everything we're testing, and why.
one page for every eval, experiment, goal and idea. what we measured, what it does not prove, and where our attention is going because a result, a paper or an idea pulled it there. negative and incomplete results stay on the page.
where our attention is going, and why
as of 2026-10-23. this is a ranked working agenda, not a claim that the loops are running. a merged library, a registered client and a verified behavior are different states. the public memory contract keeps those distinctions explicit. continuity plan
the metric and the rule
our north-star metric is context re-explanations per day. count the occasions when the user restates context an agent should have recalled. do not substitute search hits for this number. use, ignore and restatement need explicit witnesses. the daily count is not yet reported in the sources reviewed here. outcome contract
our rule for each system is a closed self-improvement loop: receipt → replay → gated promotion → rollback. capture what happened. replay real failures against a frozen baseline. require held-out evidence on the exact candidate revision and separate owner acceptance. keep the previous policy available when validity breaks. these are required boundaries, not evidence that every lane already has them. compilation and deoptimization · protected-search argument
1. useful continuity before more memory machinery
why now. the memory foundation moved. freshness, capture and the read-only facade merged into preview on october 23. the final facade integration reports 291 passing tests. the last synthesis comment still describes drafts; that status is now historical. none of these merges proves that a fresh client used a real correction. merged facade · capture · freshness
next experiment. qualify the deployed read path with explicit grants. then witness one authorized decision and later correction in a fresh second client. compare memory enabled and disabled with the same model, prompt and budget. preserve source citations, the actual model input and an independent answer check. the frozen agent-policy-to-user-correction variant must stay distinct from the still-blocked pure user-to-user pair. current synthesis
what would change our mind. if ordinary scoped source retrieval gives the same corrected behavior with fewer waits and less maintenance, keep that simpler control. a scope leak, stale required source or missing input witness blocks promotion. semantic compaction is optional. research synthesis
2. compaction economics, including the retries
why now. the frozen internal replay qualified terra at 4/5 recall against sol at 5/5. terra used 8.35% more input-plus-output tokens but 29.05% fewer uncached input tokens. the luna runs exceeded the byte target on roughly 84% of reported responses. a cheaper-family label did not mean fewer calls or fewer tokens. these are internal aggregate measurements; the public link exposes the instrumentation, not the benchmark corpus or report. usage ledger
next experiment. freeze another authorized cohort and recall key. compare the existing retry policy with a bounded size-control candidate. count maintenance calls, cached and uncached input, output, failed attempts and recall. keep missing usage unknown. inspect the actual call context rather than reconstructing it from display history. this is proposed work, not a restarted held optchat trial. ledger contract · call-context inspector · trial hold
what would change our mind. a repeated recall loss, more stale corrections or worse full-system cost would favor the control. uncached-token savings alone cannot establish subscription-quota or dollar savings when those weights are unknown. instrumentation and private-replay boundary
3. capacity for the experiments, not another scheduler
why now. the steward dry-run turns 624 job logs into 16 findings, but most historical queue, runtime and peak-memory fields remain missing. the earlier 557-job research snapshot has 13 unclassified observations. a later build stall was in controller storage before worker admission. these are evidence and infrastructure problems before they are learned-placement problems. steward · earlier coverage gap · stall receipt
next experiment. add source-bound gate-stage receipts to the existing execution path. distinguish pre-start holds from started or uncertain attempts. qualify one hermetic cpu evaluation for a reversible kueue canary only after physical health and task/image provenance pass. the existing placement smoke only prints its node; it cannot stand in for an evaluation. phones remain host-lane workers. canary audit and blockers
what would change our mind. if queueing, image work or control-plane i/o raises time to independently verified output, retain the host path. stop on duplicate execution, unsafe admission or unknown effects. do not scale optimization on top of unqualified capacity. hybrid migration decision
4. give autoresearch a real reward signal
why now. the bend story makes a useful boundary executable: a law pass is not owner acceptance. it also states the current limit plainly. the factory autoresearch replay mutates context policy against a deterministic simulated cost model. it does not rerun agent turns, train weights or establish learning gains. laws-to-autoresearch artifact
next experiment. start with read-only query rewrite or reranking on genuine memory misses. preserve the champion, a held-out cohort and failed attempts. use the existing outcome receipts rather than creating a new reward database. only then compare workflow search and supervised training on the same episode and evaluation budgets. miss-replay contract · training proposal · search-before-training idea
what would change our mind. reject a challenger that wins only on simulated cost, teacher agreement or replayed duplicate credit. require better held-out outcomes after optimization and maintenance cost. on drift, fall back rather than quietly weakening the judge. credit contract · deoptimization contract
5. local routing needs discrimination, not just calibration
why now. the observer lead did not confirm on the fresh cohort: brier worsened by 0.0094 and log-loss by 0.052. several models chose the safe argmax on every confirmatory turn. that does not prove their probabilities have no ranking value. the later comment correctly asks for discrimination metrics. confirmatory result · threshold caveat
next experiment. retain full distributions. compare simple priors, rules and local candidates on grouped fresh work items, with discrimination, calibration, risk/coverage and fallback cost. record cold loads, unavailable backends and timeouts separately. measure actual resource fit before placement. observer protocol · local shadow evidence · placement boundary
what would change our mind. if a simple prior plus deterministic gates matches held-out value, use it. a fast warm call cannot compensate for a cold-load miss, unmeasured outcome quality or false confidence. no learned score grants execution permission. selective autonomy
6. keep speculative methods small and falsifiable
why now. two ideas pull in different directions. shared-context research separates filesystem coordination from model-state reuse. the byte-distillation triage suggests exact end-of-token targets, but its larger-model advantage is a scaling-law prediction and inference economics remain unresolved. neither is a local win yet. shared-context experiments · byte-distillation triage
next experiment. test path-scoped write ownership and ordinary shared-prefix caching before cross-model cache transfer. for byte distillation, verify probability conversion and real storage/memory cost before training. retain the same legal actions, frozen outcomes and simpler controls. smallest-first context plan · conversion-first byte plan
what would change our mind. stop if leases only move contention, caching changes outputs, exact targets require unavailable logits or byte serving costs erase training savings. defer expensive reproduction until the bounded system comparison justifies it. context hypotheses · byte feasibility gates
public artifacts
- bend laws to autoresearch: executable models and a proposed protected search loop. not proof of universal interception or weight learning.
- k3s fleet handoff: recorded incidents and questions for infrastructure review. its promotion widget is illustrative. reference agreement is not independent task success.
every entry
november 2026
- frozen, not yet postedgenerative ui2026-11-02
a chart is not a web page
Does answering an inline chart as an openui-v2 program instead of an HTML page save output tokens and time, without losing validity?
Source-reported, same model and harness: 10/10 valid in both formats; openui used 3,748 output tokens versus 26,951 for HTML (7.2× fewer) and finished in 20.7 s versus 52.7 s on average (2.5× sooner), for about 922 extra input tokens. A 10-case openui sweep passed 10/10 on two models after a weekday-label oracle fix (9/10 before), and 9/9 inside a real T3 Code thread.
in focus · The cheapest measured lever on agent UI answers so far: shorter output, same validity, and a client that already renders it.
what this does not show
One model, one harness, two repetitions. The HTML arm is an unbounded interactive page checked only for the expected numbers; openui is a bounded four-chart catalog checked by a strict oracle. T3 Code does not yet instruct agents to use openui.
october 2026
- draft study PRmemory2026-10-23
Memory cannot grant permission
Can a pure memory projection preserve useful evidence without changing externally supplied authority?
The README defines nine quantified laws and 28 finite configurations over admitted, retrieved, replayed and absent memory. A later public operator receipt reports 33 tests, kernel PASS, 28 finite vectors/84 projected bits exact and rejection of the remembered-permission escalation mutant.
why it matters · Useful evidence and positive explicit grants survive, so the law is stronger than an all-deny model; remembered approval still cannot authorize a tool.
what this does not show
These are source-reported executions, not reproduced here. README execution-pending prose is historical relative to the later receipt; emitted-book proof does not attest source/compiler equivalence, trusted grant inputs, native authenticity or universal retrieval enforcement. Scope/correction laws remain future work.
- draft study PRlaws & governance2026-10-23
Bend proof-gate research program
Which deterministic policy boundaries should become laws rather than prose?
The context draft separates pure Bend laws, proof witnesses, finite parity, host-adapter tests and observed outcomes. It inventories implemented work-item L1–L7, supervision W1–W6 and integration I1–I5 families without claiming that every proposed law is proved or deployed.
why it matters · The design forces proof scope, toolchain trust and point-of-action enforcement to be audited separately instead of treating a green theorem as a permission token.
what this does not show
This is a documentation and design program, not an additional experiment. Recovered personal history is not republished; a compiler-emitted-book PASS and a CI workflow are not proof of source equivalence or protected owner-controlled branch policy.
- draft study PRfleet2026-10-23
Fleet admission fails closed
Can the selected-worker decision refuse unavailable targets and failed transfers without falling back to the desktop?
The pure decision has nine law signatures and an exhaustive 96-state Bend/Python table. A later public operator receipt reports 15 real-adapter/shell tests, 96/96 parity, kernel PASS and rejection of fallback, capacity-bypass and failed-transfer-bypass mutants.
why it matters · Missing capacity or failed source transfer produces a typed refusal rather than silently routing heavy work back to the desktop.
what this does not show
The checked-in guide retains earlier proof-execution-pending headings; later operator/CI receipts are separate source-reported evidence. Pure laws and mocked adapters cannot attest the real SSH peer, transferred source fidelity, runtime success or owner permission; no activation or merge is claimed.
- draft study PRphone & computer use2026-10-23
Bend agrees but the GPU does not win
Does CUDA Bend improve the existing per-action guard after the best CPU algorithm is considered?
All 4,096 finite guard decisions agree: 236 allow and 3,860 deny; full-vector GPU wall time is 103.8 ms versus about 1.09 ms for the equivalent Python table. Robust filtering resolves 340/340 unique groups versus 104/340 for the simple baseline, but Python makes exactly the same decisions.
why it matters · The periodicity ablation removes the apparent need for the GPU; the decision is to keep CUDA out of the per-action hot path.
what this does not show
Finite conformance is not universal proof or real GUI accuracy. Repeated batches collapse to a 4,096-case period, so large GPU throughput ratios are not verification-speed wins; public #157 does not archive a complete ABAB sequence or the separate HobbyLM comparison.
- draft study PRphone & computer use2026-10-23
A slower decision model finishes more tasks
Does HobbyLM’s semantic action loop complete more supported tasks than the small generic model?
The separate public matrix reports HobbyLM 22/31 (71.0%) supported successes versus Qwen2.5-0.5B 9/31 (29.0%). HobbyLM uses 58 decisions and 275.67 s controller wall versus Qwen’s 186 decisions and 340.20 s, despite model p50 / p95 of 1.194 / 3.464 s versus 0.247 / 0.705 s.
why it matters · Lower per-decision latency did not produce lower controller wall time or more supported success; action count and outcome matter.
what this does not show
This is #158, not Bend #157 or a demonstrated ABAB causal design. Models receive semantic text, not PNGs; adapted tasks do not establish general GUI/OSWorld competence. The strong subscription control failed before generation and is unavailable, not a zero model score; full unit/hosted qualification remains pending.
- publishedmemory2026-10-23
memory has a source-admission gate
can different harnesses recover the same source-supported state without replaying the full history?
The real-source readiness record has 4 bound source slots, 0 instantiated prompts, 0 model calls and 0 valid receipt rows; the six-case cohort is NOT RUN. The frozen correction plan also leaves all six stages in both arms and eight negative cases NOT_RUN.
why it matters · A smaller context and passing capture tests still do not establish source-supported memory; a real producer event broke a fixture-green replay reader.
what this does not show
The baseline post is published, but the executable source cohort is blocked and PR #115 remains draft and unscored; its successful replay repair is synthetic-source interoperability, not answer use.
read the postPR #57PR #59PR #115posts/unified-memory-v0.mdstudies/unified-memory-v0/manifest.yamlstudies/unified-memory-v0/README.mdstudies/unified-memory-v0/manifest.yamlstudies/unified-memory-v0/cohort.jsonstudies/unified-memory-v0/README.mdstudies/unified-memory-v0/READINESS.jsonstudies/unified-memory-v0/executable-cohort.v1.jsonstudies/unified-memory-v0/dsh/OPERATIONS.mdstudies/unified-memory-v0/manifest.yamlstudies/unified-memory-v0/README.mdstudies/unified-memory-v0/fleet-mbp-thermal.case.jsonstudies/unified-memory-v0/fleet-mbp-thermal.protocol.mdstudies/unified-memory-v0/fleet-mbp-thermal.receipts.jsonstudies/unified-memory-v0/memory_query_replay.pystudies/unified-memory-v0/test_memory_query_replay.py - in progressmemory2026-10-23
one correction, another client, no recap
Can a fresh client continue a workstream using the corrected decision without another explanation?
The October 23 synthesis reports 41 freshness tests, 81 facade tests and 159 capture tests at separate pinned heads; the frozen study variant has 8 fixture tests and 3 validated manifests.
in focus · The product test is fewer repeated explanations, not more indexed records or passing transport tests.
what this does not show
These counts are not additive independent experiments. The last issue comment predates the later merges of #198, #197 and #194 into preview. The real user-to-user decision/correction pair remains blocked; the frozen alternative is agent-codified policy followed by a user correction. Actual model-input delivery and independent use remain unqualified.
- publishedmemory2026-10-23
freshness binds to the bytes actually read
Can same-metadata edits, wrong scope and path races invalidate reuse before source disclosure?
PR #198 reports 41 passing contract/freshness/reuse tests at 8aac6cf208d6be183332fffec68d64220d497fee; the pre-feature baseline produced 3 collection errors. GitHub records it merged into preview on October 23.
in focus · A fresh hash attached to stale excerpt bytes would defeat provenance while appearing correct.
what this does not show
Published means merged source, not installed enforcement. This does not qualify live write-tool interception, authenticated caller identity, semantic sufficiency or cross-client continuity.
- publishedmemory2026-10-23
one read-only exact-source memory facade
Can multiple clients fetch the same authorized native message with current provenance?
The final #194 integration reports 291 tests passed in 33.20 seconds at d7a1a9770efda36b8fc0fc7894dfec0b20b6cb2c. Its integration RED was 1 failed and 290 passed before a mismatched source-digest fixture was corrected. GitHub records a preview merge on October 23.
in focus · This is a bounded common read path rather than another memory store.
what this does not show
AgentsView is the sole qualified exact-source backend. The checked-in ChatGPT grant is empty. Registration, approved corpus grants, host deployment and provenance-bearing queries from each client are separate receipts. The final implementation supersedes the research wave’s earlier cached-snippet characterization.
- publishedmemory2026-10-23
capture decisions without inventing user evidence
Can native decisions and corrections be admitted idempotently while rejecting injected or non-user rows?
PR #197 reports 159 passing ledger, identity, claim, scoped-capture and native-persistence tests in 49.04 seconds at 3cd637c6ce6cbe21372c166b48899ab2a0336439. GitHub records a preview merge on October 23.
in focus · A summary or assistant assertion must not become a new user-authored correction.
what this does not show
All capture fixtures are synthetic. No authorized live host/session/workstream capture hook is qualified. The accepted explicit native JSON grammar is not general prose extraction or an OMP capture adapter.
- in progressmemory2026-10-23
measure use, not just retrieval
Which retrieved memories were used, ignored, missed or restated by the user?
The public contract distinguishes hit_used, hit_ignored, miss and user_restated from denied, unavailable, stale, rate_limited, invalid and error transport states. The merged facade records retrieval dispositions only.
in focus · This supplies the real reward signal needed for memory optimization and the north-star metric.
what this does not show
No daily re-explanation count or closed real-miss replay loop is reported. Use requires an explicit client witness; user_restated requires native user-source evidence, not a generated label.
- proposalmemory2026-10-23
the OptMem wave produced a roadmap, not a benchmark win
Which OptMem ideas survive comparison with the existing source and memory contracts?
The synthesis collected 19 of 20 requested topic artifacts in an 80-minute window: 53,934 whitespace-delimited words and 501,289 UTF-8 bytes. The GraphRAG artifact was absent, not scored as a failure.
in focus · The useful distinction is retained versus discoverable versus current versus authorized.
what this does not show
These are overlapping research briefs, not independent experiments. The audit ran no compaction or handoff. Later memory PRs supersede its draft-status and cached-fetch observations.
- proposalmemory2026-10-23
scope-isolated summaries with exact expansion
Can compacted views preserve corrections and source lineage without becoming another truth store?
The synthesis distinguishes standalone OptMem’s 280-byte notes, UniiChat’s 512-byte node target and the local structural tree’s 8-event raw tail. The 512-byte target can be exceeded after up to 5 attempts.
in focus · This prevents a tree-shaped UI from being mistaken for complete, current memory.
what this does not show
The scoped forest, distributed outboxes, semantic lineage envelope and new adapters are proposals. Local structural nodes summarize counts/types/checksums, not semantic content. Exact expansion alone does not establish discovery.
- proposalmemory2026-10-23
corrections and erasure must reach derived context
Can as-of claims and deletion closure invalidate summaries, caches and compactor prefixes?
The research identifies compactor context prefixes as deletion dependencies in addition to summary ancestors, and separates valid time from recorded/known time.
in focus · Faithful historical recall and current applicability need different tests.
what this does not show
No erasure engine, historical-query runtime or cross-host deletion completion is measured. Rebuilding an index cannot recreate missing payload blobs or erase native archives and provider disclosures.
- frozen, not yet postedmemory2026-10-23
fresh-turn OptChat trial remains on hold
Does semantic fresh-turn assembly beat source-backed context after maintenance cost?
The recovery comment preserves an exact local implementation revision and an explicit skip-for-now decision. The OptMem synthesis records #128 and #142 as merged; #142 reported zero cache reads in 6 compactor completions.
in focus · A preserved experiment is not an active priority or authorization to restart it.
what this does not show
The October 4 issue baseline is historical, not proof that no implementation later landed. The held trial was not resumed or rerun. No universal continuity or production cache savings follows.
- draft study PRmemory2026-10-23
archive roots repaired; bounded latency still uneven
Does correcting archive environment propagation make original evidence reliably retrievable?
PR #45 reports AgentsView CLI FTS returning 4 results in 8.517 seconds and scoped HTTP FTS returning 4 in 0.691 seconds. ctx returned 4 once, then bounded attempts timed out at 45 and 60 seconds. Hybrid search returned 501 without configured embeddings.
why it matters · A configured archive is not a usable low-latency retrieval path.
what this does not show
Reported provider rechecks, not an independent reproduction here. The later audit is source-only and runtime NOT_RUN. The shell fragment does not automatically cover noninteractive shells, other shells, GUI clients or existing processes. Reliable latency and shared-memory continuity are not repaired merely by environment intent.
- draft study PRlaws & governance2026-10-23
binding drift stays advisory
Can desired memory bindings be inspected without turning configuration drift into CPU admission authority?
PR #44 reports 156 tests passed in 3.38 seconds at c5b29eed7c18a247174f704a0482a666283d6f56. A real dry-run reported binding DRIFT and action=none; the live configuration hash was unchanged.
why it matters · Configuration, runtime availability and verified use need independent states.
what this does not show
This is desired-state and advisory health evidence, not module installation, handshake or model-visible memory use. GraphRAG remains explicitly blocked. No live profile apply is claimed.
- draft study PRhermes & harnesses2026-10-23
native Unii conversation surface with terminal fallback
Can the existing conversation transport gain a native surface without changing session ownership?
The #2 component receipt reports 55 tests passed, 0 failed and 9,762 assertions at 205f6641fdbd385842d0571fa6593614c0f0974a, plus strict TypeScript and a terminal bundle.
in focus · The adapter preserves the original backend and fallback instead of rebuilding them.
what this does not show
Component evidence is preserved in integrated draft #4. The initial PR body says unverified, but its later receipt is more specific. Not an upstream merge or signed-in provider acceptance.
- draft study PRmemory2026-10-23
inspect actual call context, not reconstructed display history
Can the memory inspector show the context the model really received and recover original messages?
The #3 component receipt reports 18 focused tests passed, 0 failed and 118 assertions at 7653dd3f3e2012d18816f2aadd416d25be47dd6a. A display-history-versus-call-context defect was corrected with a server-owned snapshot.
in focus · Showing a plausible tree is weaker than showing the authoritative context of a particular turn.
what this does not show
This is Unii’s own memory explorer, not proof of cross-harness z0 continuity. The standalone suite once reproduced an existing performance failure; the integrated parent later passes the full suite.
- draft study PRhermes & harnesses2026-10-23
native memory explorer passes real PTY and isolated renderer checks
Do native input, memory navigation and fallback work together at the integrated revision?
Integrated #4 at b6d839005f3500b5258e530accc079f7f6d73492 reports 85 tests passed, 0 failed, 9,971 assertions across 19 files, and 4 real OS PTY input regressions. Both strict profiles and bundles passed. An isolated native renderer and terminal fallback were exercised.
in focus · Unicode, paste, duplicate sends and modal ownership are tested through actual input bytes.
what this does not show
The checks use developer-renderer/loopback and disposable state, not private user memory, ordinary signed-in GUI screenshots or provider inference. The PR remains open and draft.
- draft study PRhermes & harnesses2026-10-23
account for every model attempt before claiming savings
Can compactor and agent usage be measured without storing prompts or response bodies?
PR #5 reports 4 tests passed, 0 failed and 16 assertions. The opt-in ledger records provider usage counters, cached input, roles, retry indices, latency and failure categories; missing usage remains null.
in focus · A cheaper model label is not a measured lower cost once retries and cache behavior are included.
what this does not show
Instrumentation is disabled by default. The public PR does not publish the private replay corpus or its measurement tables. No live binary replacement or universal billing coverage is claimed.
- measured, internal evidencehermes & harnesses2026-10-23
Terra qualified for compaction, not a proven quota minimum
Can a cheaper-family compactor preserve recall while reducing uncached input?
The internal frozen replay reports the compactor switch from gpt-6.1-sol to gpt-5.6-terra: recall 4/5 versus 5/5; Terra used 8.35% more input-plus-output tokens but 29.05% fewer uncached input tokens. The two Luna variants exceeded the 512-byte target on 83.94% and 84.92% of reported responses, roughly 84%, driving repeated calls.
in focus · This reverses the assumption that fewer model dollars, fewer raw tokens and lower quota burn are interchangeable.
what this does not show
Numbers are approved internal aggregates, not public PR results. Subscription quota weights and dollar savings are unknown. Luna runs were censored, with recall unscored rather than zero. Loaded-host timings and the resumed control do not establish clean speed or cache-efficiency gains.
- draft study PRfleet2026-10-23
make fleet holds visible without a second build queue
Can CPU jobs respect host health and retry only when execution is known not to have started?
PR #41 records a 2% minimum-reclaim setting and the owner-selected 110 C source policy. A later stall receipt reports controller full I/O pressure avg300 of 72.18%, while the build worker had 9,835 MiB available at 67 C; a later bounded smoke exited 0 after controller I/O recovered.
in focus · A waiting build can be blocked before it even reaches the worker.
what this does not show
The thermal value is source policy, not a hardware-safety conclusion or direct user quotation. The later stall was controller storage, not worker admission. The branch is draft; reported live source use does not mean default-branch integration. No paired throughput gain is claimed.
- draft study PRfleet2026-10-23
624 job logs become 16 findings
Can the existing health path extract useful evidence without becoming another scheduler?
PR #46 reports 62 tests passed in 0.537 seconds and a dry-run over 624 jobs producing 16 findings. Its ledger includes 33 historical DiskPressure observations, 9 repeated missing-tool observations, 54 terminal exit-75 rows and 2 synthetic exit-137 probes.
in focus · The first payoff is actionable missing evidence and repeated failure classes, not an automated placement win.
what this does not show
Historical queue/runtime/peak-memory coverage is mostly missing and remains null. The synthetic exit-137 probes do not identify a production OOM. A native telemetry smoke failed to acquire the quiet lock within 120 seconds and was not admitted. No before/after performance win is measured.
- proposalfleet2026-10-23
learn from real attempts, not exit-code folklore
Can job receipts support scoped lessons and shadow placement hints?
The earlier research snapshot reports 557 jobs but its enumerated exit buckets total 544, leaving 13 unclassified. The later steward snapshot has 624 jobs and must not be treated as the same denominator. The proposed fleet projection currently has 4 host observations and zero offers, sessions or leases.
in focus · Identity, stage and outcome coverage must improve before any placement policy can learn honestly.
what this does not show
Owner-reported earlier counts were not reconstructed by the research authors. Exit 0 is not verified task success; exit 137 is not confirmed OOM without cgroup evidence. Physical host resources do not establish allocatable lane capacity. No learned policy benefit is measured.
- proposalfleet2026-10-23
small observability stack, only after an I/O budget
Which existing Kubernetes tools are worth adding to a small mixed fleet?
The OSS research ranks physical storage/eviction diagnosis and the existing health gate before new exporters. It proposes a narrow kube-state-metrics pilot and a single-node VictoriaMetrics collector/store; descheduler and provisioning integration are deferred.
in focus · Observability can itself worsen the constrained storage and control-plane path.
what this does not show
Resource budgets and retention settings in the research are proposed pilot values, not measured safe limits. No installation or runtime experiment was run. Earlier thermal and queue descriptions are superseded by the current admission branch.
- proposalfleet2026-10-23
one hermetic Kueue canary, not fleet-wide migration
Can queueable CPU verification benefit from Kueue while phones and interactive work keep native execution?
The migration research selects an opt-in hybrid. The later source audit is native NOT_RUN: the existing smallest placement smoke requests 50m CPU and 32 MiB, but only prints its node. It is not a source-bound evaluation. Current host retry semantics use exit 75 and 60-second pre-start waits.
in focus · The blocker is a qualified workload and receipt chain, not merely installing a scheduler.
what this does not show
The desired hermetic evaluation remains blocked by unqualified installed health, task/image provenance and missing machine-readable gate-stage receipts. Phones remain host-lane workers, not Kubernetes nodes. Queue admission and process completion are not independent eval success.
- proposalfleet2026-10-23
typed capacity offers, not movable sessions
Can provider quotas and physical resources share a read-only vocabulary without mixing authority?
Issues #131 and #133 define z0.capacity.snapshot.v1 with explicit missing sources and null unknown capacity. Semantic eligibility, placement economics, physical execution and presentation have separate owners; active sessions remain sticky.
in focus · The contract can unify inspection while avoiding a second quota ledger or scheduler.
what this does not show
Early comments describe concrete P0 wiring, but the later October 23 audit identifies #134 as closed and unmerged. A projected host is not a fresh placement offer or live actuator. This harvest does not certify deployment.
- proposalfleet2026-10-23
measure local model fit before selecting placement
Which allowed local models fit with real context and co-resident services?
Issue #50 targets a 12 GB GPU profile and asks for measured runtime, quantization, context/KV, warm/cold and co-residency evidence. The later steward handoff supplies host observations, not allocatable offers.
in focus · Local inference costs depend on residency and context pressure, not parameter count alone.
what this does not show
No comparative throughput or verified placement result is reported by the issue. z0 selects semantic candidates; Kerdoios ranks only eligible placements. Hardware capacity is not model quality.
- frozen, not yet postedsmall models & routing2026-10-23
preserve serialized serving and typed outage recovery
Can model swapping wait for an upstream response and keep transport failures distinct from parse errors?
Issue #196 identifies source that holds the supervisor lock through the response, serializes challengers, prewarms models and deduplicates outage records.
why it matters · Reliable negative evidence is necessary before local-model comparison is meaningful.
what this does not show
Source recovery only. Its delayed-transport acceptance test was not run; no live model load, provider call, new benchmark or deployment is claimed. Recovered source is not an active serving guarantee.
- idea / paperpapers & ideas2026-10-23
exact end-of-token byte distillation
Can tokenizer-independent byte targets produce an economical bounded local specialist?
Issue #203’s paper triage reports roughly 1.28B layer-parameter students, up to 1T text bytes and 64 H200 GPUs per training run. Up to about 4 percentage points of downstream advantage is a scaling-law prediction; about 6.35x unique-text and 5.19x logit-storage savings are paper-regime calculations, not local measurements.
in focus · It suggests a new target representation, but economics must be checked before expensive training.
what this does not show
The paper itself was not independently fetched in this gh-only harvest; claims are attributed to the public issue’s triage. End-to-end inference matching is unresolved. Full teacher logits are needed for exact conversion; a local tiny pilot would not replicate the large runs. No structured-output, latency or VRAM win is established.
- proposallaws & governance2026-10-23
mobile work starts with a real offline product gate
Can connector-only contributors verify bounded contract patches without implying host acceptance?
The refreshed #177 board prioritizes a secret-free preview CPU job, then receipt and serializer negative controls. The source notes that preview has coordination automation but lacks the proposed product unit workflow.
in focus · Small offline gates can make remote contribution useful without fabricating runtime coverage.
what this does not show
Cards are assignments and dependencies, not executed tests. Connector-only work does not qualify GPUs, phones, local services, private host data or default-branch promotion.
- publishedpapers & ideas2026-10-23
artifact: let the agent search; keep the laws fixed
Can executable laws and refusal receipts define a protected search problem?
The public artifact reports 10/10 law-breaking mutants rejected, parity on 400 seeded tapes and 38,454 state/event/availability parity cases for its compiled work-item subset. Its memory panel cites 28 finite vectors and 84 parity bits.
in focus · It makes the proposed exact-revision verification and separate owner-acceptance boundary concrete.
what this does not show
This is an executable explanatory artifact and linked proof/replay evidence, not universal runtime enforcement. Source translation, backend and adapter remain trusted. The factory autoresearch replay uses a deterministic simulated cost model; weight training, live promotion and learning gains are not established.
- publishedfleet2026-10-23
artifact: fleet handoff and the layer above Kubernetes
Which mixed-fleet failures should native infrastructure solve instead of custom orchestration?
The public handoff presents 624 historical job logs, 9 failure topics and 3 named ClusterQueues. It explicitly contrasts the earlier 557-job snapshot and the steward’s 54 terminal exit-75 rows with the handoff’s roughly 41.
in focus · The handoff turns specific failures into questions for infrastructure expertise without claiming a scheduler win.
what this does not show
The page was not a fresh fleet probe. Its 95% shadow-promotion widget is illustrative, not an accepted global gate; reference agreement is not independent success. Kubernetes-native remedies are research guesses, not deployed fixes. Reconcile per-snapshot denominators before comparing rates.
- proposalhermes & harnesses2026-10-22
fan-out needs a review surface
Can fan-out feed a preview card that makes the current acceptance task obvious?
The issue reports 3.4 s edit→screen and turbo motion 0.55× for its preview milestone, while the 5-fix cloud pilot remains proposed. The later ownership decision freezes PR #193 as an offline prototype, not a live deployment.
why it matters · A Now testing card ties a candidate to individually decidable criteria; the ownership correction prevents treating a cognition prototype as deployed orchestration.
what this does not show
The preview figures are issue-reported, not a frozen reproduced benchmark; cloud CLI fan-out, consumer-app automation and visible card acceptance are separate unproven lanes.
- draft study PRhermes & harnesses2026-10-22
app fan-out passed offline tests, not live qualification
Can an opt-in app fan-out prototype preserve stop, admission and rollback boundaries?
PR #193 reports 64 passed in 2.00s and synthetic Git success/revert smoke with real ≥20-second waits; no live app launch or actual preview change was performed.
why it matters · Real Git rollback evidence can coexist with no physical-device proof; synthetic qualification must not unlock live automation.
what this does not show
Stub build, fake preview logs and synthetic screen fixtures; later ownership guidance retains this as a prototype draft, and consumer-app account/compliance risk remains unresolved.
- draft study PRlaws & governance2026-10-21
Work-item laws and incident replay
Can a small proven statechart detect work that was merged, stranded or served without the required evidence?
Source-reported: 15 Bend claims under L1–L7 pass, 10/10 mutants are rejected, and the Python twin matches 400 seeded tapes over 5,077 events. The replay flags 39 integrations without a head receipt, 13 dropped interrupted workers and 32 stranded-integration candidates.
why it matters · The same transition model becomes an incident detector, but its evidence must retain candidates and unknowns rather than turn them into proven violations.
what this does not show
Emitted-book proof only; finite parity and heuristic historical mapping are not proofs of runtime behavior. The README says 174 items but the public artifact lists 111 PRs, 5 services and 59 workers; the total is inconsistent. Shared GitHub identity prevents attributing self-approval.
read the postkvnloo/z0evals PR #105studies/fsm-work-item-v0/README.mdstudies/fsm-work-item-v0/manifest.yamlstudies/fsm-work-item-v0/data/bend-receipt.jsonstudies/fsm-work-item-v0/data/bend-mutations.jsonstudies/fsm-work-item-v0/data/parity-seed7.jsonstudies/fsm-work-item-v0/data/replay-public-20261021.jsonstatechart program issue #182 - draft study PRlaws & governance2026-10-21
Worker supervision and lease replay
Can worker lifecycle laws expose lost work, lease collisions and unobserved resource waits?
Source-reported: 19 claims under W1–W6 pass, 17/17 mutants are rejected, and 400 worker tapes cover 3,370 events plus 100 lease and 100 device tapes. Replay finds 26 overdue workers and 419 writes in another session’s lease; the original 13 dropped workers are classified as 10 resume, 2 already integrated and 1 abandon.
why it matters · The replay turns a vague lost-worker problem into a bounded decision queue without pretending that a recommendation is an executed recovery.
what this does not show
Heartbeat, successor and lease mapping are heuristic; recommended actions were not performed. The README’s fleet “no queue wait” wording is not supported by end-to-end queue clocks: the public data explicitly exclude retry-queue time and record busy retries.
read the postkvnloo/z0evals PR #111studies/fsm-worker-supervision-v0/README.mdstudies/fsm-worker-supervision-v0/manifest.yamlstudies/fsm-worker-supervision-v0/data/bend-receipt.jsonstudies/fsm-worker-supervision-v0/data/bend-mutations.jsonstudies/fsm-worker-supervision-v0/data/parity-seed7.jsonstudies/fsm-worker-supervision-v0/data/supervise-public-20261021.json - draft study PRlaws & governance2026-10-21
Integration safety and debt inventory
Can an integration-owner statechart separate overdue work from missing deployment evidence?
Source-reported: 11/11 semantic mutants are rejected and Python/Bend parity covers 625 tapes, 16,225 prefix comparisons, 20 deadline cases and 300 lease cases. The inventory has 90 open drafts across 8 repositories and 79 findings including gaps: 21 definite overdue-uncontained drafts, 27 unverifiable drafts and 6 unattested runtime/build revisions.
why it matters · The checker refuses to manufacture a healthy deployment from a matching checkout and keeps missing integration ownership separate from proven policy violations.
what this does not show
A deadline monitor is not a liveness proof; 79 findings are not 79 proven failures. Checkout drift does not attest an installed build, browser probes after the initial snapshot are WHAT-IF, and the full upstream browser-project gate was blocked by missing Vitest.
read the postkvnloo/z0evals PR #113studies/fsm-integration-v0/README.mdstudies/fsm-integration-v0/manifest.yamlstudies/fsm-integration-v0/data/bend-receipt-public.jsonstudies/fsm-integration-v0/data/bend-mutations.jsonstudies/fsm-integration-v0/data/parity-seed3.jsonstudies/fsm-integration-v0/data/integration-public-20261021.jsonstudies/fsm-integration-v0/data/browser-public.json - draft study PRpapers & ideas2026-10-21
Human-capability contract fixtures
Can a synthetic outcome contract reject missingness and timing errors without pretending to measure human benefit?
The README reports 14/14 synthetic test methods; the CLI shows one observed control outcome and one missing candidate outcome. Accepted input always reports CONTRACT_ONLY, human_benefit=NOT_EVALUATED and can_promote=false.
why it matters · The missing candidate is neither filled with zero nor turned into a winner; conformance stays separate from benefit.
what this does not show
No manifest is intentionally frozen for this pre-study slice. Invented fixture revisions/timestamps are not authentic registration, a synthetic label is not a privacy guarantee, and no real human efficacy or health conclusion is evaluated.
- draft study PRlaws & governance2026-10-21
Compiled statecharts in the browser
Can the diagram replay the same compiled transitions as the proof-bound producer?
The current receipt-bound view has 17 tapes, 96 steps and 73 display edges, with zero missing or extra Mermaid edges. The README reports all five new snapshot-probe tapes completing at 412×915 and matching every compiled-replay prefix.
why it matters · The browser displays compiled-model transitions instead of maintaining a second hand-written state machine.
what this does not show
Earlier counts in the guide refer to the original two-loop snapshot, not current totals. New tapes begin at observed snapshots but all later events are WHAT-IF; full upstream check:fast was blocked by missing Vitest, and the diagram grants no runtime authority.
- draft study PRmemory2026-10-21
perfect recall, stale first
on one identical synthetic corpus, do native memory providers retrieve the answer and rank its update first?
All three measured surfaces had Recall@5 1.000 and MRR@5 0.944, but newest doc top-1 was 0/5 for each. p50 ms was 17.896 for ctx, 0.666 for Hermes and 1.267 for TencentDB; Supermemory and GraphRAG retrieval were not run.
why it matters · Perfect fixture recall coexists with stale-first retrieval; missing hosted credentials and incomparable prior extraction results stay missing rather than becoming a winner claim.
what this does not show
Draft manifest and open draft PR #114; eight synthetic documents and warm repetitions, unequal native integration overhead, incomplete build provenance, and no generated-answer or full five-way comparison.
- draft study PRsmall models & routing2026-10-21
a green suite had silently omitted the capacity tests
Does the provider-capacity export preserve quota dimensions and actually run its tests?
PR #57’s qualification found 0 discovered projection tests where 4 were expected. After converting the existing tests to the suite’s runner, 4 focused and 207 total tests passed; the earlier 203-test green suite had excluded all projection cases.
why it matters · A passing suite count can hide complete absence of the feature’s own contract checks.
what this does not show
No production implementation changed in this repair. No live provider discovery, timer freshness, consumer integration or adoption was verified. Overlapping capacity CLI paths still need reconciliation.
- proposalfleet2026-10-21
move expensive calls, not interactive sessions
Can bounded heavy tool calls use the fleet without making each turn slower?
The #185 note reports warm transfer observations of 0.62–0.71 seconds for a 33 MB repository on the build worker, about 9.7 seconds on a phone and 242 seconds when the build worker was cordoned. E2 proposes at most 1.5 seconds added latency per call.
in focus · An apparently cheap remote call becomes expensive under admission waits.
what this does not show
Transfer observations are not controlled end-to-end task latency. E2 remains NOT_RUN; its reduced off-fleet-command and lock-wait targets are hypotheses. Current interactive sessions stay sticky.
- idea / paperpapers & ideas2026-10-21
test shared-prefix reuse before cross-model cache transfer
Can existing prompt caching remove repeated shared-context prefill before a custom fuser is needed?
E3a proposes at least 4 calls over a shared context of at least 2,000 tokens, targeting at least 50% lower TTFT and 80% lower prompt_n. E3b proposes a 0.6B/4B fuser comparison on 28 local-cognition fixtures only after E3a is positive.
in focus · The smallest existing-cache control can falsify the need for a more complex mechanism.
what this does not show
All E3 arms are NOT_RUN. Targets are not savings. Cache-to-Cache shares model state between a fixed model pair; it does not provide file/DOM coherence or replace write leases.
- draft study PRhermes & harnesses2026-10-21
same model, different harness overhead
How much latency and task overhead comes from the harness rather than the model?
In s3, cold TTFT was 11.13 s for hermes-base, 3.18 s for hermes-speed, 2.22 s for fx-cli, 2.00 s for fx-base, 2.09 s for pi-base and 4.07 s for omp-base; all were 32/32 correct. All 168 counted s3/s3h task runs passed runtime verification.
why it matters · The speed profile removes much of the apparent Hermes gap; the engine-versus-CLI distinction changes the comparison.
what this does not show
One host and short synthetic tasks; custom libfx is not the fx CLI, 19 contaminated task rows were excluded, and deleted work copies prevent independent patch replay.
- proposalhermes & harnesses2026-10-21
the next-action predictor nobody committed
Does a pull-first action panel have a better starting point than the unused push predictor?
The frozen baseline recorded 7,859 shadow predictions, 19 ready-tier surfaces, 0 explicit commits and 0 accept/reject decisions in 30 days. On app-switch outcomes, recency top-1 was 0.771 against the engine’s 0.454; confidence ≥0.82 was correct 22.7% of the time (n=76).
why it matters · A plain recency list beats the predictor while its highest-confidence band is worse; changing the action space matters before adding a model.
what this does not show
One person and one desktop; the new pull-first design has no data, and recency is reconstructed from prediction-time states.
- draft study PRphone & computer use2026-10-21
a scoped agent API on a phone
Can Hermes read bounded app summaries and perform allow-listed actions while consent and audit gates hold?
The two end-to-end scripts passed 12/12, 12/12; warm data_freshness latency (n=20) was p50 10 ms and p95 14 ms. In the second run, open_screen reported landed: true in 487 ms; background opening posted a notification and left the foreground app unchanged.
why it matters · The background result is deliberately a notification, not a stolen foreground; the API distinguishes permission, delivery and landed action.
what this does not show
One phone and one grant-scope combination on a local integration lineage; advice quality was not evaluated, and installation plus grant UI verification used adb.
- proposalhermes & harnesses2026-10-21
recommendations with an interruption budget
Can bounded recommendations reduce needless prompts without pretending delivery proves benefit?
The issue proposes a no-model deterministic baseline and NO_SUGGESTION on absent or stale evidence; it reports 14 passing local test methods only for a separate synthetic evidence-shape gate, not useful-outcome validation.
why it matters · The baseline can abstain without spending inference or turning uncertain observations into health decisions.
what this does not show
Open proposal; no measured decision latency, override rate or independently verified benefit is reported, and the synthetic gate does not validate permission.
- draft study PRhermes & harnesses2026-10-20
Useful work versus faster activity
Does a faster collector retain the required facts when failures and preparation remain in the accounting?
In the bounded source-reported confirmation, control took 31.121s versus candidate 11.280s including measured preparation: 63.75percent lower observed work time. Both retained the prespecified useful facts; the candidate’s extra user-service timeout is retained, and incomplete first-use optimization cost prevents a complete ROI claim.
why it matters · The faster first candidate lost required facts and was rejected; the later bounded win is not allowed to erase failed work or partial first-use cost.
what this does not show
One related workflow family and one fresh work item per arm are not fleet-wide throughput or a causal estimate. The earlier package-loss candidate was rejected; stale-reference failures remain failures, independent custody and full assistant latency are absent.
kvnloo/z0evals PR #102studies/engineering-throughput-v0/manifest.yamlstudies/engineering-throughput-v0/data/iac-collection-v3.confirmation-facts.jsonstudies/engineering-throughput-v0/data/iac-collection-v3-confirmation.report.jsonstudies/engineering-throughput-v0/data/iac-collection-v3-first-use.report.jsonstudies/engineering-throughput-v0/data/native-facts.json - proposallaws & governance2026-10-20
act, observe, ask, abstain and escalate are different decisions
When does learned judgment add held-out value beyond deterministic gates and escalation?
RFC #55 requires per-family risk/coverage, false confidence, versioned calibration and fallback costs. Its October clarification separates deterministic validity checks, admissible agent judgment and explicitly delegated human tradeoffs.
in focus · The useful target is safe decision coverage with a costed fallback, not raw agreement.
what this does not show
No global confidence threshold or production autonomy win is established. Confidence and learned preferences do not grant authority; even an ACT recommendation can be denied downstream.
- proposallaws & governance2026-10-20
compile repeated cognition only while it remains valid
Can the cheapest verified mechanism retain outcome quality and deoptimize under drift?
RFC #56 replaces a fixed model ladder with mechanism-neutral promotion, explicit operating regions, invalidators, fallback and maintenance economics. Its October slice proposes extracting deterministic bookkeeping around existing contracts before expanding learned selection.
in focus · A specialist that cannot repay discovery and drift cost should be narrowed or removed.
what this does not show
The self-improving shadow stages have mixed historical implementation states; the RFC does not establish a closed deployed loop. A simpler rule, cache or retrieval control can win. The learner must not edit its sealed judge.
- in progresslaws & governance2026-10-20
one readiness queue, separate source and runtime acceptance
Can existing work be qualified without duplicating fixes or promoting stale evidence?
The dated queue snapshot lists 14 open PRs. Its wave reports 1 integration and 3 existing-PR repairs: file-search access tests 36 passed with 2 explicit skips, route-review tests 17/17 and scoreboard tests 21/21. A direct-source rerun passed 1,162 tests with 6 skips and 5 deselections.
why it matters · The queue exposes how much unfinished work is acceptance and integration rather than new implementation.
what this does not show
This is an October 20 snapshot, not today’s open-PR total. Counts are reported receipts, not reruns by this harvest. Repaired source, exact-head tests, merge and runtime adoption remain distinct; source-head CI and merge-checkout CI are not interchangeable.
- draft study PRphone & computer use2026-10-20
fast phone control still needs confirmation
Can in-process accessibility reduce phone-control latency without pretending a tap proves intent?
Whole phone_do p50 / p95 was 267 / 477 ms (n=16), with 16/16 landed; warm tree reads were 9.3 / 104 ms versus 98 / 165 ms for u2. On the synthetic chooser set, phone Qwen3-1.7B acted right / wrong 16 / 6 @0.95.
why it matters · Cheap lexical choice beat adding an on-phone language-model tier; waiting for confirmation dominates the fast path.
what this does not show
Not preregistered; rules were tuned on the fixture, the real-goal sample is small, and landed means an accessibility event or screen change, not independently verified intent.
- draft study PRhermes & harnesses2026-10-20
voice parity did not pass
Does z0live meet the preregistered OMP voice-parity gates on output audio?
Round 2 direct-turn p50 / p95 was 1,112 / 1,539 ms for OMP, 1,265 / 1,960 ms for z0live-bidi and 1,815 / 2,978 ms for z0live-realtime; both candidates were FAIL. The later bidi subset had 0/15 direct turns with no audible reply but 1 stale resume.
why it matters · A quicker result or fixed missing reply does not override blocking interruption and stale-speech failures.
what this does not show
Different products and measurement points; round 2 reuses the earlier OMP control, and the later bidi subset covers only two scenarios.
- draft study PRpapers & ideas2026-10-20
camera breath-phase tests remain synthetic
Can a dependency-free camera-motion estimator track breathing phase on-device?
Synthetic reproduction reported 26 JVM tests: 25 pass, 1 skipped; the full app suite reported 6755 tests, 0 failures. The recorded-clip result is null: no real-video result yet.
why it matters · The real-clip evaluator is the skipped test, so an otherwise green suite cannot support the intended camera claim.
what this does not show
Synthetic signals and rendered frames only; real-video phase accuracy and on-device latency are unmeasured, and literature respiration-rate results are not phase validation.
- draft study PRlaws & governance2026-10-19
Protected scoring and frozen incumbents
Can an adaptive evaluator refuse improvement credit when the holdout or incumbent was not independently frozen?
The frozen diagnostic reports no-skill baseline 0 / 2 and explicit-proposal body 0 / 2, both with 0 actual inspection calls and group-macro score 0.0. The scorer returned DISCARD, delta 0.0; the newer guard binds holdout bytes before queries and uses exact fractions for comparisons.
why it matters · A high score without a pinned comparable baseline receives no sealed comparison credit; an exact tie cannot become KEEP through floating-point noise.
what this does not show
The negative diagnostic used private source material; only published aggregates are retained. Same-user files are not an independent custody boundary, authentic incumbent execution still needs owner evidence, and the latest scorer guard is unmerged.
- proposalhermes & harnesses2026-10-19
bounded turbo orchestration
Can parallel lanes improve verified throughput without hiding quotas or worker failures?
The RFC specifies default bounded concurrency 3 and one outstanding request per provider; its matched one-trial A/B is proposed, not a reported speedup.
why it matters · The useful target is verified outcomes per bounded run, not launched workers.
what this does not show
Open RFC, no frozen result or measured throughput in the issue; worker coordination ownership was subsequently narrowed outside the cognition component.
- idea / paperpapers & ideas2026-10-16
separate reasoning-gate changes from label changes
Would initial and post-reasoning distributions clarify an adaptive observer comparison?
The external comment proposes preserving initial, post-reasoning and returned distributions, reasoning frequency and gate flips, alongside probability discrimination and latency.
in focus · It addresses the all-safe-threshold ambiguity without treating reference agreement as truth.
what this does not show
The commenter explicitly has not evaluated the proposed StartLux/AutoThink comparison on this workload. It is an optional shadow receipt idea, not a local result or endorsement.
- proposalpapers & ideas2026-10-13
verified replay episodes before autonomous fine-tuning
Can a routing specialist improve through training, replay, shadowing and independent outcome gates?
Issue #143 specifies a canonical episode compiler, temporal/lineage-aware holdouts and progression from offline through replay-green and shadow to verified-outcome eligibility. It retains base, SFT, State Packet, retrieval and verified-label ablations.
in focus · The reward and contamination boundaries matter before adding weight updates.
what this does not show
No trained checkpoint, held-out win or closed deployed training loop is reported here. Historical actions and synthetic teacher labels cannot silently become optimal gold. Trainer backends do not own evaluation or promotion.
- idea / paperpapers & ideas2026-10-13
prompt and workflow search compete with weight training
Does GEPA or RSI-style search beat fine-tuning after full optimization cost?
The #143 follow-up proposes incumbent, GEPA, RSI, GEPA+RSI, SFT and combined-search-to-SFT arms on the same canonical episodes and untouched temporal holdouts.
in focus · Training is one intervention, not an assumed destination for every failure.
what this does not show
Proposed experiment only. Include optimizer evaluations and amortized full-system cost per independently verified success; no prompt-search or training win is claimed.
- draft study PRlaws & governance2026-10-10
Functional repair is not qualification
Can the native launcher preserve strict qualification failures while correcting paths and error reporting?
The committed provenance records four offline reporting controls and two archived native transcripts with strict_rejection_set_unchanged: true and historical_results_unchanged: true. The independent oracle declares 4 cases; the registered wrapper’s live qualification is explicitly not rerun.
why it matters · A repaired artifact can pass functional tests while strict qualification remains unresolved; corrected reporting must not rewrite historical failures.
what this does not show
The PR’s historical functional repair pass is not a new live acceptance of the reporting revision. Independent review reports skip-with-success, escape-check and reuse-attestation gaps; same-user policy/hash controls do not establish independent evaluator custody.
- frozen, not yet postedsmall models & routing2026-10-10
negative usage and false booleans can manufacture quota
Can receipt validation keep reconstructed daily capacity and verified outcomes honest?
The public #51 receipts reproduce negative 100 tokens becoming 1,100 remaining against a 1,000 ceiling, and the string false becoming a verified outcome. A local boolean candidate passed 204 tests and 655 assertions in independent checks; the subsequent negative-token candidate passed 205 tests and 297 candidate checks.
why it matters · Accounting validation is a prerequisite to spending expiring free quota safely.
what this does not show
Accepted fixes are local/unpublished according to the receipts, not merged or activated. Missing usage, account identity, freshness, alternate aliases and executor no-paid-spill remain unqualified. No live provider calls or savings are established.
- frozen, not yet postedsmall models & routing2026-10-10
an empty import is not zero usage
Can native usage events be fed directly into the quota receipt loader?
The denied-network contract check emitted a 120-token nested usage event and loaded zero rows through the flat receipt loader. A documented flat-receipt control loaded 1 row and retained 120 tokens.
why it matters · Schema mismatch can look like unused quota unless empty coverage stays unknown.
what this does not show
This falsifies direct raw JSONL wiring only. It does not establish that no converter exists, identify an upstream defect or qualify real integration. No adapter or provider call was exercised.
- in progresssmall models & routing2026-10-10
local models need correlated shadow coverage before routing credit
Do local specialists produce bounded, observable decisions on real native tool events?
The #20 lane reports 3 of 3 observed calls for each of functiongemma_270m and hammer2.1_3b in a bounded native probe. Subsequent timeout repairs report 41 shadow tests, 207 cognition tests and 22 extension tests passed; a cold model load had previously taken about 11 seconds against a 4-second extension budget.
why it matters · Timeout and unavailable coverage must not be scored as a fast or correct decision.
what this does not show
The final small branch was independently verified against fake backends, not live models or a merged multi-feature runtime. The default supervisor was unavailable in the earlier native report. Successful invocation is not decision quality; abandoned never-returning calls retain a known thread-accumulation risk.
- proposallaws & governance2026-10-10
completion and repeated receipts are not verified credit
Can learning credit survive malformed verifier references, replay and later negative outcomes?
The #54 repair receipts report 41 focused tests for the malformed-reference guard and 44 for receipt replay accounting. The replay repair had 3 behavioral RED failures before the candidate. The RFC requires independently joined outcomes and explicit observational versus causal credit.
in focus · An optimizer can otherwise improve its own scoreboard without improving a task.
what this does not show
Those are bounded branch receipts, not full acceptance of the RFC. The later promotion queue says #153 overlaps #149 and #150 needs separately qualified combination evidence. Do not merge duplicate guards or count append-only receipt revisions as new successes.
- draft study PRphone & computer use2026-10-04
voice certification without invented actor results
What evidence can certify voice-control contracts without confusing replay with hardware performance?
Replay-only evidence can produce only PARTIAL, never KEEP; the preregistered prewarm gate requires ≥50 ms p50 useful gain and exactly zero speculative mutations.
why it matters · The wrapper makes missing runtime evidence an explicit partial decision instead of a blended passing score.
what this does not show
The manifest is draft and has no actor measurements; the numbers are acceptance thresholds, not observed gains.
- frozen, not yet postedpapers & ideas2026-10-02
the local observer lead did not transfer
Does a semantic observer add information beyond a simple prior on fresh work items?
The public receipt has 105 exploratory and 142 confirmatory turns. julia_1’s preregistered confirmatory Brier difference was +0.0094 with p=0.67; log-loss was +0.052 with p=0.81, so the test was NOT_CONFIRMED. Reported warm p50 latency was 69 ms for julia_1, 421 ms for laya_421m and 28 ms for nanojev and decider_2b. The failure-prediction target had 0 positives in 653 joined attempts.
in focus · A calibration-looking improvement can fail transfer or merely approximate the base rate.
what this does not show
Fixture-oracle proxy labels, not general real-world task quality. Paid JEV was NOT_RUN. All-safe argmax rows do not prove a lack of probability ranking: the later comment correctly requests AUROC/AUPRC. No backend was promoted.
- draft study PRlaws & governance2026-10-01
AODL admission conformance
Does the structural admission contract reject invalid authority and replay deterministically?
The frozen scaffold declares separate cases for denial families 101–106, deterministic replay and semantic identity distinct from source lineage. Its manifest has claims: [] and the branch contains no generated admission receipt dataset.
why it matters · Authority is pinned to the validated contract rather than caller-supplied limits; structural ALLOW is deliberately not labeled success.
what this does not show
This is a draft structural contract, not provider dispatch or verified task success. The PR reports a scaffold pass, but committed result receipts are absent; Bend is a differential reference, not runtime authority.
- draft study PRlaws & governance2026-10-01
AODL runtime ordering and restart
Does admission become durable before dispatch while restart and drift remain idempotent?
The runtime README specifies admission-before-dispatch, denial without dispatch, restart reuse, idempotent replay and immutable authored lineage. Provider validation is replaced with a deterministic no-op; the manifest explicitly sets provider_execution: false and network_calls: false.
why it matters · The admission-to-dispatch crash window is tested as a separate seam instead of inferring durability from a completed task.
what this does not show
The frozen runner is present, but no generated runtime-proof result is committed on this branch. Its contract cannot establish actual provider admission, credentials behavior or physical execution.
- draft study PRlaws & governance2026-10-01
AODL golden-trace collector
Can one governed root trace keep structural completion separate from an independently verified outcome?
The collector defines stages 1–6 for structural/execution completion and stage 7 for a separately joined verified outcome. The public CANONICAL_OK fixture requires that exact-output verifier; the allowlisted importer copies five sanitized JSON files and never imports raw state.
why it matters · A completed provider response cannot silently become verified success, and a missing joined outcome remains visible.
what this does not show
This branch freezes the collector, schemas and importer, not an imported completed live-canary dataset. Synthetic collector regression passes reported in the PR do not supply missing runtime stages.
- in progresshermes & harnesses2026-10-01
savings depend on the task
do the earlier context savings survive pre-registered larger samples and fresh coding tasks?
The 496-run primary study reports -34.2% billed tokens, -45.4% estimated cost, and success 121/124 -> 124/124. The 208-session long follow-up reports -30.7% estimated cost over lean, but directed recall was -64.2% while multi-bug debugging was +1.3%.
why it matters · A pooled saving hides opposite packet effects and output-heavy versus debugging regimes; later addenda qualify the original optimistic default.
what this does not show
Manifest remains draft with no matching PR found; one host/model and shared QA cache. Prompt gating failed both registered tests; the long follow-up recall subgroup quality guard failed (12 -> 11).
read the postposts/claude-code-savings-v1.mdstudies/claude-code-savings-v1/manifest.yamlstudies/claude-code-savings-v1/README.mdstudies/claude-code-savings-v1/data/summary.jsonstudies/claude-code-savings-v1/ADDENDUM-v1b.mdstudies/claude-code-savings-v1/ADDENDUM-obspack-long.mdstudies/claude-code-savings-v1/data/results.jsonstudies/claude-code-savings-v1/data/trials.csvstudies/claude-code-savings-v1/data/v1b/results.jsonstudies/claude-code-savings-v1/data/obspack-long/results.json
september 2026
- draft study PRpapers & ideas2026-09-30
Causal attribution remains unresolved
What component attribution is actually supported by the published compiler-versus-unfiltered results?
Four same-backend contrasts cover the same 28 state groups: Hammer2.1 3B +7.14 percentage points, Nemotron Orchestrator 8B -10.71 percentage points, Qwen3.5 4B 0 percentage points and Qwen3.5 9B 0 percentage points. Every compared arm has 0/84 known costUsd readings.
why it matters · A positive contrast, a negative contrast and two nulls survive the same audit; repeated attempts are not promoted into independent experimental units.
what this does not show
These are recomputed source-label contrasts, not reproduced model measurements, independent outcomes or causal effects. Work-item lineage, folds, label-generator identity and original runtime/configuration pins remain unknown.
- in progresshermes & harnesses2026-09-30
less prefix, less reported cost
which context levers reduce claude code reported cost while preserving checked task quality?
The pinned held-out comparison reports stock+raw $2.186 21/22 versus lean+packet $0.861 22/22 (-61%). ObservationPack cut long-recall total cost 30.3% with 12/12 verified in both arms; short-task medians were -1.3% / +0.3%.
why it matters · Prefix reduction and useful state compose, while compression is null on short tasks; a cheaper model without state can cost more.
what this does not show
Manifest remains draft with no matching PR found; small samples and interim looks. Costs are Claude Code estimates, not invoices; packet results are manifest/README source reports rather than rows in the imported summary.
- in progresslaws & governance2026-09-30
three small decision tests
which local tool-caller, recovery controller and merge rule survive frozen comparisons?
No tool-calling candidate beat Qwen3-8B (152/184). The recovery controller cost 0.288 versus constant 0.368, while the selected merge rule auto-merged 65/91 holdout events with 11 bad (0.169 versus bar 0.115).
why it matters · The recovery primary endpoint passed without passing the MLP comparison; green checks often were not code checks, and the chosen auto-merge rule failed holdout.
what this does not show
Manifest remains draft with no matching PR found; private single-operator histories are aggregate-only. The tool-call gate had a gold defect, recovery content audit is absent, and merge labels mostly measure fixes within three days.
- in progresslaws & governance2026-09-30
authority belongs at the action
can local routing and prompt-level grants make bounded agent work cheaper without unauthorized actions?
The learned router tied generic Qwen3-8B at 23/59 versus deterministic 8/59. Action-level authority caught 112/134 unauthorized privileged calls (83.6%), with 12.2% sample friction and 34 ms hook p95; both safety and sample-friction bars failed.
why it matters · A turn-level grant cannot inspect actions chosen later in the turn, but moving the check to the tool call did not by itself make enforcement safe.
what this does not show
Manifest remains draft with no matching PR found; one operator, LLM-labelled gold and concentrated historical calls. Shadow-only; resource-posture results remain pending; local operating cost was unmetered.
- in progresssmall models & routing2026-09-30
a format pass failed faithfulness
can local summaries replace frontier work, and are there enough verified outcomes to learn a gate?
The format-and-facts summary check passed 60/60 locally, but faithfulness found 42/160 (26.3%) hallucinated items versus Haiku 21/159 (13.2%); no class remained speed-qualified. The verified loop had 6 analysis rows, 1 not-success and 1 group, so its result was INSUFFICIENT_DATA.
why it matters · A format verifier admitted unsupported summaries, and a pipe-exit bug contaminated success labels; learning is evidence-bound, not learner-bound.
what this does not show
Manifest remains draft with no matching PR found; one principal/day, small fixtures and one LLM judge. Loop timelines are projections; utilization is an unpinned source-reported operator run.
- draft study PRphone & computer use2026-09-30
agent-web boundaries passed; model evidence was blocked
Can deterministic integration evidence establish a safe boundary without being mistaken for live quality evidence?
The context campaign recorded n_rows 18 and n_live 0, with eligible_for_assist_review false; reliability recorded 7 observed receipts and passed true. The model-backed smoke recorded credentials_available false and executed false.
why it matters · Green transport and privacy checks coexist with a blocked model study; the evidence classes cannot substitute for each other.
what this does not show
Authored deterministic fixtures are not live customer samples or model-quality evidence; active routing and assist promotion remain blocked.