I am an independent AI systems researcher focused on reliable machine agency. My foundation is decades of hands-on systems engineering. How can models become systems that perceive, reason, act, coordinate and recover under real constraints, with verification and measured cost?
Researcher Profiles & Identifiers: ORCID 0009-0005-7810-5077 · Hugging Face (@Cisco801) · Kaggle (@cisco801) · GitHub (@cisco801) · Lab: RoamingPigs
My frontier-model research focuses first on inference-time reasoning, verifier reliability, and evaluation, followed by controlled model adaptation and learning experiments. Current findings are bounded studies; they do not establish frontier-scale training, frontier-model performance, or independent peer review.
Portfolio status · 21 October 2026 Pacific: this agenda connects acoustic research, synthetic-speech detection, a design-stage energy project, documented specialist-model experiments and a verifier-reliability proposal. TuneHarness includes experiments run under its former VastOps name; the latest portfolio update also records a separate Bootscry setup of a Libre Computer single-board computer and newer Noise or Voice diagnostics. These project updates do not change the verifier study status: the frozen 24-task, 576-call H1 collection has not run, serving identity and collector admission remain pending, and no confirmatory result is claimed. Completed experiments retain their own methods and limits.
How I test a research claim
I start with a testable hypothesis, compare it with simpler alternatives, and keep the evidence that changes my conclusion.
- Define the comparison. State the hypothesis, task and failure criterion. Compare simpler baselines under matched information and budgets; distinguish my implementation work from upstream models, data and methods.
- Freeze conditions before outcomes. Record exact source, model and data identities, splits, scoring rules, stopping decisions and resource limits. Keep development and calibration separate from held-out evaluation.
- Retain and expose the evidence. Keep failures, abstentions, unrun cases and unknown costs in the registered accounting. Link versioned artifacts and descriptive uncertainty, and state which records an authorized reviewer can actually rerun.
A source-only guard, a load-qualified runtime and a completed task experiment support different claims. My self-published reports and same-workflow reviews remain separate from outside peer review and independent reproduction.
Architecture draft: machine agency infrastructure
How can separately operated models, agents, compute tools and physical interfaces cooperate with explicit permissions, evidence and recovery? My machine agency infrastructure draft maps implemented interfaces and proposed connections. The components do not yet constitute one integrated autonomous platform. This is an architecture draft, not an empirical reliability result or a claim of frontier-model transfer.
Canonical preprints & working papers
Working papers, registered study protocols, and technical reports authored for reproducible peer scrutiny. Each paper includes explicit mathematical formulations, bounded empirical findings, replication instructions, and BibTeX citations.
Working paper · research status and methodsAbstract: This working paper records an open question about verifier dependence under fixed inference budgets, an authored information ceiling, and missing-outcome bounds. The separate 120-task HumanEval engineering pilot is descriptive; the prospective H1 study remains unrun. A prior 504-task companion is disclosed as a seeded synthetic simulation, and its covariance, policy-rate, distribution-shift and token-conservation outputs are withdrawn as empirical evidence.
Artifacts & Replication: Guarded local collector, networkless verification replay, exact synthetic mathematical diagnostics, reproducible finite-regime maps, and a retained seeded simulation package for exercising analysis code. Its 504 task-metadata rows and 4,032 synthetic candidate records are not generated or tested benchmark solutions; the three labels reuse the same fourteen domain names.
Evidence boundary: The separate 120-task HumanEval engineering pilot reports 109/120 first-candidate successes and 91.39% under a higher candidate cap, with a descriptive interval including zero gain and one task of candidate-pool headroom. It does not test the joint-verifier hypothesis. The prior 504-task companion is a seeded synthetic simulation: its reported covariance and pass rates are not empirical results, its labels do not establish held-out-family shift, and its 0.85 early-stop threshold is unreachable. The proposed H1 collection remains unrun.
@article{caceres2026verifier, author = {Caceres, Cisco}, title = {Verifier Reliability Under Budgeted Code Generation: Research Status and Simulation Limits}, journal = {Cisco Caceres Research Working Papers}, year = {2026}, url = {https://ciscocaceres.com/downloads/verifier-reliability-budgeted-code-generation/}, orcid = {0009-0005-7810-5077}, note = {ORCID: 0009-0005-7810-5077} }
Preprint · Working Paper & Empirical BenchmarkEarlier-pilot diagnostic summary: E1 retains 27 authored synthetic input clips across 6 models and two decoding regimes (324 output evaluations). Overall per-model ungated rates are 24–55 words per minute; 160 is a condition-cell peak. The gate reduces overall WPM by 19.5–25.0%, leaving hallucinations on 108/144 non-speech evaluations. E2 retains 5 authored synthesized utterances across 30 declared conditions and 6 models (900 model-by-condition records). Mean WER across recorded models and paths is 129.4% at 0 dB. These are repeated views of a small synthetic input set, not independent samples or authentic Coast Guard recordings. No population confidence interval, genuine-voice preservation, codec-compliance guarantee, operational-radio accuracy or proprietary AMBIE model accuracy is established.
Artifacts and scope: Retained E1/E2 output ledgers and a separate 1,050-file authored TRCS fixture package. No claim of authentic Coast Guard test recordings, independent population sampling, codec compliance or proprietary-model accuracy.
Full-study update — 11 October 2026 Pacific: I completed a separate registered evaluation of six third-party models: 5,760 E2, 1,620 E1 and 1,440 speech-cost records from 1,050 authored fixtures. These are repeated model/condition evaluations, not 8,820 independent samples.
For each model, the pipelines emitted normalized words on 71/71 distinct non-speech clips without gating and 50/71 with energy gating. The neural gate rejected all 71 distinct non-speech clips before ASR; this is a gating result, not model robustness on those inputs. On synthetic speech, pooled word error rose from 58.85% to 67.47% with neural gating; the paired bootstrap resamples 40 authored utterances, giving a 95% interval of 1.62–16.00 percentage points for the increase. Energy gating left the pooled point estimate unchanged. These fixture intervals establish no population inference.
Local snapshot replay reproduced the analysis and report, and all input-file hashes matched. This was verification within my workflow, not independent replication or peer review; it establishes neither field nor proprietary-model performance. Full-study results and limits · Codec study.
The PDF and online paper linked below retain the earlier pilot; the registered full study uses separate data and results.
@techreport{caceres2026trcs, author = {Caceres, Cisco}, title = {Tactical Radio Channel Simulator (TRCS): Characterizing Non-Speech Audio Hallucination in Speech Foundation Models}, institution = {Cisco Caceres AI Systems Lab}, year = {2026}, url = {https://ciscocaceres.com/downloads/tactical-radio-channel-simulator/}, orcid = {0009-0005-7810-5077}, note = {ORCID: 0009-0005-7810-5077} }
Historical diagnostic · Automated annotation illustrationHistorical Coding Diagnostic and Programmatic MAST Illustration
Scope: The separately recorded development diagnostic retained 48 attempts on reused, exposed fixtures: single agent 8/16; ordinary two-worker messaging and Echo each 0/16. It establishes no coordination benefit or fresh benchmark performance.
Correction · 11 October 2026 Pacific: The companion MAST script produces two label columns using deterministic rules in one program. Those columns are not independent annotator judgments. Its per-attempt token and latency rows allocate aggregate totals by hash weights rather than measuring each attempt. The prior independent-coding, kappa 0.941 and causal failure-taxonomy claims are withdrawn. The original aggregate results remain separate.
Next evidence: Independent blinded coding of permitted retained records and a prospectively frozen fresh-task comparison.
@techreport{caceres2026mast, author = {Caceres, Cisco}, title = {Historical Coding Diagnostic and Programmatic MAST Illustration}, institution = {Cisco Caceres AI Systems Lab}, year = {2026}, url = {https://ciscocaceres.com/research/#paper-mast-coordination}, orcid = {0009-0005-7810-5077}, note = {ORCID: 0009-0005-7810-5077} }
Exploratory profiling · Mixed summaries and calculated scenariosLocal Inference Profiling: Measurement and Modeling Boundaries
Scope: The profiler contains a live local-model and smart-plug sampling path, plus calculated quantization, bandwidth, power and GPU-cost scenarios. Most retained benchmark rows are formula outputs.
Correction · 11 October 2026 Pacific: The prior 38.4 tokens/second, 108.4 W and 2.82 J/token headline came from a calculated profile, not the retained live 9B summary. Three separately labeled live summaries remain available, but their raw timestamped power arrays were not retained in the named ledger. Their integration windows and sensor selection cannot be independently reconciled from that ledger alone. Bandwidth saturation, efficiency and datacenter-cost comparisons are modeling assumptions, not measured superiority.
Next evidence: Freeze an acquisition protocol and retain synchronized model, token, elapsed-time and raw power records before reporting measured energy or matched hardware comparisons.
@techreport{caceres2026sovereign, author = {Caceres, Cisco}, title = {Local Inference Profiling: Measurement and Modeling Boundaries}, institution = {Cisco Caceres AI Systems Lab}, year = {2026}, url = {https://ciscocaceres.com/research/#paper-sovereign-economics}, orcid = {0009-0005-7810-5077}, note = {ORCID: 0009-0005-7810-5077} }
Synthetic simulation · Empirical routing study pendingAuthored Routing Simulation: Cost and Selection Mechanics
Scope: The routing harness generates task descriptions, token counts and tier-success outcomes from seeded assumptions, then compares selectors against those generated outcomes. It makes no benchmark model calls.
Correction · 11 October 2026 Pacific: The earlier production-trace benchmark, measured accuracy, confidence-interval and statistical-superiority descriptions are withdrawn as empirical evidence. An appended small real-trace summary did not determine or calibrate the generated success labels. The retained simulation can exercise routing mechanics; it cannot establish model quality, production savings or an architecture advantage.
Next evidence: Evaluate frozen policies on permitted real task outcomes with pinned models, complete usage accounting and a separately declared holdout.
@techreport{caceres2026routing, author = {Caceres, Cisco}, title = {Authored Routing Simulation: Cost and Selection Mechanics}, institution = {Cisco Caceres AI Systems Lab}, year = {2026}, url = {https://ciscocaceres.com/research/#paper-gateway-routing}, orcid = {0009-0005-7810-5077}, note = {ORCID: 0009-0005-7810-5077} }
Authored outcome simulation · Real execution study pendingRepeated-Trial Reliability: Simulation and Prospective Evaluation
Scope: The current harness draws Bernoulli outcomes with assumed success rates and explicit architecture multipliers. It does not execute agents, merge repositories or observe coordination failures.
Correction · 11 October 2026 Pacific: The earlier empirical trial-count, pass-rate and Echo-benefit claims are withdrawn. Their differences are injected simulation assumptions. Intervals computed from those draws describe the simulator and do not establish real-world reliability, merge-conflict mechanisms or coordination gains.
Next evidence: Freeze fresh task environments and repeated-run conditions, then record actual executions, failures, accepted artifacts and cost across matched arms.
@techreport{caceres2026repeated, author = {Caceres, Cisco}, title = {Repeated-Trial Reliability: Simulation and Prospective Evaluation}, institution = {Cisco Caceres AI Systems Lab}, year = {2026}, url = {https://ciscocaceres.com/research/#paper-repeated-trials}, orcid = {0009-0005-7810-5077}, note = {ORCID: 0009-0005-7810-5077} }
Research projects
Each project has its own evidence and next experiment. Research includes proposals, design work and empirical investigation; their stages remain distinct.
Research proposal · 120-task engineering pilot complete; H1 study unrunQuestion: Under a fixed total inference budget and task-family shift, does modeling cross-verifier error dependence improve hidden-test success when selecting candidates and deciding when to stop, beyond strong single-judge, fixed-consensus and matched-retention filter baselines?
Method: A completed 120-task HumanEval engineering pilot is reported separately. The earlier 504-task companion is a deterministic synthetic policy simulation, not a benchmark model evaluation. The prospective H1 collection separately freezes MBPP source development/calibration and a BigCodeBench target diagnostic; that collection remains unrun.
Artifacts: Updated proposal, related-work matrix, task-label manifest, guarded local collector, networkless VM outcomes and public aggregate results for the separate 120-task HumanEval engineering pilot (480 candidates and 480 verifier calls). The earlier 504 task-metadata rows and 4,032 candidate records are retained as a seeded synthetic simulation: candidate correctness, public-test outcomes, verifier scores, costs and latencies were generated, not measured. No models or benchmark tests ran. The three named splits share the same 14 domain labels and are not a held-out-family evaluation. The stated 0.85 early-stop threshold is unreachable under the implemented score formula. The prospective 576-slot H1 collection remains unrun.
Evidence: The separate 6 October 2026 HumanEval pilot passes 109/120 tasks with the first candidate; higher-cap SoftBoN reaches 91.39%, a descriptive +0.56-point difference with interval [0, +1.67]. Candidate-pool headroom is only one task in that pilot. The earlier 504-task Pillar 1 figures (including covariance, pass rates and cost outputs) are withdrawn as empirical evidence: source audit shows they are deterministic outputs of a seeded synthetic simulator using benchmark task metadata. They do not measure model/verifier behavior, code correctness on benchmark tests, held-out-family generalization or a VEC-SCR gain. The simulator’s three splits reuse the same 14 domains; its 0.85 stopping threshold cannot be reached (the implemented public-pass score is at most 0.75), so no early stopping or token conservation occurred. The joint-verifier and natural-shift H1 study remains unrun.
Next milestone: Reconcile serving model/tokenizer/template identity, qualify and execute the frozen 576-slot development collection locally, then fit source-only controllers and matched baselines, measure candidate-pool headroom and cost, and freeze the preregistration before any confirmatory run.
Read the evidence and method
Separate 120-task HumanEval pilot: methods, results and limits.
Model research · Bounded synthetic-input speech diagnostics recordedQuestion: Which acoustic and deployment choices make noisy-speech recognition more reliable under radio and environmental constraints?
Method: Retained local speech-model output diagnostics on authored non-speech inputs, synthesized speech and declared channel/noise transforms. Exact launch/model-byte provenance and AMR/GSM codec versus emulation paths remain unqualified.
Artifacts: Dated third-party speech benchmark, reviewed planning assumptions, commercial-engine passthrough service, E1 and E2 per-model/condition output ledgers and a separately released 1,050-file TRCS fixture package. The release and the smaller evaluation input sets have distinct denominators.
Evidence: E1 retains 27 authored synthetic input clips across 6 models and two decoding regimes (324 output evaluations). Overall per-model ungated rates are 24–55 words per minute; 160 is a condition-cell peak. The gate reduces overall WPM by 19.5–25.0%, leaving hallucinations on 108/144 non-speech evaluations. E2 retains 5 authored synthesized utterances across 30 declared conditions and 6 models (900 model-by-condition records). Mean WER across recorded models and paths is 129.4% at 0 dB. These are repeated views of a small synthetic input set, not independent samples or authentic Coast Guard recordings. No population confidence interval, genuine-voice preservation, codec-compliance guarantee, operational-radio accuracy or proprietary AMBIE model accuracy is established.
Next milestone: Retain model, launch and codec-path receipts; freeze licensed, speaker-diverse real-domain inputs and critical-field scoring before assessing proprietary or operational performance.
Read the evidence and method
Whitepaper and design · Implementation pendingQuestion: What does machine intelligence cost in energy, and which proposed computational mechanisms can be tested against physical constraints?
Method: Proposed continuous-time dynamics studies, followed by a digital implementation and comparisons with conventional computation before any hardware experiment.
Artifacts: Written whitepaper and design materials; public project description.
Evidence: No software or hardware implementation and no measured results. Proposed energy figures are not demonstrated performance.
Next milestone: Define a minimal falsifiable experiment, implement it and measure it against a conventional baseline.
Experimental acoustic research · DSP and SSL diagnostics completedQuestion: Can interpretable acoustic features help distinguish synthetic speech across generators, speakers and recording conditions?
Method: Complete-recording DSP analysis with separately frozen human-only stress, ODSS and ASVspoof 5 two-class diagnostics; historical detector snapshots, unchanged thresholds and retained failures.
Artifacts: Browser demo, feature analysis, internal test methodology and documented failures. Completed 64-record ODSS and original 192-record ASVspoof 5 diagnostics; full acquisition, PCM/WAV replay and blind scoring deposits precede label analysis. Aggregate reports retain false flags and misses. A separate developer-selected, known-label 168-trial paired diagnostic compares the frozen JavaScript DSP rule with an author-released SSL checkpoint; the committed per-trial ledger reconciles all overall and dataset-level counts.
Evidence: Scores are uncalibrated and verdicts do not authenticate speakers. The ODSS diagnostic evaluates a pinned historical JavaScript research snapshot, not automatically the current deployed detector. Its four test-speaker clusters and descriptive interval are fragile. Earlier human-only failures remain separate; two-class results do not validate the detector, streaming, radio use or general-population performance. ASVspoof selected labels cover two base TTS families, a related adversarial derivative and shared vocoder. In that separate internal paired diagnostic, DSP classified 123/168 correctly versus SSL 140/168. DSP falsely flagged 30/65 human recordings and missed 15/103 synthetic recordings; SSL had 26/65 false flags and 2/103 misses. DSP analyzed complete WAVs while SSL used the first 64,600 resampled samples or repetition/padding. Nonrandom selection, previously exposed labels, related sources, unresolved upstream label custody and training exposure prevent blind-validation, population-performance or general-superiority claims. This is separate from the original 192-record ASVspoof diagnostic.
Next milestone: Independently reproduce the historical-detector diagnostics, qualify the current deployed implementation separately, and extend acoustic coverage under a new prospective protocol. Resolve synthetic-content reuse scope, source linkage and training exposure before broader claims.
Read the evidence and method
Selected-168 DSP/SSL diagnostic artifact.
Operating research infrastructure · Documented ML experimentsQuestion: How can specialist-model training, adaptation and evaluation improve task outcomes within measured compute budgets?
Method: Specialist task-method comparisons, adaptation and compression ablations, speech and diagnostic benchmarks, and bounded single-node multi-GPU trials with per-arm evaluation and cost accounting.
Artifacts: Training and provisioning tooling, a dated cost ledger, model experiment matrices, and per-arm ablation reports, including the former VastOps work. Completed work includes saved-output tool-calling comparisons, a 120-task local verifier pilot, native FunctionGemma checkpoint diagnostics, a 160-case authored full-schema comparison, and bounded native environment export and restoration checks. These are distinct experiments and prerequisite stages; their full numerical record is retained below.
Evidence: The comparisons show task-specific differences and failures, including FunctionGemma invocation collapse in a lossy native condition and a separate JSON endpoint audit. Training-content overlap, earlier probe exposure, omitted schemas, authored fixtures, and unidentified original SFT lineage prevent causal training-benefit or broad generalization claims. Single-node multi-GPU work does not establish multi-node or arbitrary-scale training or marketplace-wide rates. Environment export and finite restoration checks qualify prerequisites, not serving identity, exhaustive ELF closure, collector admission, or a verifier-study result. The complete dated record and unresolved dispositions remain below.
Next milestone: Validate specialist findings on independent real data and held-out domain shifts with a frozen protocol.
Full experiment and prerequisite record
Artifacts: Training-run tooling, dated provisioning ledger, model experiment matrices, ablation reports and per-arm metrics, including the former VastOps work. Exact saved-output rescore, recovered ablation provenance, local verifier diagnostics and a checked mathematical reference. A completed 120-task local verifier pilot with public aggregate analysis and explicit evidence limits. A separately reconstructed names-ablation CPU contract probe completed eight cases: six correct, all eight parseable and strict-format compliant. This is not the missing original SFT checkpoint or a historical generation reproduction. A completed matched 400-record comparison retained 69.7222% class-balanced correctness for the reconstructed names-ablation versus 34.7222% for publisher Qwen3-1.7B; pooled correctness was 254/400 versus 74/400. A separate FunctionGemma base/checkpoint-600 diagnostic qualified finite tensors and repeated synthetic forward outputs, not task accuracy. A later tokenizer-only native fit retained all 400 historical FunctionGemma inputs without oversize; a separate incomplete schema inventory covered 381 records with 19 unsupported, not task accuracy. A completed same-lossy-native FunctionGemma comparison retained both 400-case arms: base 228/360 call and 0/40 no-call correct; checkpoint-600 0/360 call and 40/40 no-call. This was checkpoint invocation collapse, not improved tool calling: macro 31.6667% to 50%, pooled 57% to 10%, with eight checkpoint output truncations. A subsequently completed publisher full-schema + instruction control retained 235/360 call and 27/40 no-call correct, macro 66.3889% and pooled 262/400, with 138 errors. A separately frozen argument-sensitivity rescore reconciled all original labels and changed zero of 1,200 correctness labels across the three Qwen conditions. A post-hoc retained-output audit subsequently replayed all 800 native labels unchanged. Exact stored strict JSON scored 0/400 in each arm. Removing exactly one terminal <eos> only when the recorded stop ID is 1 gave checkpoint-600: 159/360 call and 36/40 no-call correct, pooled 195/400 (48.75%) and class macro 67.0833%; base remained 0. Typed/legacy equality agreed, and all eight call truncations stayed incorrect. This alternate JSON endpoint qualifies blanket incapacity interpretations, not the original native invocation-collapse table; it establishes no causal improvement, equivalent full-schema comparison or external peer review. A later matched full-schema and instruction diagnostic on 160 authored structural fixtures retained reconstructed names-ablation 160/160 versus publisher Qwen3-1.7B 132/160 correct (strict contract validity 160 versus 137). All 320 requests completed without truncation. These fixtures were authored after prior findings; unresolved training lineage and toy structure prevent causal training-benefit or frontier-generalization claims.
Evidence: Completed experiments retain task, dataset, simulator and hardware limits. The reported structural-pruning/recovery arm was planner-refused before training. Single-node multi-GPU execution does not establish multi-node or arbitrary-scale training; provisioning samples are not marketplace-wide rates. The eight-case probe remains separate. The completed names-only comparison retains five exact training-content overlaps and eight earlier probe exposures; omitted argument schemas may favor a catalogue-trained specialist. The original selected SFT remains unidentified. This establishes neither causal training benefit, general tool-calling superiority, frontier transfer nor production readiness. FunctionGemma evidence includes load/finite-forward qualification, native input fit and incomplete schema inventory; these earlier load/fit/audit stages did not establish task accuracy. The later completed base collection remains a separate exposed, lossy-native condition; the completed comparison showed checkpoint invocation collapse with zero correct calls and eight truncations. No-call measures invocation absence, not semantic refusal quality; the cause is unknown from aggregate review. Equivalent full-schema information is not established. Checkpoint-400 remains truncated. Full schemas and instruction both changed; the exposed historical prefix and different model catalogue priors remain. This follow-up is not causal training benefit or fresh generalization; strict argument sensitivity is distinct from strict-format validity. The original native invocation-collapse finding is endpoint-specific; the later one-terminal-EOS JSON audit does not replace it or establish general model incapacity. The 160-case follow-up is authored after earlier findings, not a blind broad benchmark or proof of causal training benefit. A V5.1 native environment diagnostic matched 168/168 pinned package versions but failed 1 of 12 ABI/import checks: Numba 0.55.0 could not import pkg_resources with setuptools 84.0.0 pinned. All 24 references and 9 controls were skipped; no H1 evidence resulted. A prospectively amended V7.2 run changed only runtime setuptools to 81.0.0: all 168 package identities, five source builds, twelve fixed ABI checks, 24 references and nine expected control outcomes qualified. No model calls occurred. V7.2 retained only the runtime inventory. A subsequent V5 run retained 2,646,208,612 regular-file bytes and 45,730 runtime entries in private host custody after repeated full prerequisite qualification and owned guest shutdown. A subsequent owned, networkless restoration replay reconciled eight registered native-file byte identities, all 168 package versions, 12 ABI checks, 24 references, nine registered controls and the unchanged 45,730-entry retained tree. The disposable VM was removed. Five historical source-build proofs were reused, not rebuilt. These finite checks supersede the earlier pending restoration descriptions for that replay; they do not establish exhaustive ELF dependency closure, serving-model/tokenizer/template identity or collector admission. The earlier 1,850 unresolved native metadata dispositions remain retained. Candidate supervision and real collection qualification remain pending. This is prerequisite qualification, not a verifier-study result.
Read the evidence and method
Planned model-method studies
- Inference-time reasoning: how generation, verification and stopping interact under fixed budgets. First proposal published; experiments pending.
- Evaluation and reliability: hidden-test quality, correlated judgment errors, distribution shift and uncertainty. These are questions to measure, not solved problems.
- Efficient model adaptation: planned controlled fine-tuning and ablations against an unchanged base model. No results claimed.
- Learning and scaling: planned small-model training studies that expose data, optimization and compute tradeoffs. Small-model evidence will stay separate from frontier-scale claims.
- Speech and acoustic intelligence: an existing applied specialty and an ongoing research direction; the dated speech benchmark records what changed my assumptions.
Research infrastructure and education
These working systems enable experiments or explain mechanisms. Their engineering evidence stays separate from the findings of studies run with them.
Pre-release coordination · Local path-lease and bounded guest reboot-resume checks passed; worker admission and coding-task benefit unestablishedHow can durable task ownership and path-level coordination prevent collisions and support recovery in cross-machine coding-agent work? Cross-machine, cross-vendor coordination through durable messaging, leased task ownership, repository-scoped path claims, persisted results and advisory file reservations. Echo development itself uses repeated bounded, source-pinned local test cycles. Self-hosted deployments require a user-managed NATS server and a credential for each participating machine. Echo coordinates cooperating processes; it does not sandbox agent processes or synchronize repositories.
Artifacts: CLI/daemon, transport measurements, repository-scoped path leases and bounded dogfood/recovery tooling. The original synthetic local-model run retained 48 attempts with zero accepted pairs. A separately qualified V2 tool-contract revision completed another 48 attempts: single 8/16; ordinary two-worker messaging and Echo each 0/16. Current local tests cover overlap refusal, disjoint-path admission and daemon/broker restart. On 21 October, a network-restricted QEMU TCG guest hard-reset test resumed from the same qcow2 overlay and ext4 checkpoint; all nine phases passed, including 95 runner unit tests and 29 offline guards. No agents or model/API calls were used. Companion MAST labels are generated by deterministic rules, not independent annotations; per-attempt resource rows are allocated estimates.
Evidence: The reused-fixture V2 diagnostic had 8 tasks and 16 attempts per arm: single agent 8/16, ordinary two-worker messaging 0/16, and Echo 0/16. It retained 4 merge conflicts in the ordinary arm and 9 in Echo (13 combined); the older fixture predates the current shared repository path-lease protocol. Current bounded tests refuse overlapping repository path claims, admit disjoint claims and preserve lease behavior across daemon/broker restarts; the restart test passed three repeated runs. The 21 October disposable QEMU TCG guest was hard-reset during stage_ecc and resumed from its same system overlay and ext4 checkpoint while the host remained running. All nine phases, 95 runner unit tests and 29 offline guards passed; five staged binaries matched the verified prebuild. The host scope capped memory at 8 GiB with zero swap, two CPUs and 96 tasks; peak use was about 1.91 GB with no OOM or memory-limit events. No agent or model/API calls occurred. This qualifies guest reboot-resume behavior only, not physical-host power loss, live-worker admission, production deployment or coding-agent task success. The companion MAST script does not provide independent annotations; the prior independent two-pass coding and kappa 0.941 claims are withdrawn. The native host worker guard and finite aggregate cgroup remain unqualified, so worker admission stays closed.
Next milestone: Qualify the native host-wide worker resource guard and finite aggregate cgroup before admitting workers; qualify physical-host reboot recovery separately. After host admission and local serving identity pass, freeze a fresh path-lease-aware task set and compare one worker, uncoordinated candidate selection and Echo-coordinated disjoint work under the same acceptance rules and aggregate resource budget, retaining failures and interruptions.
Read the evidence and method
Working physical-machine toolingSupports studies of observable physical-machine setup, recovery and agent interventions. Out-of-band KVM screen reading and HID input alongside exact text and network checks where available.
Artifacts: Documented ARM-board bring-up, older-PC installation and x86 rescue sessions; Libre Computer setup also recorded.
Evidence: Each ARM and x86 record states its method and limits. The Libre card was prepared on another host, not through Bootscry delivery. Hangs and manual power interventions are retained; these are not a general recovery success rate.
Next milestone: Define repeated hardware trials with intervention counts, verified end states and a scripted baseline.
Read the evidence and method
Educational implementationSupports understanding of network learning, tokenization, backpropagation and inference mechanics. Interactive small-model implementations in the browser.
Artifacts: Working educational demonstrations. A separate causal-attention educational reference passed nine tests on Haywire and Feral; numerical gradient and KV-cache checks are recorded.
Evidence: Teaching implementations do not establish a novel model method or frontier-scale training contribution.
Next milestone: Use a separately versioned training experiment for any future learning or scaling claim.
Read or download the research and systems portfolio for the contribution record and planned publication sequence.
The research agenda and study status live here. Long-form technical reporting also belongs in the RoamingPigs Field Manual.
Verifier study milestones
- Proposal, literature review and the limited-headroom engineering pilot — published 6 October 2026.
- Scoped restoration replay — completed. Serving model/tokenizer/template and candidate qualification, new-family candidate diversity and matched baseline feasibility — next.
- A separate source/calibration pilot, power analysis and frozen preregistration for the joint-verifier hypothesis — planned.
- Confirmatory evaluation, technical report and independent reproduction — conditional on those gates.
I welcome specific criticism of the protocol and collaborators interested in verification, code generation and reproducible inference research. Contact me about the study.