Obsah: :: Library Catalog

Uloženo v:

Podrobná bibliografie
Hlavní autor:	pulsaride
Médium:	Recurso digital
Jazyk:
Vydáno:	Zenodo 2026
On-line přístup:	https://doi.org/10.5281/zenodo.20107855
Tagy:	Přidat tag Žádné tagy, Buďte první, kdo vytvoří štítek k tomuto záznamu!

Obsah:

<h1>Release v0.5.0-h7-κ-transfers</h1> Title: H₇-κ Transfers: The One-Tick-Memory Signature Survives a Calibrated Change of Environment, Under a Publicly Audited Runner Date: 2026-05-10 Scientific status: <code>KAPPA_TRANSFERS</code> — the κ signature published in v0.4.0 (one tick of local-pressure memory breaks the σ inversion on E₀) survives an out-of-distribution change to a calibrated diffusive environment E₁ on a fresh, pre-registered seed pool [2000-2029]. Both branches of the pre-registered double reporting (Miller-Madow corrected primary, plug-in transparency) agree at the same p-value with <code>verdicts_agree = true</code>. Operational status: the κ pipeline is now end-to-end publicly executable. The runner is a public reimplementation gated by a bit-level audit against the v0.4.0 reference seeds [1500-1529] under fixed tolerances (atol = 1e-9, rtol = 0, <code>==</code> strict on integer columns). ADR-033 ACCEPTED on 2026-05-09; tag <code>audit-passed-v1</code> placed before the portability tirage. Predecessor: <a href="#release-v040-h7-%CE%BA-reverses">v0.4.0-h7-κ-reverses</a> (κ on E₀; v0.5.0 tests whether the same signature transfers to a calibrated E₁). DOI (v0.5.0): TBD — to be filled after Zenodo deposit; concept DOI inherited from v0.1.0 lineage. <h2>1. Abstract (no marketing)</h2> H₇-κ (v0.4.0) published a paradox resolution on E₀: a single tick of memory on the previous local-pressure observation (M_κ, zero parameters, zero learning) reversed the H₇-σ inverted-coupling signature with $d = +2.66$ at $p = 9.3 \times 10^{-10}$ on n = 30 seeds [1500-1529]. The closing section of v0.4.0 explicitly listed cross-environment portability as out of scope and requiring its own pre-registration. ADR-031 (post-κ trilemma) ranked branch B (portability E₁) first among the three legitimate continuations. ADR-031.bis clarified the binding conditions. ADR-032 pre-registered H₇-κ portability on E₁ under the same statistical chain as ADR-027 (paired Wilcoxon, both alternatives, $\alpha = 0.005$, $d \geq 0.5$, Miller-Madow + plug-in double reporting), on a fresh seed pool [2000-2029] never used in any prior tirage, with M_κ inherited bit-identical from ADR-030. Two independent guarantees were required before the portability tirage was admissible: <ol> <li>Audit gate (ADR-033). The public reimplementation of the κ runner (<code>src/experiments/portability_draw.py</code>) had to reproduce the v0.4.0 per-seed CSV <code>research/h7_kappa_run_results.csv</code> on the reference pool [1500-1529] under fixed tolerances. Iter1 failed strictly on the obs-shuffled witness branch S; R, M_κ and M were already bit-identical. The investigation prescribed in ADR-033 §6.2 #5 (BLAKE2b convention) was executed by exhaustive S-only brute force on witness seed 1500 (90 combinations: key format × digest_size × byteorder). A unique solution reproduced <code>KL_S_M_corr</code> to $3.0 \times 10^{-11}$: <code>key = "obs_shuffle::{seed}"</code>, <code>digest_size = 8</code>, <code>byteorder = "big"</code>. Single-bit fix in <code>src/agents/obs_shuffled_agent.py</code> (commit <code>c15f313</code>). Iter2 PASS: max $|\Delta|$ across all 14 float columns $\leq 4.5 \times 10^{-11}$; 0 strict mismatches across all 15 integer columns. Tag <code>audit-passed-v1</code> placed.</li> <li>Pool freshness. The portability pool [2000-2029] was reserved in ADR-032 and remained frozen until ADR-033 reached <code>ACCEPTED</code>. No pre-tirage observation of E₁ pressure on these seeds was permitted.</li> </ol> The portability tirage was then run on E₁ under <code>--pool portability --i-have-read-adr-033</code>. Result: | Metric | Corrected (primary) | Naive (transparency) | |---|---|---| | Wilcoxon $p_\text{greater}$ | $9.31 \times 10^{-10}$ | $9.31 \times 10^{-10}$ | | Cohen $d$ | $+3.0906$ | $+3.0907$ | | $n_\text{post-drop}$ | 30 / 30 | 30 / 30 | | seeds with $\Delta_s > 0$ | 30 / 30 | 30 / 30 | | total clip events | 0 | 0 | <code>verdicts_agree = true</code>. <code>inconclusive_reasons = []</code>. Verdict: <code>KAPPA_TRANSFERS</code>. The diagnostic numbers replicate the qualitative pattern of v0.4.0 under the new medium: median $\delta_\sigma^R_\text{corr} = -0.18$ (R remains on the σ-inverted side under E₁) versus median $\delta_\sigma^{M_\kappa}\text{corr} = +0.40$ (M_κ still crosses to the structural side); the action-repertoire diagnostic shifts from median $K_R = 35$ to $K{M_\kappa} = 57$, a +22-cell recovery — of the same order as the +23 reported on E₀ in v0.4.0. The κ mechanism — local temporal differentiation — is therefore not an artefact of E₀'s specific pressure dynamics: it transfers across a calibrated change of medium with no loss of effect size. <h2>2. What this release ships</h2> | Path | Description | SHA-256 | |---|---|---| | <code>research/MANIFEST.v0.5.0.yaml</code> | full release manifest (audit gate + portability) | <code>2232e037…ca13b6ee</code> | | <code>research/h7_kappa_portability.csv</code> | per-seed E₁ tirage (30 rows × 29 cols) | <code>b532d938…c02829ee</code> | | <code>research/h7_kappa_portability_verdict.json</code> | adjudicator output | <code>1b7f4ee0…58218dc3d8</code> | | <code>research/h7_kappa_audit_v04_iter2.csv</code> | iter2 audit candidate (30 rows × 29 cols) | <code>4a6815c2…61bd231b0</code> | | <code>research/h7_kappa_audit_v04_iter2.report.txt</code> | audit verdict (PASS, max $|\Delta| = 4.5 \times 10^{-11}$) | <code>386d993e…4fed65a</code> | | <code>src/experiments/portability_draw.py</code> | public κ runner (E₀ audit + E₁ portability) | <code>3c4a7df4…716aa79d</code> | | <code>src/agents/obs_shuffled_agent.py</code> | obs-shuffled witness (BLAKE2b byteorder=big per §11.2) | <code>c5e77295…34f37007</code> | | <code>src/agents/adaptive_agent.py</code> | R policy | <code>19042669…07176d3</code> | | <code>src/agents/memory1_agent.py</code> | M_κ (bit-identical to v0.4.0) | <code>2684b171…e727c7870</code> | | <code>src/analysis/sigma_chain.py</code> | KL chain, histograms, Markov fit, Wilcoxon | <code>414e1463…05d49084</code> | | <code>src/analysis/audit_compare.py</code> | bit-level audit comparator (CLI + dataclass) | <code>a55e3df1…17fac681</code> | | <code>src/analysis/verdict_v05.py</code> | v0.5.0 adjudicator (frozen before tirage) | <code>389528eb…34a36be</code> | | <code>tests/adt/test_portability_draw.py</code> | 12 ADT (runner) | <code>f25c1c07…342e8500abc59b4</code> | | <code>tests/adt/test_audit_compare.py</code> | 12 ADT (audit gate) | <code>a27cb107…182b6cdca2f644</code> | | <code>tests/adt/test_sigma_chain.py</code> | 21 ADT (KL / Markov / histograms) | <code>5a599fad…ccefa943570353b</code> | | <code>tests/adt/test_obs_shuffled_agent.py</code> | 15 ADT (witness; assertion updated to byteorder=big) | <code>91653b03…0c8fc50ce</code> | | <code>tests/adt/test_verdict_v05.py</code> | 13 ADT (adjudicator) | <code>5523cd4d…3a3189c89e5</code> | | <code>docs/adr/ADR-031-post-kappa-trilemma-b-first.md</code> | branch B selection | <code>fb1a52ea…57648e841</code> | | <code>docs/adr/ADR-031.bis-supersede-section-5-3.md</code> | ADR-031 §5.3 supersession | <code>f4628531…fc920f3f9</code> | | <code>docs/adr/ADR-032-h7-kappa-portability-e1.md</code> | portability E₁ pre-registration | <code>4e87aa4e…7567f4e749</code> | | <code>docs/adr/ADR-033-audit-gate-public-runner.md</code> | audit gate, ACCEPTED 2026-05-09 | <code>938aa792…994d3f2f</code> | Now public (no longer behind the private freeze chain): the entire κ pipeline. Unlike v0.4.0, v0.5.0 ships a runner that an auditor can execute end-to-end against the v0.4.0 CSV with a single command and verify the bit-level reproduction (max $|\Delta| \leq 4.5 \times 10^{-11}$ on the worst float column, 0 mismatches on all integer columns) before the portability claim is even considered. The private v0.4.0 runner (<code>.forge_private/h7_dev/src/h7/kappa_runner.py</code>) is no longer required to reproduce the κ statistical claim. <h2>3. Where the value is</h2> This release ships three independent guarantees stacked on top of each other: <ul> <li>Portability of the signature. The v0.4.0 κ effect is not an E₀-specific artefact. Under a calibrated diffusive E₁, on a fresh pool [2000-2029] never observed before the tirage, the same paired Wilcoxon test returns $p_\text{greater} = 9.3 \times 10^{-10}$ with $d = +3.09$ on both reporting branches simultaneously. Thirty positive seeds out of thirty. The mechanism — local temporal differentiation — survives the change of medium with the effect size increasing slightly rather than decaying.</li> <li>A public bit-level audit. The runner that produced <code>h7_kappa_portability.csv</code> is the same runner that, on the v0.4.0 reference pool, reproduces <code>h7_kappa_run_results.csv</code> to $4.5 \times 10^{-11}$ on the worst float column and exactly on all fifteen integer columns. The audit was a precondition for the portability tirage being admissible, not a post-hoc justification. Tolerances were frozen in ADR-033 §4 before iter1 was executed and were not relaxed when iter1 failed.</li> <li>A reproducible failure-then-fix trail. Iter1 failed strictly on the obs-shuffled witness branch. The diagnostic was published (commit <code>2719191</code>), the investigation was bounded by ADR-033 §6.2 with the candidate hypotheses ordered by prior probability, and the fix was a single-bit byteorder change in <code>src/agents/obs_shuffled_agent.py</code> (commit <code>c15f313</code>) — not a tolerance relaxation. The v0.4.0 reference CSV was not modified at any point. This is what an audit gate is for.</li> </ul> For an operational reader interested in detection-of-anomaly use cases, the practical reading is narrow and intentional: the κ test is a portability stress test of a structural-coupling probe, not a detection product. What v0.5.0 demonstrates is that the probe's positive answer on E₀ does not collapse when the environment is replaced by a calibrated diffusive variant; the signal-to-noise on $\Delta_\kappa$ (Cohen $d \approx 3$) survives the swap. Whether that property is useful for a downstream anomaly-detection pipeline is a separate engineering question that is not answered here. <h2>4. What this release is NOT</h2> <ul> <li>Not a claim about cognition, agency, intent, intelligence, emergence, thought, or any cousin term. ADR-020 §3 vocabulary remains binding; permitted vocabulary remains operational only.</li> <li>Not a claim that v0.5.0's $d = +3.09$ is larger than v0.4.0's $d = +2.66$ in any meaningful sense. The two effect sizes were measured under different environments, on different seed pools, and the comparison is not pre-registered as a test.</li> <li>Not a claim that the κ mechanism is the only mechanism that would transfer to E₁, nor that it is the minimal transferring mechanism. Alternative agents are not tested here. ADR-032 only pre-registered M_κ.</li> <li>Not a claim about portability to environments other than the calibrated E₁ specified in <code>research/calibration_e1.json</code>. Generic "robustness" or "out-of-distribution generalisation" beyond the pre-registered E₁ is not claimed.</li> <li>Not a detection product, not a security product, not a compliance product. The runner is a research instrument; using it as a decision-making component requires its own qualification process, which is out of scope of v0.5.0.</li> <li>Not preliminary. Per ADR-032 the portability pool [2000-2029] is now spent. No further runs of the κ pipeline on E₁ under the present probe are authorised on this pool.</li> </ul> <h2>5. Reproducibility note</h2> The H₅ Docker image (<code>cae-research-kit:0.1.0</code>, fingerprint <code>406ce26e…008e7e5f</code>) is unchanged. The development environment is unchanged from v0.3.0 / v0.4.0 (Python 3.12.3, <code>numpy==1.26.4</code>, <code>scipy==1.12.0</code>, <code>pytest==8.1.1</code>). All v0.4.0 private-chain SHAs are copied into <code>research/MANIFEST.v0.5.0.yaml</code> for cross-reference. To reproduce the audit gate (cold): <pre><code>python -m src.experiments.portability_draw \ --pool audit \ --output research/h7_kappa_audit_v04_iter2.csv python -m src.analysis.audit_compare \ --reference research/h7_kappa_run_results.csv \ --candidate research/h7_kappa_audit_v04_iter2.csv \ --report-output research/h7_kappa_audit_v04_iter2.report.txt # exit code 0 on PASS, 1 on FAIL. </code></pre> To reproduce the portability verdict on E₁ (cold; only after audit PASS): <pre><code>python -m src.experiments.portability_draw \ --pool portability --i-have-read-adr-033 \ --output research/h7_kappa_portability.csv python -c "from src.analysis.verdict_v05 import compute_verdict; \ compute_verdict('research/h7_kappa_portability.csv', \ 'research/h7_kappa_portability_verdict.json')" </code></pre> Wall-clock on the H₇ development environment: ≈ 3 min 35 s (audit [1500-1529], E₀) + ≈ 9 min 50 s (portability [2000-2029], E₁) on a single CPU. The portability runner enforces a hard refusal on <code>--pool portability</code> without <code>--i-have-read-adr-033</code>. The 30-row per-seed CSVs (29 columns each: header verified by <code>audit_compare.EXPECTED_HEADER</code>) are sufficient to reproduce both verdict computations independently of the runner. The 73 new ADTs introduced between v0.4.0 and v0.5.0 (<code>test_sigma_chain</code> 21, <code>test_obs_shuffled_agent</code> 15, <code>test_portability_draw</code> 12, <code>test_audit_compare</code> 12, <code>test_verdict_v05</code> 13) all pass against the shipped sources. <h2>6. Next steps</h2> This release closes the portability question raised in v0.4.0 §6 (bullet Cross-environment portability). It does not pre-empt the other two follow-on questions of v0.4.0 §6: <ul> <li>κ-stability sweep. Whether the same effect size holds for variants of M_κ (longer windows, EMA, leaky integrator, multi-tick lookback). Not pre-registered; pool not reserved.</li> <li>H8 deferral confirmation. Whether the H8 architectural pivot remains permanently unnecessary. ADR-028's Option A vindication — now extended to E₁ by v0.5.0 — does not pre-empt H8 on its own merits; a separate ADR is required.</li> </ul> A new question opens with v0.5.0: <ul> <li>Portability beyond E₁. Whether the κ-transfers signature replicates under environments other than the specific calibrated E₁ used here (e.g. recalibrated diffusion coefficients, alternative topologies, perturbed observation maps). Out of scope; would require a new pre-registration, a new pool, and a new audit gate against v0.5.0.</li> </ul> The portability pool tail [2030-2099] remains reserved and frozen pending any of the above ADRs. No further H₇ runs are authorised on it until then.

Podobné jednotky