code-v107: SNN Sound Companion learning and memory

Sound Companion now learns labeled microphone sounds with an SNN. Both input synapses and the class readout train, older taught examples are replayed for retention, and trained parameters reload without fitting. The SNN chooses action labels; a separate example-support guard can reject them. The previous matcher runs as a comparison control. Teaching, correction and forgetting run outside the Tk event loop.

Existing sound memory is imported while preserving the original file. Fresh test recordings score both backends without training or triggering actions.

Verification for source c306986c3f5035ccb2bd98fc0a195ee46b466c1b:

• 87 distinct affected checks verified: 77 unchanged non-window checks reused from an 86-test pass, plus all 10 final native-window checks after layout fixes. Exact receipts and source hashes are retained; this was not another combined 87-test run.
• Frozen comparison: each of three SNN seeds scored 48/48 after teaching four generated sound families and after reload. The matcher tied. The same 48 clips were reused across seeds; these are not microphone accuracy results.
• Mean software decision time, including scheduled controls: SNN plus guard 2.07 ms; matcher 0.81 ms. No SNN speed or energy advantage is claimed. Raw neural unknown choices still require the rejection guard.
• Independent aggregate verification checked 1,708 numeric assertions and rejected three corrupted reports. All 18 plotted values/denominators were checked.
• Physical microphone calibration, level delivery, clean shutdown and lock release were checked. Personal audio and learned parameter arrays are not included.

This pre-release also includes the microphone checkpoint and training-memory work committed since code-v106; its earlier negative findings remain in the saved reports. No full local suite or current-source full-CI acceptance is claimed, and remote CI was not awaited.

Run and teach sounds · Comparison and limitations · Verification evidence

code-v58: audit amendments verified

Complete the affected audit validation on the code-v53 runtime and close the repaired rich-surrogate issue #66.

• Fast: 270 tests OK in 222.641s.
• Slow: 64 tests OK in 3399.418s.
• Retry: 3 tests OK in 653.939s.

All lanes ran sequentially with no exclusions on 105 unchanged inputs (476d213151b3998477739267dee834bf4ac604cae8e41c81870fdd7cba3c8e73). CI fast passed on Python 3.12 and 3.13. Local full discovery was not run. Both full-CI jobs remain running at the final status observation and are not counted as passes. Earlier complete failures and interrupted runs are preserved.

Includes the verified nonvacuous rich-surrogate caller, paginated hub asset lookup, and fresh three-seed author/beam and ranker measurements. Conditional neural ordering survives on the measured fixtures; unique spiking benefit, broad transfer, and guided signature-work savings remain unestablished. Hard surrogate fitting/selection has one-task coverage; full search acceptance still runs all eight tasks. Search remains bounded and incomplete. Issues #51-#71 receive the completed results and their remaining limits.

Changes, raw data, logs, and limits.

Observed-context result: frozen comparison and independent replay

Publishes #84's second frozen 8,640-episode comparison and corrected independent replay. The observed-example adapter is measured at ff9a0f3 / code-v97; the repaired auditor is be79fe9 / code-v98. Both source scopes and the original failed audit are retained without rewriting measurements.

At budget 144, context-conditioned persistent SNN accepts 390/864 repeated task/order/seed attempts versus frozen 395, reset 425 and same-run streaming 414. The previous no-context persistent count was 378. All gates fail in all six orders. The small descriptive change does not establish that the persistent-learning regression is fixed.

The cold fixed encoder represents 48 contexts with only 10/6/8 distinct LIF codes across seeds, versus 48 for each continuous control. At full grammar, persistent SNN saves 2.7% pooled candidates but uses 4.29 times streaming CPU; reset saves 13.7% but uses 3.89 times CPU. No SNN-specific, total-work or hardware-energy gain is demonstrated.

Verification: all 8,640 episodes independently replayed, eight corruption controls rejected, measured execution-source hashes preserved. The corrected checkout differs only in the auditor and its test, recorded in auditor-repair.json. All 436 fast tests passed in 235.965 s; 62 affected tests passed in 0.204 s. No skips/exclusions; checks ran serially under the builder lock. Slow/retry and local full discovery were not rerun for this additive path. Newer full CI remains pending and tracked separately.

The adapter stays opt-in. #84 remains open for encoder/reward/exposure mechanisms and real sensor/hardware contracts. Original ranking and pilot results, both frozen input corpora and the first counterbalanced result remain intact.

Observed-context auditor: repaired guard and verified frozen replay

Corrects the independent observed-context auditor's obsolete metadata-key guard and adds an actual pinned-protocol parser regression test.

Verified after the fix: 62 affected tests passed in 0.204 s; all 436 fast tests passed in 235.965 s with unchanged source hashes and no skips/exclusions. Corrected replay passed all 8,640 frozen episodes. Eight semantic/provenance corruption controls were rejected after a positive control.

The measured code-v97 runtime, protocol, raw data, manifest and numerical summaries remain unchanged. Verification uses its immutable source archive plus a separate current-execution hash check; only the auditor and its test differ, and their hashes are retained. The initial failure is preserved. No learning or performance improvement is claimed by this auditor fix. CI for this corrected source is run 39166387866; pending jobs are not credited as passing. Slow/retry and local full discovery were not rerun for the additive diagnostic path.

Observed-example context: tested adapter and frozen comparison

Adds an opt-in observed-example context adapter for #84. It supplies the six current observed integer triples through a fixed, separately seeded projection while keeping the original encoder initialization, action RNG, readout, reward and current-decision update unchanged. Fresh context is required each episode; boundary snapshots retain learning without retaining task examples. Opaque episode IDs are guards only, with an invariance regression test.

Freezes the next comparison before measurement: all six original phase orders, three seeds, five matched controls including same-run streaming, both budgets, 8,640 episodes. Protocol SHA-256: d8897dc237ceed1a8781dd2dd828b4507842bbdf53473558edffa928fe6485fb. It includes a fixed-encoder capacity diagnostic, context/diagnostic CPU costs, and independent replay. Giving the learner current examples changes its information set; no learning improvement or hardware gain is claimed before measurement.

Validation: 61 affected tests passed in 0.210 s; all 435 fast tests passed with unchanged source hashes and no skips/exclusions. The initial synthetic fixture failure is retained. An earlier 434-test run passed before CPU-scope wording and opaque-ID contract clarification. Slow/retry and local full discovery were not rerun for this additive path.

Also retains source-specific GitHub fast CI logs for f4b97fd / code-v95: 420 tests passed on Python 3.12 in 458.615 s and Python 3.13 in 476.132 s. Those results do not verify the new adapter. Existing full CI jobs remain pending. Original runtime, benchmark corpora and negative ranking/persistence results are preserved.

Persistent-learning diagnosis: six frozen phase orders audited

Publishes #84's frozen six-order comparison, measured at f4b97fd / code-v95. All 8,640 episodes were independently replayed; seven semantic/provenance controls were rejected. Original-order traces, acceptance and weight hashes exactly match the prior 1,440-row pilot.

At budget 144, persistent SNN accepts 378/864 repeated task/order/seed attempts versus frozen 395 and reset 415. Persistent learning loses to reset in four orders, wins one and ties one; pooled first-choice counts equal frozen. All frozen gates fail. This supports further mechanism work, not the claim that persistence loses in every order.

At full grammar, reset uses 14.2% fewer pooled candidates than streaming, with fewer candidates in all 18 order/seed runs, but 3.56 times its CPU. Persistent saves 1.5% pooled candidates with mixed per-run results and 4.04 times streaming CPU. No total-work or hardware-energy gain is established.

Validation remains 46 affected and 420 fast tests passed with unchanged source hashes, no skips/exclusions. Slow/retry and local full discovery were not run for this additive diagnostic. The measured source's CI run 39164398208 is separate from earlier source scopes; full CI is not credited as passing. #84 remains open for mechanism and sensor/hardware work.

Counterbalanced diagnostic: replay auditor and 420 fast tests

Adds the independent auditor for #84's frozen six-order comparison. It checks all planned rows, policy/execution/feedback replay, fresh state between orders, diagnostic purity, RNG offsets, and CPU/gate summaries. Raw-attempt replay streams without retaining every graph in memory.

Validation: 46 affected tests passed; all 420 fast tests passed in 234.598 s, with unchanged source hashes and no skips or exclusions. Commands ran serially under the builder lock. Slow/retry and local full discovery were not run for this additive diagnostic path. The initial synthetic-test failure and its correction are retained.

The comparison has not been measured yet. Original runtime, frozen corpora and negative results remain unchanged. #84 remains open.

Authentication source: full Python 3.13 verified

Full Python 3.13 CI now passes for the authentication-fix source: 441 tests in 5910.417 seconds, OK, job 117753752870, run 39154146448, source 59ef1abe06a5a9d0d8e9813252f05c902fdfc729 (code-v87).

Published code-v93, commit 7c0120c, retains the exact completed-job metadata, compressed full log, source tree/input/workflow hashes, dependency checkout SHA and independent verification receipt in artifacts/ci-full-followup-20261018/code87/. The full discovery command remained intact. No local tests or runtime changes were made for this evidence-only checkpoint.

On the same source, both fast jobs already pass 374 tests each. Full Python 3.12 job 117753752836 remains in progress at the retained snapshot. #44 remains open until that version also passes. #82's full-matrix verification remains pending; its completed negative benchmark and issue closure are unchanged.

The newer code-v90 online runtime is a separate source scope: its two fast jobs pass 409 tests each, and both full jobs remain pending. The older source's full result does not verify those newer files/tests. Monitoring continues for both scopes; no pending job is credited as passing.

Online runtime: both Python fast CI jobs verified

Both fast CI jobs now pass for the online-learning runtime at source 5cdb9b5782a37bf0c32bb15d22098399f3ce12df (code-v90), run 39160627937:

• Python 3.12: job 117772980247, 409 tests OK in 459.645 seconds.
• Python 3.13: job 117772980160, 409 tests OK in 469.247 seconds.

The exact compressed logs, job metadata, run snapshot and source/hash receipt are published in code-v92, commit ce02608, under artifacts/ci-full-followup-20261018/code90/. Independent verification confirmed both compressed/decompressed hashes and the RSI and Spikeforge checkout SHAs. Runtime, tests, scripts, fixtures, requirements, justfile and workflow match the tested code-v90 source; this checkpoint changes evidence only. No local tests were rerun.

Both full jobs for the online runtime remain in progress. The separate authentication acceptance run 39154146448 at code-v87 also still has both full jobs pending. #44 remains open; #82's full-CI status is unresolved. This updates #83's status to fast matrix verified, full matrix pending. Neither full matrix is credited as passing.

All published ranking and online-learning conclusions remain unchanged. Monitoring continues for both source scopes.

Online outcome learning: independently audited pilot

Completed the first online graph-learning pilot in code-v91 (commit e601381), following the published code-v89 protocol and code-v90 runtime. The SNN updates its readout directly from executed-graph error without an LLM in the runtime/update loop. The spiking encoder remains fixed.

All 1,440 episodes completed with unchanged source inputs. Independent replay passed, and six corrupted full-bundle copies were rejected after their raw checksums were recomputed. Local validation passed 35 focused tests and all 409 fast tests; no exclusions or skips. The initial audit-fixture failure and its correction are retained. Slow/retry and full discovery were not rerun locally for this additive path.

The result is negative for useful persistent adaptation. At cap 144, persistent SNN accepts 56/144 task-seed attempts versus frozen 67/144, reset SNN 68/144, nonspiking 45/144 and streaming 69/144. All arms accept 144/144 at cap 432. All three preregistered outcome gates fail.

A descriptive secondary observation: task-reset SNN uses 21,560 candidate executions versus streaming 25,128 at full acceptance (14.2% fewer, with the same direction in each seed), but takes 3.41 times the recorded CPU. This is not a total-work, chip-energy or sensing win. The three model seeds repeat one 48-task stream.

The complete result, per-seed/phase summaries, raw evidence, source hashes and reproduction commands are in docs/online-graph-result.md. The supplied paper was read in full; its assessment and exact PDF hash are in docs/online-learning-direction.md.

This closes the implementation/evaluation pilot, including its negative result. #84 tracks the unresolved persistent-learning regression and sensor/hardware feedback contract; #74's broader action/rollback work remains open. No existing negative benchmark conclusion was changed and no weights/databases were committed.

CI for the new code-v90 runtime, run 39160627937, is still pending at publication. The code-v87 authentication acceptance run 39154146448 remains separately scoped to #44, which is not closed by this pilot.