Reliability and Safety

Improve precision, abstention, robustness, reproducibility, and safe artifact handling.

What we have learned

Seed Experiments

Current Read

Reliability standards must be enforced by process and validation, not just written as advice.

Scorecard

Program: charter
Current read: reliability depends on precision, abstention, artifact hygiene, and explicit hidden-label boundaries.
Best next experiment: create a reliability scorecard applied to selectors, verifiers, and tool controllers.
Strong anchors: qwen35_4b_reliability_exec_opsd_audit, qwen35_4b_real_sample_verify_commit, qwen_readable_candidate_verifier.
Avoid repeating: impressive accuracy reports without commit-rate, abstention, artifact, and leakage checks.
Evidence that advances the program: reliability metrics expose regressions that ordinary accuracy hides.

Charter

Show charter.md

Purpose

Improve precision, abstention, robustness, reproducibility, artifact hygiene, and operational safety for small-model experimentation.

Why This Is A Program

As the repository grows, failure to handle artifacts, false confidence, hidden-label leakage, and unverifiable claims becomes a research risk. Reliability is not housekeeping; it determines whether findings can compound.

Progress Signals

Validation catches unsafe artifacts and weak metadata.
Reports clearly distinguish oracle-only and deployable evidence.
Selectors can abstain rather than commit uncertain outputs.
Experiments are reproducible without private weights or accidental adapter check-ins.

Boundaries

This line overlaps with selection and infrastructure but owns reliability standards and failure-prevention patterns.

Backlog

Show backlog.md

Next Experiments And Infrastructure

Add validation gates for oracle/deployable labeling in new reports.
Create artifact manifests for every trained model or adapter that is not checked into git.
Build abstention benchmarks for visible-only selectors.
Add reproducibility scorecards to high-impact experiments.
Audit generated catalogs for stale, missing, or misleading references.

Required Controls

Explicit hidden-label boundary.
Artifact manifest when outputs are external.
False-positive and false-negative slices.
Re-run or deterministic seed notes for major claims.

Stop Conditions

Do not accept a new result as shared knowledge if it cannot be audited for hidden-label use, artifact location, and baseline comparison.

Experiments 85

2026-07-27 → 28 Self-Written Verifier Fidelity
Fill this after the run. Separate deployable evidence from oracle/hidden evaluation.
2026-07-19 Qwen35 4B Agentic RLVR Feasibility
Yes to the machinery: with three fixes (force eager mode so the hybrid architecture doesn't hang, load the model in bf16 not fp32, and split GPU memory 55% to the inference engine with sleep-mode) the loop CLOSES - a mov
2026-07-18 Qwen35 4B Self-Repair Install
The first bet that moved the needle at all — gently. We taught the 4B to debug by training on 504 examples of [buggy code + the real test-failure message] -> [diagnosis + fix], all self-generated by injecting bugs into c
2026-07-16 Repair-Verifier Signal Probe
The gate said no, cleanly. Handed two candidate repairs and the full failure evidence — a task solvable by mentally running each candidate through both trials — the model picked the working one 51.5 percent of the time w
2026-07-15 Retention-Screen Calibration Study
The gap wobbles with a standard deviation of 4.3 tasks — the five-task pass/fail margin was barely one wobble wide, so single-quiz forgetting verdicts were close to coin flips. Every historical 'this model forgot 5-10 ta
2026-07-15 Rank-Capacity Vehicle Cell
The trial could not answer the capacity question — by design. Its built-in guard required the known nine-point forgetting case to reproduce before trusting any comparison, and on this fresh screen that case measured only
2026-07-15 Interleaved-Replay Dose with Medium Pilot
The review round did not prevent forgetting: the dosed model lost nine to ten retained answers against both comparisons — almost exactly the cost of dosing directly — so the theory drawn from comparing old receipts is re
2026-07-15 Dose-Diversity Mechanism Cell
Variety does not protect memory: the larger varied practice set also cost nine retained answers, the known ten-point case reproduced exactly, and even pure review cost five on this fresh screen — so the one past case wit
2026-07-15 Axis Stack Re-adjudication with Medium Pilot
Judged fairly on brand-new tasks — with the ceiling quirk handled — the stacked model won the overall skill test for the third straight time (22 vs 15 and 18 of 40) with the cleanest finishing behavior again. But it won
2026-07-15 Axis Corpus V2 with Staged Repair
Even with lessons rebuilt from the autopsy — walking through the search step by step instead of asserting the answer — the repair skills did not stick (a tie and a loss against the comparisons), so the preregistered kill
2026-07-14 Policy-Supported Successful-Sibling Universal Curriculum
This version cannot run its second step. Across 624 first attempts it found 227 failures, but counting and routing had none and selection had only two; the plan required four failures in every skill. It therefore stopped
2026-07-14 Natural-Language State-Table Universal Curriculum
Not known yet. CPU construction produced 80 truth-checked lessons and two 320-row streams with exactly 286,814 tokens each. All 48 smoke tests pass, but no model has been trained or evaluated.
2026-07-14 Search-Scaffold Universal Curriculum
No. On 26 fresh paired cases, the staged-search model tied replay at 16 correct and trailed its parent at 18. All three parsed 23 answers and hit the limit three times; the candidate solved none of four execute/induct ca
2026-07-14 Failure-Selected Counterfactual Restart Curriculum
The data-selection stage passed: from 624 fresh attempts, the frozen rule selected 52 clean restarts, exactly four for each of 13 skills. Forty correct accuracy or cap failures and twelve correct-but-too-long cases are n
2026-07-13 Validation-policy counterexample curriculum
The test could not answer the question. Before the newly trained model was allowed to run, the starting model and the equal-size old-lesson control each fixed all 48 practice cases. The required gains were therefore impo
2026-07-13 Low-Density Token-Matched Universal Curriculum
No. Neither 40 nor 80 abstract lessons passed the fresh local screen. The stronger 80-lesson stream matched the inherited anchor at 54% accuracy and parsed 62% of answers, but still hit the answer cap 10 times; broad eva
2026-07-13 Qwen3.5-4B: Installing Universal Features via Designed Synthetic Curricula
Not with either tested training geometry. Continuing from the strong policy raised fresh synthetic accuracy from 50% to 69% and beat the untrained base overall, but it scored 31% aggregate versus 45% for the strong start
2026-07-13 State-Formation Capacity Adjudication
Not run yet. The reviewed test starts with three compact-update training runs. If any required state check misses, six matched controls become mandatory; the final three runs open only if the full-size version also misse
2026-07-13 Semantic-policy headroom tournament
No rule produced a stable training target after a failed test: every one of 72 cases reached a correct patch. But before seeing test output, the model got zero of 54 ambiguous first patches fully right and never opened t
2026-07-13 Qwen3.5-4B: The Compute-Optimal Confidence Policy (select + abstain + escalate)
Partly. Picking the answer with the highest single-token P(True) confidence is the best grader-free selector at every compute level, and abstaining below a confidence threshold trades coverage for accuracy cleanly — both
2026-07-12 Transaction-invariant recovery curriculum
Partly. Focused training lifted practice repairs from 52% to 82% and preserved the test-and-revise loop. But on different transaction APIs it reached 72% versus 70% before training, far below the required ten-point gain.
2026-07-12 Verifier-conditioned recovery banking curriculum
Yes — replaying real failures taught recovery, lifting success from about half of cases to more than nine in ten. But the best version also added a little "explain your fix" text, and that supposedly tiny slice actually
2026-07-12 Qwen3.5-4B Same-Prefix Advantage Routing
Only when it replicates, and here one helper didn't. The test actually ran: the deep specialist's local edge held up on a fresh, separate batch, but the quick specialist looked strong on one batch of stuck points and the
2026-07-12 Repository search-compress-bank coding curriculum
No. It became flawless on the six codebase types it trained on, solving every task in exactly four clean steps. But on four never-seen codebase types its success collapsed from 68% to 35% — worse than just running the un
2026-07-12 Public-verifier recovery branch tournament
No. On four new problem types, both policies repaired 74%, and their combined best reached only 75%—exactly tied with two action-only attempts. The selector was correctly stopped before scoring.
2026-07-12 Qwen3.5-4B Pareto Policy Integration
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 d
2026-07-11 Qwen3.5-4B Specialist Policy Integration
No — we never found out. Each of four required skills had to gain ten points before merging, but scores top out at 100. One tool-use skill already sat at 99.4%, so its target of 109.4% was impossible. The experiment stop
2026-07-11 Interactive policy curriculum: oracle DAgger to execution-reward RL
No — it backfired badly. Success on the practice tasks fell from about 60% to 35%, and on tasks never touched in training from about 69% to 35%. It kept editing fluently but stopped running its checks: the verify-and-com
2026-07-10 Qwen3.5-4B Verified Macro Invention Long-Context Rerun
No. The model was not short on thinking space, it was stuck repeating itself, and more space only bought longer loops. Handed the blueprint, all 16 tasks finished cleanly. Asked to invent the program from examples alone,
2026-07-10 Qwen3.5-4B verified-macro exact CUDA-graph vLLM rerun
No, it just loops harder. At both budgets, all 48 reasoning attempts ran out of thinking room and had to be force-stopped, none finishing on its own. Widening the budget from 49,152 to 61,440 tokens pushed the share stuc
2026-07-10 Qwen3.5-4B verified-macro capacity-fit vLLM rerun
No. Two hidden limits bit first. Configuring the server for 64 simultaneous long prompts demanded about 2.4 million tokens of fast memory from a pool holding roughly 1 million, forcing constant re-reading. Cutting to 19
2026-07-10 Qwen3.5-4B Answer-Potential Trace SFT
No. The score carried genuine signal — it tracked what the reasoning actually said, not its length or wording — but its top pick raised fresh-answer success only from about 13% to 20%, short of the useful margin needed t
2026-07-09 → 10 Gauntlet round 1: breadth-first agentic expert iteration
Mostly the second. On a blind benchmark the model scored about 14 percent, with six task types near zero, but it had usually reasoned correctly. It simply hit its thinking limit, restarted explaining instead of writing t
2026-07-09 Qwen3.5-4B Verified Macro Invention
No. Even handed the finished plan and asked only to re-express it — no problem to actually solve — the model got it exactly right just one time in four, short of the three-in-four bar. Every output looked flawless: valid
2026-07-08 → 09 Does the installable hypothesize-and-verify skill move the structure wall?
No. Fine-tuning the model on 1,476 worked guess-and-check traces doubled its success at the two-step depth it practiced on (lists jumped from 37% to 70%), yet did nothing one step deeper: three-step success stayed near 5
2026-07-08 Qwen3.5-4B: Does Code Confidence Replicate on HumanEval?
Ask it directly. Averaging the model's confidence across every code token barely helps, nudging correct picks from 77% (blind guessing) to just 79%. But handing the model its own code and reading its confidence in one ye
2026-07-07 → 08 Qwen3.5-4B: Can Confidence Replace the Verifier in the Banking Flywheel?
No. When training on answers checked by actually running the code lifted single-shot accuracy from 8% to 24%, confidence-filtered data — fifteen times purer than a random grab of the model's own outputs — landed right on
2026-07-07 Qwen3.5-4B: Does the Confidence Toolkit Survive on Real Code?
Yes, but not the obvious way. Averaging the model's certainty across every token of a program barely beats a plain majority vote among the tries. The real winner: make the model write the code, then answer one yes/no que
2026-07-03 Qwen3.5-4B Coverage Banking: does banking shift the proposal distribution?
Yes, but it depends on difficulty. On easy one-step problems it just pulls answers it already knew into its top guess (60% to 80%), gaining no new ground. On harder two-step problems something new happens: on brand-new t
2026-06-30 Qwen3.5-4B Verifier vs Visible Selector Showdown
No. When you can run even a single example test on each candidate, that filter alone lifts the share of shipped programs that fully work from 77% to 85%. Adding a free, instant self-confidence rating reaches 87% — matchi
2026-06-30 Qwen3.5-4B Thinking-Budget Scaling
Yes, but with a sweet spot. Turning on the model's built-in reasoning lifted the accuracy of the answer you would actually ship from 76% to 91% on basic Python tasks — a genuine capability gain, not just a wider pool of
2026-06-29 → 30 Qwen3.5-4B Thinking Separability Probe
Yes, but not for the reason you would expect. From one snapshot of internal activity, whether the model's own code is correct is readable well above a coin-flip, about 64 to 76 percent of the time. Thinking first sharpen
2026-06-29 → 30 Qwen3.5-4B Generator-Verifier Gap
Only after it thinks. Judging on sight, the model rubber-stamps 91% of its tries as correct while just 77% truly pass — barely a check, mostly agreeing with itself. Given room to reason first, it becomes a real critic: l
2026-06-28 Qwen Oracle-Distilled Acquisition Policy
No. The trained picker does read real signal: it beats revealing nothing (50 to 57 percent of tasks fully solved) and crushes a version fed scrambled answers (27 percent). But it lost to simply grabbing a varied spread o
2026-06-28 Counterexample-Guided Ephemeral Program
No. Answering each row directly solved three-quarters of tasks completely, while the best rule-program the method could pick solved only about four in ten — and a perfect picker that peeks at the answers did no better. T
imported 2026-07-12 Qwen Register Trace Refiner
Rarely. Even a flawless picker that always grabbed the correct edit reached only 37% on plainly worded problems and 7% on reworded ones, because the correct program usually isn't among the roughly 1,300 nearby edits at a
imported 2026-07-12 Qwen Progressive Repair Compiler
Partly. The correct fix sits in the candidate list almost nine times in ten, yet the judge finds it only about half the time, lifting exactly-correct programs from 30% to 49% against the 88% a flawless chooser would reac
imported 2026-07-12 Qwen Learned Repair Verifier
Partly. Without ever seeing the true answer, the judge lifted correct execution from 30% to 47%. But a checker allowed to peek at the answer key found a correct fix already sitting in the candidate pile 88% of the time,
imported 2026-07-12 Qwen Candidate-Trace Verifier
Yes. A small checker that reads each candidate's worked-out steps — never the true answer — lifted correctly-running programs from 30% to 54%, and to 56% when it cross-checks two wordings of one task. The twist: a picker
imported 2026-07-12 Qwen 3.5 4B Verified Edit Closure
Yes, on the hardest unseen tasks. Testing small edits of the model's own near-miss program and keeping whichever passes the example cases raised fully-correct answers from 39% to 52%, while re-sampling the model and re-r
imported 2026-07-12 Qwen 3.5 4B GraphIR Self Repair
No. The plain one-line formula fully solved 29% of brand-new tasks; the step-by-step diagram managed only 22%, and the fix-it pass recovered part of that gap to 24% — still behind. That pass genuinely works: on randomly
2026-06-27 Real Transform ABI Gate with Counterexamples
It depends on how messy the data is. For clean, spreadsheet-style pipeline jobs the toolbox covered every one (100%) and held firm even against deliberately tricky examples. For irregular date, ID, and text cleanup, cove
2026-06-27 Qwen Verified Skill Memory RAG
No. Handing the model a matched, verified example made it slightly worse, not better: every value came out correct on 47.5% of jobs, versus 50% when it worked through each value alone with no example. The lookup itself w
2026-06-27 Pairwise Table Judge
No. Letting the model run a knockout comparison among its own candidate tables did not help: fully-correct tables slipped from 50% (just trusting its first attempt) to 47.5%. On the hard cases where a correct table exist
2026-06-27 Counterexample-Guided Consistency Projection
No. Distilling the model's messy guesses into one reliable formula solved only 20% of tasks completely, versus 50% when it simply transformed each row directly, a 30-point drop. For 29 of the 40 tasks no formula even fit
2026-06-27 Qwen3.5-4B Foofah Program Ensemble Consensus
No. The simplest rule won: run the first program that passes a single worked example. It solved 52% of tables versus 44% when the model just answered directly, rescuing 23 tables it had otherwise botched while breaking o
2026-06-27 Qwen3.5-4B Code ABI Oracle Coverage Ladder
Yes, mostly. Snapping together at most three verified functions produced a passing solution for 84% of 160 small programming tasks, up from just 13% with a bare-bones library. But the library is a foundation, not a solve
2026-06-26 Qwen3.5-4B Verified Algorithm Retrieval Adaptation
Yes, but modestly. Fetching the most similar already-solved problem and adapting its code cracked 3 of the 8 problems that repeated tries kept missing, lifting overall success from 67% to 79%. Matching by meaning clearly
2026-06-26 Qwen3.5-4B Substrate Coverage Ladder
Only when a building-block was hand-carved for each specific problem. A shared library that reshapes old solutions to fit new ones solved zero of the nine stuck tasks, no better than nothing. Custom-written parts solved
2026-06-26 Qwen3.5-4B Retrieval Adapt Verify Scale
It splits. Pulling a similar solved problem from a 364-entry library and rewriting it recovered 8 of 24 stuck tasks — a third, double the 4 that a random-library control found. But the only safe check, the visible sample
2026-06-26 Qwen3.5-4B Reliability Exec OPSD Audit
No. All three cheap fixes failed. Ranking programs by the model's own confidence picked working code less often than just taking the first one that passed the visible sample tests (6 of 24 versus 8), and let more bugs sl
2026-06-26 Qwen3.5-4B Offline Hard Negative Coverage DPO
It depends — and even the gentle version failed its own bar. A light touch matched the task-solving reach of generating twice as many candidate answers at roughly half the compute, and beat a shuffled-pair placebo. But i
2026-06-24 → 26 Qwen3.5-4B Oracle Probe Synthesis MDP
It's the menu. Just curating which eight test inputs the model chose from raised success from about 43% to 49% — a bigger jump than any training gave. Supervised coaching added a bit more (48% to 51%); preference- and re
2026-06-25 Qwen Readable Candidate Verifier
Barely. Training a picker on the model's reading of each fix's pseudocode plus its claimed output raised the share of tasks fixed correctly from 44% to 51%, closing only about 15% of the distance to a perfect picker's 90
2026-06-25 Qwen Candidate-Conditioned Trace Verifier
No. Doing nothing already solved about 44% of tasks, and always choosing a correct candidate could reach 90%. Yet no trained picker captured any of that headroom; the best merely tied doing nothing. Left unchecked, the m
2026-06-25 Qwen3.5-4B Verifier-Guided Self-Improvement Report
No. After a four-billion-parameter model retrained on its own test-passing code, its success on unseen problems barely moved, going from about 65% to about 65% (a hair lower). Retraining made its several attempts look mo
2026-06-25 Qwen3.5-4B Real Sample Verify Commit
It's the writing. When a 4-billion-parameter model writes several code attempts per problem, choosing is not the bottleneck. On easier problems nearly every attempt is already correct, so the best result sits near 97%. O
~2026-06-25 Qwen3.5-4B Oracle-Distilled Semantic Verifier
Yes, but only on home turf. On the problem set it trained on, the trained judge picked a genuinely-correct program 81% of the time, up from 72% for the same model untrained and just 44% for grabbing the first candidate t
~2026-06-25 Qwen3.5-4B HumanEval Adaptive Evidence Budget
No. Every strategy landed at the same 16.7% correct-pick rate — running no tests, running all eight, and even a strategy allowed to peek at the right answer to time its stop. The trained model did learn to quit sooner, u
2026-06-24 Qwen3.5-4B Sketch Coverage Shift Probe
Only when the template explicitly names the new operation. On tasks needing a brand-new operation, hand-written templates that named and typed it kept the correct program every single time; auto-generated templates kept
2026-06-24 Qwen3.5-4B Oracle Process GRPO
Yes, but with sharp limits. Teaching the model to copy a flawless test-picker lifted its success from about a third of puzzles to 43%, versus roughly 31% for picking blindly. Layering on preference-based and reward-based
2026-06-24 Qwen3.5-4B Adaptive Evidence Budget Policy
Yes. Trained just for this single decision, the model matched the accuracy of running all ten checks — solving about 92 of every 100 tasks — while using only about five checks on average, near a perfect-hindsight stopper
2026-06-24 Qwen3.5-4B Active Counterexample Trace Selection
Yes, but choose them well. Committing on the visible examples alone left about a quarter of picks secretly wrong, even though every one passed all the examples shown. Requesting six new test cases where the surviving pro
2026-06-23 Qwen Typed Bytecode Expert Iteration
It depends. Training the model only on its own attempts that landed on the correct final answer lifted unaided first-try accuracy from 62% to 73% on fresh problems — a real gain that sticks when it writes programs alone.
2026-06-23 Qwen Prefix-State Process Verifier
Barely. The judge got genuinely good at telling promising partial programs from dead ends, yet it lifted first-try accuracy only a few points, with hard problems going from 41% to 44%. The real gap: on hard problems a co
2026-06-23 Qwen On-Policy Repair-to-Compiler
Yes, but the self-correction isn't the secret ingredient. On brand-new requests the model climbed from about 29% correct programs to 99% after training on its own corrected drafts. The catch: training on plain correct pr
2026-06-23 Qwen Mixed-Domain Trace Verifier
Yes, partly. On fresh tasks the frozen model alone got 46% right; the proofreader lifted that to 57%, and a cross-check that compares reworded versions of the same task reached 61%. But a correct recipe was already among
2026-06-23 Qwen Hidden VM Curriculum Repair
No. Feeding it nearby worksheets that merely land on the correct answer wrecked it. Plain step-by-step training scored 72% on the main mixed test; the same model after answer-chasing repair fell to 35%, and collapsed on
2026-06-23 Qwen Context-Conditioned Trace Verifier
Mostly no. A correct program was almost always sitting in the candidate pile, for 91 to 100 percent of questions, yet the judge barely helped: it nudged easy questions from about 69 to 70 percent and actually made the ha
2026-06-23 Qwen Complete-Program Trace Reranker
Barely, and it backfires on hard cases. A correct program sat in the candidate pile 91 to 100 percent of the time, yet the trained picker nudged easy prompts only from 69 to 70 percent and actively hurt longer ones, drop
2026-06-22 Qwen Verifier-Guided Slot Repair
Yes, mostly, but with a catch. A small model copying 24-step calculations got only about 27% exactly right on its own. A checker that knows the correct running number after every step, allowed to swap one or two bad step
2026-06-22 Qwen Teacher-Distilled Slot Compiler
No. A model trained to copy numbers and operations out of text and run a 24-step calculation got 27% of final answers exactly right; adding the pointing signal landed at 28%, a tie. Worse, agreement between two rewording
2026-06-22 Qwen 3.5 4B Model-In-Loop Counterexamples
No. Building practice cases from the model's actual wrong answers matched, but never beat, simply hand-picking the tricky categories in advance. Both lifted the hardest problems from 64% to a perfect 100% passing every h
2026-06-22 Qwen 3.5 4B Counterexample-Directed DSL
It depends. Hand-picked examples lifted the model's single-best-guess repair rate from 51% to 58% over random examples. But when it generated several candidates and kept the best, the edge vanished (64% slipped to 61%).
2026-06-20 Counterexample Rule Repair Experiment
Yes, but only for rule shapes it has practiced. Given the failing examples, a small fine-tuned model fixed 91% of bugs on familiar tasks and got unseen inputs right too, versus 0% untrained and near-zero for every versio