.490), 74ac16e (F1 audit-time schedule reconstruction + F2 constant, 0.1.491), 404bf62 (lease stray-positional JSON refusal + test, 0.1.492). Passed: tsc --build clean; eslint clean on touched files (direct binaries); node --check on lease script; live lease refusal probes (stray create/release/show all exit-2 JSON, no sidecars). Still blocked/unrun: full suite and F1 negative-path tests (A0 admission: agents uncontained, preflight/lease fail-closed gaps under repair); frozen pnpm path (engines wants pnpm 12.3.4, host has 10.33.0/11.9.0); repo-wide eslint carries 2 pre-existing errors in untouched isolation.test.ts. Host containment and whole-dyad qualification are still open, so this is implementation progress, not a research result. |
0.1.493. Preflight compares every effective slice limit with policy rev 2, verifies member nice and role-aware lease residents, samples all readings at every 5s sample of a filled 60s window (short windows diagnostic-only), checks bavail storage, device/IO binding, --health-file evidence, and OOM recovery (global vmstat counter, 10-min history, incident + explicit reconcile, monitor-block). Lease acquires atomically (hard-link, exactly one concurrent owner) and binds receipt policy/source/task/command/peak/ path/io-class plus a content hash. New a0-launch.mjs supervises from outside the aggregate (kernel forbids non-leaf descendant migration): receipt re-verify, launch readback/membership/nice verify, enforced deadline, policy stop conditions, owned cleanup, no auto-restart. Checks: tsc clean; node --check clean; live disposable qualification (concurrency 10->1, trivial exit 0, emergency-stop, refusals, contained 16M OOM + reconcile, full 60s 12/12 valid). Suites unrun (admission).
Author
|
A0 safeguards are defined for this source (fde1020, 0.1.493): fail-closed preflight (effective-limit compare, member/resident checks, full-window sampling, health + OOM-recovery gates), atomic strictly-bound lease, and a supervised launcher with enforced deadline, stop conditions, and owned cleanup. Passed: tsc clean; node --check clean; live disposable qualification (10-way lease race yields one owner, trivial run exit 0, emergency-stop with verified readbacks, owner/tamper refusals, contained 16M OOM with incident + reconcile, full 60s preflight 12/12 valid). Still blocked/unrun: A0 vitest suites and the F1 negative tests (admission: 1.3G headroom, health evidence unbound, lead containment unproven); frozen pnpm path (engines wants pnpm 12.3.4, host has 10.33.0/11.9.0). Specific work remains paused until headroom and the research gates pass. Host containment and whole-dyad qualification are still open, so this is implementation progress, not a research result. |
This branch has not been deployed
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Successor to closed #1 on rewritten history (plans/ purged, see comments). R01/R02/R03/R04.1 batch.