Tend learned a distinction in a real macOS pipe system. Repairing its model changed a later experiment—and saved one acquisition against its preceding model. A competent source-aware procedure tied it.
This is one selected native result, not a complete account of current research.
This is a curated account of completed result records. The website review checked the committed reports and key comparisons; it did not rerun the native experiments, installed acceptance or release qualification. Supporting downloads below are public summaries, not the original execution archive.
Two selected updates on generating alternatives and testing when guidance can change a useful decision. These are research findings with different evidence boundaries from the native case below.
Mechanism demonstrated
Generate alternatives from an earned model
A Rust research prototype generated four alternatives from a model learned through observation. Each alternative changed one output in that two-state model and retained its parentage.
What happened
In the authored example, a contrary observation refuted the parent and narrowed two possible target forecasts to one. A separate actual target observation was still required to establish the answer.
Cost and comparison
Direct observation was cheaper: 2 queries and 6 attempted steps versus 3 queries and 8 steps with the diagnostic, for the same supported answer. Donor preparation cost another 10 queries and 22 steps.
What remains open
The mutation rules, output alphabet and question were supplied. The family stayed fixed after generation. This demonstrates model-descendant generation, not autonomous evolution or a proven improvement in later investigations.
Evidence status and next question
Research-only prototype; no supported runtime or installed release changed. Fifteen focused checks passed, but the producer retained a failed closeout after a stale excluded-test count. A separate saved-evidence reader checked the named passes and trace without rerunning them; the earlier compile failure remains recorded.
A useful limit
Different guidance can share the same blind spot
In a study of how to allocate investigation time, guidance from past experience and current progress pointed to the same next action. A read-only reconstruction confirmed the agreement at both decisions under review.
What happened
The nominations agreed before and after the shared affordability filter. A gate that only chooses between those nominees cannot change either action. Agreement also did not certify a good choice: one prior recorded outcome was worse than a competent ordinary procedure.
Cost and comparison
The check reconstructed 27 saved decision views. It made no native calls, fitted no gate, and acquired no new observations. Retained preparation still carried 2,680.504419 recorded seconds of prior cost.
What remains open
This closes that particular two-ranking intervention, not the broader model-population or transfer idea. Earlier disagreements and unsupported current scores remain in the record; they did not establish an actionable difference at this selected boundary.
Evidence status and next question
This is a reconstruction of exposed saved evidence by one coordinator, not an independently executed applicability test, prospective learning result or installed acceptance. The broader question remains open: when advice agrees, what permitted observation could test a shared assumption before a useful later action?
The target was an owned nonblocking pipe with macOS kqueue notifications. The supplied actions were Write, Drain, Poll and Enable, with defined reset behavior, observation projection and servicing goals.
Starting empty, Tend constructed a three-state model from 52 native queries and 148 actions. It predicted 60 of 64 traces in the first catalogue correctly. An actual contradictory observation led refinement to separate a fourth state: an empty pipe with its dispatched notification still disabled behaves differently from the initial empty, enabled state.
After repair, the cumulative cost was 86 paid queries and 362 actions. The original twelve-action per-query limit remained in place. This is useful structural learning within a supplied representation, not discovery of the action vocabulary or general investigative skill.
02 / Recorded finding
Manual exploration
Follow what changed.
Narrated from completed result records, not a live Tend run. These explanations summarize the method; they are not a captured reasoning trace. State markers show counts, not a transition graph.
1 / Prediction
A three-state starting model
1233 learned states
Starting empty, the learner constructed a three-state model from 52 native queries and 148 actions. It predicted 60 of 64 traces in the first catalogue correctly.
Why this next step?
Evidence
A model learned from actual queries; the catalogue exposed four mismatches.
Uncertainty
A compact model can fit its history and still miss a relevant state. Catalogue accuracy does not prove completeness.
Cost
52 paid queries and 148 actions at construction.
2 / Observation
Reality supplies a counterexample
1233 learned states
The first actual mismatch in the catalogue gave refinement a concrete contradiction to investigate. A forecast remained distinct from what the pipe actually did.
Why this next step?
Evidence
The recorded result identifies the first actual mismatch at catalogue index 5.
Uncertainty
This is a summary of the result record, not a replay of raw trace events or a captured reasoning trace.
Cost
Counterexample acquisition and refinement are included in the cumulative repaired totals below.
3 / Contradiction
Empty does not mean enabled
1233 learned states
An empty pipe whose dispatched notification remains disabled behaves differently from the initial empty, enabled state. The model needed to represent that distinction.
Why this next step?
Evidence
The native repair separated behavior using the distinguishing suffix Poll, Poll, Drain, Write, Poll.
Uncertainty
This distinction depends on the supplied action vocabulary, reset behavior and observation projection.
Cost
The original twelve-action per-query limit remained in place.
4 / Revision
A fourth state earns its place
12344 learned states
Refinement produced a four-state model. Both the preceding and repaired models later scored 64/64 on a different prediction catalogue; that endpoint did not measure the useful difference.
Why this next step?
Evidence
One refinement, two model revisions, and an executable four-state model are recorded.
Uncertainty
A model revision is not automatically a practical win. Its effect on a later decision must be measured separately.
Cost
86 paid queries and 362 actions cumulatively: 34 queries and 214 actions beyond initial construction.
5 / Later choice
Skip a genuinely unresolved dead end
12344 learned states
After Write then Poll, both methods rejected Drain using already paid evidence. The preceding model next tried Drain, Drain; its decisive Poll was still unobserved and returned no notification. The repaired model chose Drain, Enable directly and verified service.
Why this next step?
Evidence
Across four later service jobs: two acquisitions for the repaired model, three for the preceding model. All methods established the same four service outcomes.
Uncertainty
The source-aware procedure ties the repaired model. The saved acquisition does not establish a complete-cost advantage or cross-system transfer.
All methods had the same current paid evidence for checking candidates and stopping. The repaired and preceding models differed only in guidance. Across four later service jobs, every method established the same supported outcomes.
In the distinguishing case, both rejected Drain using an already paid prefix. The preceding model then chose Drain, Drain. Its decisive Poll had not yet been observed; a new acquisition returned no notification. The repaired model chose Drain, Enable directly and verified service. The saved seven-action acquisition was genuinely unresolved beforehand.
Additional work after common history. Preparation is excluded from this table.
Later service comparison
Method
New acquisitions
New actions
New syscalls
Verified jobs
Repaired model
2
12
38
4
Preceding model
3
19
58
4
Direct procedure
2
13
39
4
Source-aware procedure
2
12
38
4
Shortlex search
3
19
58
4
What this does—and does not—establish
The source-aware ordinary procedure ties the repaired model. Superiority over the strongest ordinary method is not established.
The comparison measures additional investigation work after common history. It does not show that learning and repair pay for themselves over the complete lifecycle.
Both models achieve 64/64 predictions on the later catalogue. The demonstrated benefit is query selection at this endpoint, not higher prediction accuracy.
The interface, actions, reset contract and task predicates were supplied. Tend did not discover these representations.
The earlier two-versus-six comparison was invalid because its preceding-model control ignored an already paid prefix. The corrected, evidence-aware result is two versus three.
The producer receipt binds the aggregate result, not every individual supporting file. Later audits do not retroactively authenticate the entire producer chain.
04 / Separate engineering evidence
Retain the model. Continue the same job.
In a tested installed macOS workflow, Tend retained its learned model and an earlier verified service result through a forced interruption, then continued the same job.
macOS arm64; an owned nonblocking pipe and kqueue; a supplied fixed action plan. Continuation used the same job, executable, clock window and original limits. The interruption was a SIGKILL after a native action and before its journal commit.
The original four journal entries and four-state model survived. Continuation charged the interrupted work, retried only when the original reservation allowed it, and established a second verified service result. The installed binary ran outside the source checkout without Cargo, rustc, Python or the research archive.
This demonstrates recovery at the exercised cut. It does not promise recovery at every instruction, after a reboot, under power loss, or after moving a job. It is engineering evidence, separate from the learning comparison.
The supplied actions and schedule are part of the boundary. This was not attachment to an arbitrary application or event loop. There is no supported public installation download on this page.
05 / Supporting material
Readable here. Portable for inspection.
These downloads preserve the selected native comparison, newer research notes, scopes and source identifiers. They are derived summaries, not native logs, raw snapshots, an authenticated execution bundle or an independent replication. No repository access is needed to read them.
The selected native case is based on committed records in jkordish/tend at revision 12255a772a54595200a97969aea63acdb59215f5. Source identifiers support traceability; they do not make the unpublished original archive publicly auditable.
Installed recovery result docs/native-service-session-result-v1.md Blob ad682955ae8435519e159e52e61a5947d0d8f1c8
Installed acceptance record research/native-service-release-v1/installed-v2-result.json Blob 8448b699aba9fa78312b1c555f8839b4694994de
Qualification acceptance record research/native-service-release-v1/completion-audit-v3-result.json Blob 4a6da561505144ee05fd21a51688b0b1cceef7b7
The newer research notes identify their separate source revisions below.
Generate alternatives from an earned model docs/bounded-model-variation-result-v1.md Revision a80bcbe3f1a29369983b85970c1f76ab86362857 · blob f9c9abfd8fc8c76b57a746f1391c61d0852a047a
Different guidance can share the same blind spot docs/learning-curve-applicability-premise-result-v1.md Revision 06246f160e8ec8d4966cf0ba27a5fc7ec57c8b59 · blob b7c064ab3444324b767bd5a0d7129f5d014bd4ca
The installed acceptance record and aggregate qualification record concern engineering delivery. They are not a new independent learning-benefit study. Public summaries omit private local paths and authorization records.
Research on predictive models guiding behavior. It inspires questions and comparisons; it is not evidence for Tend’s capabilities or a claim that Tend uses Dreamer’s architecture.
Have a system worth understanding?
Bring a concrete question, observable behavior and a way to tell whether an answer helped.