Qwen3.5-4B Pareto Policy Integration

Finished2026-07-12Agentic Breadth Installation Evidence-Conditioned Selection Posttraining and Adaptation Process Control and Tool Use+2 moreGitHub ↗

The one idea you need

To fold two specialist models into one, each coaches the tasks it owns — a sprint coach and a marathon coach training a single runner. That only pays off if each is genuinely fastest in its own lane. So first you check.

The question

Two versions of the same small model each looked best at either quick or long, involved tasks — can you fuse them into one model good at both?

What we found

No. Retested on fresh, uncontaminated tasks, the version built for quick work actually lost at quick work by about two points, while the version built for long work won at both quick AND long tasks. One version quietly dominated everywhere, so there was no genuine trade to fuse — the whole merge was cancelled before any training ran.

Why it matters

Before spending compute to merge or distill two specialists, prove on your own tasks that each is genuinely the stronger coach on the exact slice it will teach. A specialist label earned on another benchmark can silently invert.

Quick version's margin on quick tasks-2.2%wrong direction — it lost in both test rounds

Long version's margin on long tasks+4.6%the only specialty label that held up

Worst skill the quick version broke-22%on a task type it was never trained on

Skills the long version quietly lost6 of 32retention checks it slipped past, worst about -9%

On this page

Results at a glance
Overview
Report
Experiment log
Data files
Reproduce
Related

Results at a glance 2

Each specialist versus its rival on its own tasks, by test round

How to read

The horizontal axis lists four test rounds: two quick-task rounds, then two long-task rounds. The vertical axis is how much the version assigned to that task type beat the other version; zero is a tie, and above zero is better.

Takeaway → Both quick-task bars dip below zero while both long-task bars rise above it: the quick version lost its own specialty in both rounds, the long version won its own in both.

Data table

frozen capability block	intended teacher minus alternate
quick block 0	-0.00693
quick block 1	-0.0379
deep block 0	0.0425
deep block 1	0.0487

Numbers from experiments/qwen35_4b_pareto_policy_integration/analysis/specialist_qualification.json

Technical framing

Only the deep specialist won both qualification blocks — Positive means the policy assigned to that stratum beat the alternate checkpoint. Both quick blocks were negative; both deep blocks were positive. No effect-size floor was used.

Pooled score and its cautious floor for each specialist

How to read

The horizontal axis shows the two roles, quick version and long version. For each, one bar is its overall margin over the rival and a second is a cautious worst-case estimate of that margin; both bars above zero means a solid pass.

Takeaway → The long version's two bars both sit above zero; the quick version's both fall below, so even its best-case estimate is negative and only one specialist qualified.

Data table

intended specialist role	paired macro delta	one-sided 95% lower bound
quick blend	-0.0224	-0.049
deep apex	0.0456	0.034

Numbers from experiments/qwen35_4b_pareto_policy_integration/analysis/specialist_qualification.json

Technical framing

Pooled capability evidence passes only for the deep policy — The deep policy clears the capability criterion but fails six retention cells; the quick policy fails the capability sign in both blocks. The gate stops before teacher audit or MOPD.

In the author’s words from the Report · “Interpretation”

The later C54 model-soup sweep strengthens the parameter-interpolation negative, but it does not repair this missing teacher crossover. This is not evidence for or against MOPD, OPSD, or the corrected top-k loss. Those mechanisms require a better teacher at the student's actual state. Here the prerequisite failed before teacher scoring or any integration update. It would be scientifically invalid to run MOPD and call the outcome an integration test when one coarse route is already worse on both replicated capability blocks. The useful next hypothesis is narrower than “try the same teachers harder.” The cell table contains real heterogeneity: each checkpoint wins some local states, but the assumed quick/deep labels are too coarse. … Read the full result →

Overview

Status: stopped negative on 2026-07-12 before teacher audit. This clean successor removed the prior experiment's arbitrary +0.10 teacher hurdle and gave any replicated, statistically credible positive paired gain a path forward. The regenerated C54 policies did not form the required complementary pair on the clean procedural proxy: blend lost its intended quick comparison in both blocks, while apex won deep capability but missed retention.

Research Programs

Primary: agentic_breadth_installation.
Supporting: posttraining_and_adaptation, test_time_reasoning_budget, benchmark_generalization.
Closest near-duplicate: qwen35_4b_specialist_policy_integration, preserved as the design-feasibility negative that motivated this correction.
New anchor: C54 in qwen35_4b_gauntlet_frontier, which measured a non-convex quick/deep Pareto pair from the same pinned 4B base.

Question and Hypothesis

Can one Qwen3.5-4B policy consolidate two same-origin policies that separately win on short/quick and interactive/deep work, or is their tradeoff a genuine shared-parameter capacity frontier?

The student starts from the quick policy and generates its own continuations. At the exact visible student prefix, the quick teacher supplies retention pressure on short atoms and the deep teacher supplies capability pressure on long atoms and interactive states. Corrected top-50 MOPD should preserve dense policy choices that data union and parameter interpolation lose.

Corrected Gates

No fixed absolute specialist delta.
A complementary teacher advantage is delta > 0, both independent blocks positive, with a one-sided 95% stratified-bootstrap lower bound above zero.
Saturated cells such as ferrier are retention anchors. Equality is fine; regression is not. They never veto another teacher or integration.
The final one-checkpoint system—not each teacher—must beat matched-compute sampling.
Integration seed 42 (the first frozen seed) is the deployable primary; seeds 43 and 44 are directional replications, never a checkpoint-selection pool.

Exact frozen rules are in reports/preregistration.md, with adversarial review in reports/design_review.md.

Substrate

The copied C54 procedural gym contains 12 training families and two never-trained transfer families (brinework, spindle). Quick evaluation is atoms L1-L2; deep evaluation is atoms L3-L6 plus episodes L2/L3/L5. All generation seeds are disjoint across calibration, qualification, rollouts, and confirmation.

The two specialist datasets are committed provenance-clean C54 artifacts:

quick: qwen35_4b_gauntlet_frontier/data/sft_blend.jsonl;
deep: qwen35_4b_gauntlet_frontier/data/sft_apex.jsonl.

Both policies are regenerated independently from the identical pinned base and explicitly merged before evaluation.

Firewall

The only model is Qwen/Qwen3.5-4B at revision 851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a.
Nothing under benchmarks/ is read, imported, or used for training.
Programmatic state and scores never appear in prompts.
Comparable evaluation arms use the same pinned vLLM backend and metadata.
Runtime LoRA is forbidden; every evaluated adapter is explicitly merged and behavior-gated.

Run

CPU scientific smoke:

python3 experiments/qwen35_4b_pareto_policy_integration/scripts/run.py --smoke

Reached model stages will be resumable:

python3 experiments/qwen35_4b_pareto_policy_integration/scripts/run.py --stage model-smoke
python3 experiments/qwen35_4b_pareto_policy_integration/scripts/run.py --stage specialists
python3 experiments/qwen35_4b_pareto_policy_integration/scripts/run.py --stage specialist-canary
python3 experiments/qwen35_4b_pareto_policy_integration/scripts/run.py --stage calibrate
python3 experiments/qwen35_4b_pareto_policy_integration/scripts/run.py --stage qualify
python3 experiments/qwen35_4b_pareto_policy_integration/scripts/run.py --stage teacher-audit
python3 experiments/qwen35_4b_pareto_policy_integration/scripts/run.py --stage locality
python3 experiments/qwen35_4b_pareto_policy_integration/scripts/run.py --stage integrate --seed 42
python3 experiments/qwen35_4b_pareto_policy_integration/scripts/run.py --stage controls
python3 experiments/qwen35_4b_pareto_policy_integration/scripts/run.py --stage confirm

The benchmark stage is intentionally unavailable until every procedural gate passes.

The reached qualify command now exits nonzero by design after writing the terminal receipt. All later commands fail closed on that receipt and must not be run in this experiment.

Current Evidence

Prior result preserved: the old +0.10 rule was impossible at a 0.994 tools baseline and did not test MOPD. This successor actually tested the corrected delta > 0 prerequisite.
Both specialists were independently regenerated, explicitly merged, and behavior-gated. Calibration and all four qualification arms passed exact model, engine, seed, scope, and pairing checks.
On 768 pooled quick capability pairs, blend - apex = -0.02241; both block means were negative (-0.00693, -0.03789) and the one-sided 95% lower bound was -0.04897. The failure is the sign of the effect, not an arbitrary minimum magnitude.
On 4,032 pooled deep capability pairs, apex - blend = +0.04563; both block means were positive (+0.04254, +0.04871) and the lower bound was +0.03401. However, six deep retention cells regressed by more than the frozen 0.02 allowance.
Therefore the C54 quick/medium Pareto labeling did not transport into a clean quick/deep teacher crossover. This result says nothing about MOPD's efficacy: no teacher audit, locality pilot, MOPD update, control, confirmation, or benchmark invocation ran.

Artifacts

configs/default.yaml: frozen splits, seeds, statistics, and controls.
idea_intake.md: relation to the stopped predecessor and C54.
reports/preregistration.md: decision rules.
reports/design_review.md: adversarial pre-run review.
reports/literature_review.md: primary-paper map behind the social-post acronyms and the experiment's collapse safeguards.
analysis/specialist_qualification.json: terminal paired gate receipt.
runs/policy_eval/*qualification*: all four raw qualification arms and provenance metadata.
src/gym/: contamination-safe procedural substrate.
src/mopd_loss.py: corrected teacher-top-k reverse-KL objective.
reports/artifact_manifest.yaml: external checkpoint policy.

Report

Rendered from reports/report.md

Status

Stopped negative on 2026-07-12 before teacher audit. The corrected specialist gate had no fixed effect-size floor: any replicated, statistically credible paired gain above zero could qualify. The regenerated blend policy nevertheless lost its intended quick comparison in both independent blocks. The apex policy won deep capability but missed the frozen retention rule. There was no clean complementary teacher pair, so MOPD was not run.

Research Program Fit

This is the clean successor to qwen35_4b_specialist_policy_integration, whose fixed S0 + 0.10 rule became impossible at a saturated baseline. It directly tests the prerequisite behind the MOPD path suggested by C54: whether C54's same-origin quick-optimal blend and medium/deep-optimal apex checkpoints actually cross over on a contamination-safe state distribution.

Method

Both rank-32 policies were regenerated independently from the identical pinned Qwen/Qwen3.5-4B revision, explicitly merged, weight-hashed, and behavior-gated through the same vLLM 0.24 backend. Calibration assigned saturated, explicit anchor, and never-trained transfer cells to retention; all remaining cells were fixed as capability cells before qualification.

Qualification used two disjoint procedural seed blocks. Quick capability was atoms L1-L2; deep capability was atoms L3-L6 plus interactive episodes at L2/L3/L5. For each intended teacher, qualification required:

pooled paired macro delta greater than zero;
a one-sided 95% family/level-stratified bootstrap lower bound above zero;
both block means above zero; and
no retention-cell regression greater than 0.02.

There was no minimum positive effect size. Every arm used greedy decoding, identical paired items, the exact frozen engine geometry, and the same local composite provenance checks.

Reached Evidence

The installation canary passed before any task score was accepted: both specialists differed from base on 8/8 fixed prompts and from one another on 7/8. Calibration then completed 1,488 items per policy and assigned 56 cells to capability inference and 52 to retention. All ferrier, brinework, and spindle cells remained retention-only as frozen.

Qualification completed all four 4,416-item arms—17,664 arm-item evaluations, or 8,832 paired item identities—with no exclusions. It consumed 18,049,063 sampled tokens and 12,713.08 evaluation seconds.

Block	Policy	Broad quick	Broad deep	Sampled tokens	Wall seconds
96200	`blend`	0.7876	0.5297	4,748,497	3,387.61
96200	`apex`	0.8089	0.5813	4,244,434	2,970.32
96300	`blend`	0.7892	0.5191	4,737,106	3,341.25
96300	`apex`	0.8268	0.5819	4,319,026	3,013.91

All nine protocol checks passed: both seed lists, qualification scope, greedy decode, family identity, calibration authorization and partition, and presence of capability cells in both strata.

Terminal Qualification Result

Intended teacher	Capability pairs	Block deltas	Pooled delta	One-sided 95% LCB	Retention	Decision
quick `blend`	768	−0.00693, −0.03789	−0.02241	−0.04897	12/20 cells pass	fail
deep `apex`	4,032	+0.04254, +0.04871	+0.04563	+0.03401	26/32 cells pass	fail

The quick result is decisive for this design: blend did not merely miss a large or practical-gain threshold; its effect had the wrong sign in both blocks under a rule that would have accepted any credible positive delta. Broad raw scores agree—apex was higher on both quick and deep strata in both blocks—although the preregistered capability-cell macro, not the broad mean, made the decision.

The deep result is separately informative. apex has a stable intended-stratum capability advantage, but it is not a clean dominance result. Six retention cells exceeded the 0.02 regression allowance: brinework atom L6; glyphgate episodes L2/L3/L5; and kilnrite episodes L2/L5. Its worst retention delta was −0.09375. Conversely, blend failed eight quick retention cells, with its worst delta −0.21875 on never-trained spindle L2.

The machine receipt therefore records gate.passed=false and downstream_authorization=stop_before_teacher_audit.

Interpretation

C54's quick/medium Pareto result is real evidence on its menagerie instrument, but its labels do not identify a transportable quick/deep teacher routing rule on this clean procedural proxy. A checkpoint can cross an external aggregate target while failing to supply a locally better teacher on the state distribution where distillation must occur. The later C54 model-soup sweep strengthens the parameter-interpolation negative, but it does not repair this missing teacher crossover.

This is not evidence for or against MOPD, OPSD, or the corrected top-k loss. Those mechanisms require a better teacher at the student's actual state. Here the prerequisite failed before teacher scoring or any integration update. It would be scientifically invalid to run MOPD and call the outcome an integration test when one coarse route is already worse on both replicated capability blocks.

The useful next hypothesis is narrower than “try the same teachers harder.” The cell table contains real heterogeneity: each checkpoint wins some local states, but the assumed quick/deep labels are too coarse. A new experiment may pre-register outcome- or state-routed same-prefix distillation: score both same-origin teachers on disjoint calibration prefixes, route only where a teacher has positive continuation advantage, and confirm the routing rule on fresh cells before training. That is the principled version of replacing a hinted log-probability slogan with an advantage estimator. It must remain a new experiment with fresh splits and must still beat both checkpoints, a visible two-checkpoint router, and matched-compute sampling.

For immediate deployment, C54's visible tier router remains the supported upper reference. It is not a one-checkpoint capability-installation result.

Unreached by Design

same-prefix correct-versus-wrong teacher audit;
exact-logit locality pilot;
MOPD, wrong-route, and off-policy updates;
parameter-merge and compute-overmatched union-SFT controls;
confirmatory evaluation and matched best-of-8;
every benchmark invocation.

No benchmark content, item, transcript, or result detail was read.

Artifact Manifest

The terminal receipt is analysis/specialist_qualification.json; all four raw qualification arms are under runs/policy_eval/. Large specialist adapters and explicit merged composites remain external as documented in artifact_manifest.yaml.

Experiment log 14

Show the running log (14 entries, 2026-07-12)

Scaffold

Created as a new experiment scaffold.

2026-07-12 — corrected successor accepted

User rejected the predecessor's fixed S0 + 0.10 specialist gate as an obvious scientific-design error. The correction is not to lower that number; it is to remove arbitrary effect-size qualification entirely.
Teacher existence is now paired delta > 0 with two positive seed blocks and a one-sided stratified-bootstrap lower bound above zero. Saturated cells are retention anchors, not vetoes.
C54 landed between the two experiments and materially changed the best test: rather than speculate about four not-yet-trained domain specialists, this run attempts to consolidate the already evidenced same-origin quick/deep Pareto policies (blend, apex).
New experiment directory created rather than rewriting the predecessor.
No task-model output existed when the config, preregistration, and design review were authored.

2026-07-12 — design lock

Pre-output design commit: 6bb8530ac5b1c289fbf9682846317607d46e9673.
runs/preregistration_receipt.json freezes SHA-256 digests for the config, intake, preregistration, and design review. Every non-smoke stage verifies both those digests and commit ancestry before loading a model.

2026-07-12 — live model preflight

Pinned base vLLM semantic/runtime smoke passed 4/4 tasks; Transformers prompt parity, finite logits, causal-conv and flash-linear-attention fast paths all passed under the frozen training lock.
The scaffold runner had accidentally omitted its local-composite CLI path. The first reload therefore stopped at argument parsing after training/merge, before producing a local-model score. Reintroduced the proven explicit model_override path with mutual-exclusion and model-type validation plus a regression test.
Weighted-training smoke completed 8/8 steps on the quick data shape with no skips. Explicit merge applied 128/128 nonzero deltas (summed Frobenius norm 23.00) on CUDA FP32 with TF32 disabled.
The merged composite then loaded through vLLM, produced the 4/4 semantic smoke outputs, and preserved the requested full CUDA-graph decode geometry.

2026-07-12 — integration harness locked before policy evaluation

Corrected teacher-top-50 MOPD now caches full-softmax probabilities at the exact student token prefix and consumes 160 distinct trajectories per round; no rollout or target span is split or replayed to manufacture update count.
The five-update locality pilot measures centered non-target logit drift and full-vocabulary entropy change before authorizing the four-round run.
Wrong-route and off-policy controls use the primary arm's deterministic rollout selection and rescale backward loss to the primary arm's measured initial corrected-top-k pressure in each round. Update count, data pressure, and initial objective magnitude are therefore matched explicitly.
Non-finite loss or gradients and the frozen round-loss ceiling now preserve an auditable stopped adapter receipt instead of disappearing as a crashed process.
The two-block final analyzer uses equal quick/deep macro weight, paired one-sided bounds against both source policies and every one-checkpoint control, separate anchor/transfer retention checks, three training seeds, and the execution-filtered best-of-8 hurdle.

2026-07-12 — quick specialist regenerated

The committed C54 blend corpus encoded 2,233/2,240 rows; the seven omitted rows exactly match the pre-run encoding audit.
QLoRA completed 350/350 optimizer steps (2.5 epochs), final logged loss 0.8077 and whole-run train loss 2.0220 on the NVIDIA L40.
Explicit composite installation applied 128/128 nonzero LoRA deltas with summed Frobenius norm 164.55 (CUDA FP32, TF32 disabled). The merged model weight hash is 97bb30362c130fab6525586b39dff9d72ca31be57b72edc879bf03e304ce91cb.
No task score was inspected before the independent deep specialist began.
Before either full specialist was behavior-scored, an eight-prompt same-prefix installation canary was fixed (four quick, four deep; maximum rendered prompt length 438 tokens). Calibration now requires both merged specialists to differ from base and from each other under the same greedy vLLM runner; the canary has no task-success threshold.

2026-07-12 — pre-evaluation protocol hardening and upstream closure

Before any calibration or qualification score existed, audit found that the imported procedural harness did not forward the config's explicit CUDA-graph capture list even though CLI generation did. Commit ce45c383 forwards the frozen geometry and adds a regression test.
Commit 437a1b87 additionally makes every procedural score fail closed unless all generation calls prove the exact local composite, config hash, engine dimensions, and resolved full-decode graph geometry. The scientific suite is 31/31 and repository CI is green.
C54 concurrently closed its simple adapter-capacity alternative after this experiment's design lock: rank-128/alpha-256 APEX scored quick +0.249 and medium +0.229, below rank-32 APEX (+0.308/+0.345). Extra LoRA capacity therefore did not dissolve the frontier. This strengthens the motivation for the already-frozen policy-space test but does not change an arm, seed, gate, or hypothesis here.

2026-07-12 — deep specialist regenerated

The committed C54 APEX corpus encoded 4,662/4,669 rows; the same seven pre-audited non-training rows were omitted.
QLoRA completed the frozen 730/730 optimizer steps (2.5 epochs) with whole-run train loss 0.882575, 7,156.98 seconds wall time, and 16.13 GB peak allocated CUDA memory on the NVIDIA L40.
Explicit composite installation applied 128/128 nonzero LoRA deltas in CUDA FP32 with TF32 disabled. Their summed Frobenius norm is 248.5323 (maximum 4.1774), and the merged model weight hash is 3bf936150a0a68e80a7a2ef3564334503fdfc3dfb3dcc5f6dfa0b1b12b7cf28b.
Both specialist receipts now bind the pinned base revision, exact external adapter/composite paths, and model-weight hashes. No specialist task score was inspected during either training run.

2026-07-12 — specialist installation canary and evaluation-engine preflight pass

On eight fixed same-prefix greedy prompts, quick and deep each changed all 8/8 base continuations; the two specialists differed from one another on 7/8. All prompt, runner, decode, merge-receipt, and nonzero-delta checks passed, authorizing independent calibration. This is an installation test, not a task-success measurement.
A separate one-token quick-composite preflight loaded the exact frozen evaluation geometry: 16,384-token model/batch limits, 48 sequences, 0.85 GPU utilization, and explicit capture sizes [1,2,4,8,16,24,32,40,48]. vLLM resolved full decode graphs at all nine requested sizes without a Mamba clamp or process re-exec, reported 807,029 KV-cache tokens and 49.26x maximum full-length concurrency, and bound the output to the quick merge receipt.

2026-07-12 — first calibration attempt exposed post-generation ledger bug

The quick policy completed all 1,344 calibration atoms and 144 interactive episodes through turn 17, then failed before writing a score because the token ledger expected turn["policy"]["n_sampled_tokens"]. The copied harness actually stores the slim policy fields directly on each turn.
This was a bookkeeping-only failure after generation: no score, cell role, gate, or analysis artifact existed, and the output directory remained empty. The fix reads turn["n_sampled_tokens"] through a schema-specific helper; a regression test exercises both atom and episode token layouts. The suite is now 32/32. The identical frozen calibration seed and protocol will be rerun.

2026-07-12 — independent descriptive calibration passes

The corrected rerun completed 1,488 paired items per specialist. Quick used 1,642,252 sampled tokens in 1,296.25 seconds; deep used 1,433,797 in 1,163.68 seconds. Every exact-model, engine, runner, and resolved-graph check passed.
Calibration assigned 56 cells to capability inference (8 quick, 48 deep) and 52 to retention. Every ferrier, brinework, and spindle cell is retention-only as frozen; both strata retain capability cells.
Descriptively, the quick policy scored 0.7939 on the broad quick stratum and 0.5256 on deep, while deep scored 0.8224 and 0.5662. On the cell-balanced capability subset, intended-teacher deltas were -0.0605 (quick) and +0.0332 (deep). These calibration outcomes do not qualify or disqualify a teacher and were not used to alter cells, seeds, or thresholds; the two frozen qualification blocks remain the decision instrument.

2026-07-12 — qualification execution in progress

Quick policy block 0 (seed 96200) completed all 4,416 items in 3,387.61 seconds using 4,748,497 sampled tokens. Broad raw means were 0.7876 quick and 0.5297 deep; all evaluator provenance checks passed.
This single-arm output is checkpointed without an advantage interpretation. The paired deep arm and second independent block remain frozen and required.
Deep policy block 0 completed the identical 4,416 items in 2,970.32 seconds using 4,244,434 sampled tokens. Broad raw means were 0.8089 quick and 0.5813 deep; every provenance check passed. The run proceeded to seed 96300 without an interim paired decision or any protocol change.
Quick policy block 1 (seed 96300) completed in 3,341.25 seconds using 4,737,106 sampled tokens. Broad raw means were 0.7892 quick and 0.5191 deep, with every provenance check passing. Deep block 1 remains required before any qualification statistic is computed.

2026-07-12 — upstream C54 model-soup context arrived during qualification

After this experiment's design lock and while the frozen qualification run was active, C54 added a five-point weight-space sweep between blend and apex. Its best reported joint point (soup40) reached paired menagerie deltas of +0.331 quick and +0.304 medium; the source experiment reports disjoint mixture ranges for clearing its chosen +0.32 target on both tiers.
This is relevant negative evidence for parameter interpolation, one of this experiment's already-frozen controls. It neither tests same-prefix MOPD nor establishes that the two regenerated policies cross over on this clean procedural proxy. No arm, seed, cell role, gate, or interpretation rule was changed in response.

2026-07-12 — terminal specialist-qualification stop

Deep policy block 1 completed all 4,416 items in 3,013.91 seconds using 4,319,026 sampled tokens. Broad raw means were 0.8268 quick and 0.5819 deep, and every exact-model, runner, engine, seed, and scope check passed.
Across all four frozen qualification arms, 17,664 arm-item evaluations (8,832 paired item identities) used 18,049,063 sampled tokens and 12,713.08 seconds. No item, family, seed, or arm was filtered after observation.
The quick policy's intended-stratum capability delta was negative in both independent blocks (-0.00693, -0.03789) and pooled to -0.02241 with a one-sided 95% stratified-bootstrap lower bound of -0.04897. It therefore fails the corrected delta > 0 existence gate without any effect-size floor.
The deep policy did show replicated capability: block deltas +0.04254 and +0.04871, pooled +0.04563, one-sided lower bound +0.03401. However, it exceeded the frozen 0.02 regression allowance on six retention cells, so it is not a clean dominating teacher either.
The pair is not complementary on this contamination-safe procedural proxy. analysis/specialist_qualification.json authorizes stop_before_teacher_audit; no teacher audit, locality pilot, MOPD update, integration control, confirmation, or benchmark invocation ran. Per the preregistration, this is a teacher-transport/qualification negative, not an MOPD result.

Data files 1

Result tables and metrics copied from the experiment folder — preview inline or open the raw file.

runs/smoke/summary.json2.4 kB

Reproduce

Smoke test

python3 experiments/qwen35_4b_pareto_policy_integration/scripts/run.py --smoke

Full run

see the staged Run section in README.md

Run steps are documented inside the experiment folder (README and scripts).

Browse the experiment folder on GitHub ↗