Qwen3.5-4B Tokenizer-EOS Answer Commit Factorial

Results at a glance 1

Overview

Calibration qualified the tokenizer-EOS-only interface; mechanics transport passed 24/24, but visible selection ended in terminal instrument failure after all five transactions. Residual capability remains unadjudicated.

This fresh successor tests whether the prior strict answer-seam failure was caused by waiting past Qwen3.5's tokenizer chat-end token. It registers the first tokenizer EOS only during answer generation, preserves every pre-commit byte under the strict grammar, and compares it with the existing HF-model-EOS boundary on new calibration and residual-mechanics identities.

Research Program

• Program: structured_execution_and_compilers
• Program question: can a token-native output boundary expose a deployable structured compiler interface without supplying answer identity?
• Closest result-bearing predecessor: qwen35_4b_materialized_residual_answer_seam_factorial.
• Prior evidence: the predecessor was 0/48 strict in every registered arm, but suffix-only removal gave 48/48 frozen-parser exactness in both paired no-think cells; thinking remained 38/48 and 24/48 because of extra closes.

Question

On fresh known-answer rows, does stopping the answer stage at the first tokenizer EOS (248046) create a strict exact program interface that the matched HF-model-EOS (248044) boundary does not—and, only if it qualifies, does that interface permit a disjoint materialized-residual capability test?

Hypothesis

Qwen3.5 naturally commits short structured answers with <|im_end|>. The predecessor treated that token and the following newline as answer content because it waited for <|endoftext|>. Registering first-248046 answer-stage stopping should lift no-think exact/parse from the predecessor's 0/48 to at least 44/48 without changing prompts, answer identity, or thought generation. This is an interface hypothesis, not yet a capability hypothesis.

Setup

• Model: only Qwen/Qwen3.5-4B revision 851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a, bf16.
• Backend: pinned repository vLLM for every paired arm and matched-compute baseline.
• Dataset/task source: fresh procedural exact-depth-three list transforms; never benchmarks/.
• Frozen fresh splits: 48 known-answer calibration tasks and 24 disjoint residual-mechanics tasks, with new functions, task IDs, record IDs, token sequences, and seed domain.
• Factorial: answer boundary {first tokenizer EOS, HF model EOS control} x thinking {off, forced close at 512} x answer prefix {freeform, PROGRAM:}.
• All 192 boundary pairs must share prompts, answer seeds, persisted thought IDs, request adjacency, and sampled token prefix through the earliest stop or cap; one mismatch terminates the experiment before qualification.
• Every task has one persisted thought transaction reused by all four thinking answer cells. Natural post-close answer content is discarded and cannot bypass the answer-stage comparison.
• Answer-stage tokenizer EOS never applies to thought generation.
• Calibration gates: >=44/48 strict exact echoes, >=44/48 parses, <=2/48 answer-cap contacts, plus >=22/24 exact/parse and <=1/24 cap contact in each arity.
• Parse is membership in a frozen token-ID grammar for any A-X program of the registered arity; exact is equality to the known answer. A stop on sampled token 24 still counts as a cap contact.
• Eligible winner: first qualifying tokenizer-EOS arm in the frozen no-think PROGRAM:, no-think freeform, think PROGRAM:, think freeform priority.
• Boundary controls: matched HF EOS, first-stop uniqueness, early/interior/ missing tokenizer EOS, extra pre-commit bytes, and exact stop/token/cost receipt authentication.
• Conditional mechanics is fully frozen at 24 tasks, 24 candidates per suffix arm, and a 96-row-per-task direct ceiling. Controls are materialized state, name-only, shuffled materialized state/target, candidate-blind direct samples, exhaustive CPU ceiling, and taskwise sampled/logical-token matched- compute prefixes.
• Every suffix is bound to its semantic candidate first operation before any deduplication or scoring, yielding the same canonical full three-operation proposal type used by direct sampling.
• Primary capability metric if opened: hidden exact accuracy of a frozen visible-only selector versus all structured controls and matched-compute direct sampling.
• Hidden-label boundary: mechanics remains sealed until a committed-green calibration decision and second winner-bound lock; hidden labels remain sealed until a committed-green visible-selection receipt.

The prefix cells share tasks and seed derivation and are paired conditions, not independent replications. Only tokenizer-pass/HF-fail in the selected matched cell supports a causal termination-boundary claim. Dual qualification within one matched thinking/prefix pair is impossible under the authenticated-prefix, exactness, and cap gates: shared exact-cap overlap is <=2 globally and <=1 per arity, yielding 44+44-2>48 and 22+22-1>24. Any observed dual qualification terminates as a scoring invariant violation.

Historical reproduction commands

This experiment is finished and its result-bearing mechanics transaction must not be executed or rerun. The commands below document historical provenance only. Use the fresh successor named in the report for any further model call.

Model-free smoke:

python3 -B experiments/qwen35_4b_tokenizer_eos_answer_commit_factorial/scripts/run.py --smoke

Fresh construction:

python3 -B experiments/qwen35_4b_tokenizer_eos_answer_commit_factorial/scripts/construct.py

Historical calibration entry points:

experiments/qwen35_4b_tokenizer_eos_answer_commit_factorial/scripts/calibration_launcher --stage lock
experiments/qwen35_4b_tokenizer_eos_answer_commit_factorial/scripts/calibration_launcher --stage run
experiments/qwen35_4b_tokenizer_eos_answer_commit_factorial/scripts/calibration_launcher --stage analyze

Historical winner-bound mechanics entry points (do not execute):

experiments/qwen35_4b_tokenizer_eos_answer_commit_factorial/scripts/mechanics_launcher --stage lock
experiments/qwen35_4b_tokenizer_eos_answer_commit_factorial/scripts/mechanics_launcher --stage run
experiments/qwen35_4b_tokenizer_eos_answer_commit_factorial/scripts/mechanics_launcher --stage analyze-visible
# Hidden scoring was never authorized and must not be executed:
experiments/qwen35_4b_tokenizer_eos_answer_commit_factorial/scripts/mechanics_launcher --stage score-hidden

calibration_launcher is a 12.6-KiB static x86-64 ELF with no dynamic interpreter. It retains a waiting static parent, opens its exact executable on inherited descriptor 198, discards the inherited environment, and directly execves the pinned Python interpreter with -I -B in a child protected by a parent-death signal. The Python bootstrap requires the live parent executable, the inherited descriptor, and the tracked launcher path to be the same stable inode with the reviewed SHA-256. This kernel-carried provenance cannot be forged by an environment marker. Its source is scripts/calibration_launcher.S; its exact reproducible build command is:

/usr/bin/gcc -nostdlib -static -no-pie -s -Wl,--build-id=none \
  -o experiments/qwen35_4b_tokenizer_eos_answer_commit_factorial/scripts/calibration_launcher \
  experiments/qwen35_4b_tokenizer_eos_answer_commit_factorial/scripts/calibration_launcher.S

The mechanics launcher applies the same kernel-carried provenance design to the winner-bound runner. Its exact reproducible build command is:

/usr/bin/gcc -nostdlib -static -no-pie -s -Wl,--build-id=none \
  -o experiments/qwen35_4b_tokenizer_eos_answer_commit_factorial/scripts/mechanics_launcher \
  experiments/qwen35_4b_tokenizer_eos_answer_commit_factorial/scripts/mechanics_launcher.S

Direct Python invocation—even with a caller-supplied marker and an open copy of the launcher on descriptor 198—fails because its live parent is not the static launcher. Missing/malformed stages, non-isolated Python, altered launcher bytes, and dynamic-loader variables also fail before local imports.

Results

The model-free protocol smoke passes with zero model calls. It accepts only a clean expected answer followed by one terminal registered stop token. The matched HF boundary preserves tokenizer EOS/newline as strict answer content and therefore fails exactness. Early stops, interior-plus-terminal stops, missing claimed stops, and extra pre-commit bytes all fail exactness or their registered contract as appropriate. A unique final early stop is authenticated but scored; an exact-cap length trace is authenticated with all tokens retained as content.

After six prospective holds, global independent rereview returned PASS_DESIGN for exact pushed/green commit abd2ffcd. No model or capability result exists; live execution remains sealed behind implementation release review and a committed-green lock.

Model-free construction then passed with 48 calibration and 24 mechanics tasks, the frozen 8/8/4/4 strata in each calibration arity and mechanics, every A-X alias once per answer position, zero overlap with all 72 predecessor public fingerprints, zero request/seed/prompt/derived-seed overlap, and a distinct transport namespace. Mechanics gold exists only as tracked AES-256-GCM ciphertext; its key is local and ignored. Model calls and sampled outputs remain zero.

Before any model call, the prospective runner was security-hardened to pin child executables/environment and preserve isolation across recovery re-exec. The tokenizer receipt's sole runner binding was correspondingly refreshed from cbbfae3e... to 4ce61e64...; its grammar, prompts, termination IDs, freshness inventories, and zero-call declarations were unchanged.

A later clean implementation rereview held live calls because an in-Python LD_PRELOAD rejection occurs after interpreter startup and because several zero counters still accepted JSON Boolean aliases. The prospective static launcher now sanitizes before Python exists, and exact integer predicates plus typed-canonical engine comparisons reject those aliases. These are model-free repairs; no calibration request or sampled output has yet occurred.

The next clean rereview held again because the first static-launch repair used a caller-controlled environment marker to assert entry. The parent-plus-open- executable proof above replaces that assertion with kernel process/file state that remains valid across the sanctioned Mamba recovery execve. The repair is still model-free; live authorization remains absent.

The seventh clean review then returned PASS_IMPLEMENTATION for exact pushed-green commit d70756122bc768e82fa4d77a61e05522ef5bca79: 95/95 permitted tests passed, the launcher rebuilt byte-identically, all protected read inventories were empty, and model/GPU calls remained zero. The PASS does not itself authorize generation; its canonical hash-bound receipt and the subsequent implementation lock must each be committed, pushed, and green.

The independently committed receipt and implementation lock both passed their two required workflows. The sealed live calibration then completed 48 shared- thought requests plus 384 paired answer requests. All 192 boundary pairs and the five-invocation transaction chain authenticated. Results were:

The frozen decision is TOKENIZER_EOS_ONLY_INTERFACE_QUALIFIED; the winner is tokenizer_eos_no_think_program_slot, and its matched HF control is hf_model_eos_no_think_program_slot. This opens the preregistered conditional mechanics branch only after this calibration result and a second winner-bound lock are committed, pushed, and green.

The first exact-commit conditional-mechanics review returned HOLD_IMPLEMENTATION for fd06b5053c9327c61775d07061c9a84e070cdcb6. It found four live-stage deadlocks (an incomplete path-audit allowlist, an invalid calibration-verifier call, and two Python-tuple/JSON-list durable comparison failures), partial type/schema authentication, a missing registered direct-pool-exhaustion terminal, and incomplete resource receipts. The prospective repairs add exact calibration/review support reads, a scoped and tested immutable-verifier adapter, JSON-domain receipts, exact typed engine/prompt/seed/token/terminal/cost authentication, omission of unrequested likelihood diagnostics, a durable DIRECT_RESOURCE_MATCH_POOL_EXHAUSTED receipt, and explicit overshoot/row-ID inventories. The model-free suite now passes 124/124. These repairs do not authorize mechanics; a fresh pushed-green exact-commit adversarial review and second lock are still required.

The round-two review of pushed-green commit 3d2f051203c56456fad716e20950b55d5717afd5 verified all seven earlier repairs and held on three new exact-type replay failures: the recorded mechanics preflight authenticated only a runtime subset, visible authorization inherited Python's Boolean/integer equality, and the calibration-era generic transaction envelopes had the same alias at STARTED and bundle schema boundaries. The prospective repair now reconstructs the entire clean preflight from the live runner and exact-compares every JSON-domain field after authenticating recorded CI. Visible authorization also uses recursive exact-type comparison. Because the generic transaction source is an immutable calibration-critical file, mechanics routes through a new additive exact-typed transaction layer rather than altering the calibration anchor. Durable hostile tests cover all ten preflight mutations, visible true versus 1, STARTED/bundle aliases, receipt aliases, and registered row-count aliases. The full model-free suite passes 130/130; mechanics calls and protected reads remain zero. A third fresh pushed-green exact-commit review is still required.

That round-three review of pushed-green 9b527cbfd6934f579b0ebcb07cb4b695c370798b returned HOLD_IMPLEMENTATION despite confirming all round-two closures. It found that production tuple-valued logprob_token_ids would mismatch the JSON-native STARTED sampling after the first call, that hidden scoring reread the visible receipt after authorization, and that the low-level transaction primitive authenticated only a predecessor's state before a fresh successor call. The repairs normalize the complete fresh generated bundle before its first validation/write, pass the exact in-memory authorized visible object directly into scoring, and fully authenticate plus recheck the predecessor chain inside the primitive itself. Tests now use the actual frozen mechanics sampling plan, assert no second visible read, and corrupt a predecessor before a fresh successor. The full model-free suite passes 134/134; mechanics calls and protected reads remain zero. A fourth pushed-green exact-commit review is required before any lock or generation.

Round-four probes confirmed all production-shape and static-corruption repairs, then held on two narrower concurrent-mutation windows: a predecessor could change inside generation or promotion after the pre-call recheck, and visible authorization compared HEAD:path before separately resolving the commit recorded in its receipt. The primitive now rechecks the authenticated predecessor immediately after generation and again before returning from every promotion/recovery path. Visible authorization resolves one commit exactly once and uses that immutable commit ID for its blob comparison and receipt. New callbacks mutate the predecessor during generation and at COMPLETE publication; both are detected before bundle publication or successful return. The full model-free suite passes 136/136 with zero mechanics calls or protected reads. A fifth pushed-green exact-commit review is required.

Round five confirmed every earlier regression family but found that the pre-return predecessor recheck covered only the terminal COMPLETE file. A concurrent change to an already-authenticated predecessor STARTED, bundle, or GENERATED artifact could therefore escape the primitive's successful return until final chain authentication. The primitive now reruns exact authentication over the entire predecessor prefix at each recheck and compares the resulting receipt with the original authenticated snapshot. New regressions change each non-terminal predecessor artifact during generation, publication, or recovery; all fail closed without a successful successor return or recovery resample. The model-free suite passes 139/139 with zero mechanics model/GPU calls or protected reads. A sixth pushed-green exact-commit review is required.

Round six returned PASS_IMPLEMENTATION for exact pushed-green commit df096d330f09847ce844af6255b349b4f707f464. It passed 139/139 tests, rebuilt the static launcher byte-identically, rejected all full-prefix changes across pre-call, callback, publication, and recovery checks, and reconfirmed the production tuple path, hidden object/commit binding, exact typed gates, resource receipts, immutable calibration files, and bootstrap routing. All protected-read arrays and model/GPU counters were zero. The hash-bound review report and canonical receipt must now be committed, pushed, and green before the mechanics lock can be published; this PASS alone does not authorize a model call.

After that receipt was published green, the first lock-only invocation failed closed before creating a lock: calibration decision recomputation retained 32 integer dictionary keys while canonical JSON represented those object keys as strings. There were zero actual tuple/list mismatches. The lock compared across that serialization boundary without first entering the JSON domain. The prospective mechanics verifier now JSON-normalizes the recomputation and then applies recursive exact- typed comparison, preserving rejection of Boolean/integer aliases. A direct model-free recomputation now authenticates the frozen TOKENIZER_EOS_ONLY_INTERFACE_QUALIFIED decision and winner. No mechanics data, model, or GPU call occurred. Because this changes reviewed mechanics code, a fresh exact-commit implementation review and replacement receipt are required before retrying the lock. The model-free suite passes 140/140.

Round seven verified that real decision repair, rejecting 20/20 typed aliases and four semantic mutations, and completed a read-only prospective lock build. It nevertheless held on the next exact-type boundary: durable sampling-plan validation used ordinary Python equality, accepting five n=1 to true aliases and 25 Boolean-to-integer aliases across the five arms. Sampling-plan validation now uses recursive exact JSON equality. The real-plan regression mutates both directions for all 30 fields and requires every case to fail. No mechanics payload, model, or GPU call occurred; another exact-commit review and replacement receipt remain required.

Round eight returned PASS_IMPLEMENTATION for exact pushed-green commit 3e7b650a90ff1d65fe371552354895756efcf728. It passed 140/140 tests; rejected all 58 real-plan type/schema cases and all five semantic seed mutations; reauthenticated the real calibration decision with the correct 32 key conversions; and completed read-only lock build/validation over 29/29 current critical files. Full-prefix, production durability, hidden binding, preflight, resource, calibration, routing, launcher, and static checks also passed. All mechanics/protected read inventories and model/GPU counters remained zero. The replacement hash-bound receipt must be committed, pushed, and green before the lock-only stage is retried.

After that replacement receipt was published green, the next lock-only attempt failed closed while hashing reviewed critical evidence: the pre-import path audit allowed all 22 runtime files and 11 support files but omitted the seven reviewed mechanics test files in the 29-file critical inventory. The lock was not written, and no mechanics payload, model, or GPU call occurred. A separate exact seven-file critical-test allowlist now permits only those review inputs; it does not widen runtime imports, prepared/mechanics data, or hidden access. The bootstrap test proves that this tuple is exactly the critical-minus-runtime set and is actually consumed by the path audit. Because bootstrap-reviewed code changed, another exact-commit review and replacement receipt are required. The model-free suite passes 141/141.

Round nine returned PASS_IMPLEMENTATION for exact pushed-green commit c0075a019fd0f202c3b0e6cf0be5528e08c61649. It passed 141/141 tests; proved the exact seven-file critical-test inventory; hashed all 29 critical files through the active audit while denying neighboring files; preserved the 22/11 runtime/support inventories; and passed every prior typed, semantic, full-prefix, durability, hidden-binding, calibration, launcher, and resource regression. The prospective lock build/validation was read-only, and mechanics payload reads, model/GPU calls, and protected-read inventories remained zero. The replacement hash-bound receipt must be committed, pushed, and green before the lock-only stage is retried.

The replacement receipt was committed, rebased, pushed, and green. An initial lock retry then failed closed because the shared-main rebase advanced HEAD to a commit whose Validate Repository workflow was still running; it wrote no lock and made no mechanics/model/GPU call. After both workflows for that exact HEAD passed, lock-only publication succeeded. The canonical implementation lock has SHA-256 d42e2db4b589e470f42d963b19e01a8b880fa7858a40b10966150c775c3d925b, binds all 29 critical and 22 runtime files, and records zero pre-lock generation requests, sampled outputs, and protected reads. Live mechanics remains unauthorized until this lock is itself committed, pushed, and green.

The lock was committed and both exact-commit workflows passed before live mechanics. The selected interface then transported perfectly: 24/24 exact echoes, 24/24 parses, and zero answer-cap contacts, balanced 12/12 in each arity. All five durable transactions completed and preserved 4,056 sampled outputs. Visible analysis nevertheless failed before selection because it recomputed the already-recorded transport decision through the initial authorization path, which correctly requires every later invocation to be absent; at replay time direct and all suffix invocations were already complete. No visible selection, resource decision, hidden score, hidden read, or benchmark read occurred. This is terminal instrument failure, not evidence for or against residual capability. The generated result is preserved, and a fresh-identity successor must separate initial transport authorization from post-chain replay authentication under a new review and lock.

Interpretation

The fresh result isolates a real termination-boundary effect: with identical paired prompts/seeds and authenticated shared prefixes, stopping on tokenizer EOS yields 48/48 strict structured answers while waiting for HF model EOS yields 0/48 because tokenizer EOS/newline remains pre-commit content. This is an interface qualification, not a capability gain. Thinking is actively worse at this interface (38/48 and 30/48, including 16 freeform cap contacts), so the no-think PROGRAM: winner was the only branch that advanced from calibration. That branch is now frozen after terminal instrument failure; only the fresh-identity successor may continue it.

Any conditional mechanics pass is scoped to a 24-task contamination-free large-effect pilot; it is not a confirmatory or general deployability claim.

Knowledgebase Update

• Program evidence, backlog, scorecards, and shared synthesis now record the committed-green calibration and transport evidence, terminal mechanics instrument failure, and fresh-successor requirement.
• Claim-ledger allocation is deferred: the repository's knowledgebase protocol forbids adding or promoting claims until the outstanding adversarial re-grade checklist is processed. No claim ID is consumed, avoiding a collision while preserving the result at program level.

Artifacts

• src/protocol.py
• scripts/run.py
• configs/default.yaml
• runs/smoke/summary.json
• reports/preregistration.md
• reports/design_review.md
• reports/mechanics_failure_review.md
• reports/artifact_manifest.yaml
• runs/calibration/implementation_lock.json
• runs/calibration/live_preflight.json
• runs/calibration/decision.json
• src/mechanics_runtime.py
• src/mechanics_stage.py
• src/mechanics_lock.py
• scripts/run_mechanics.py
• scripts/mechanics_launcher
• runs/mechanics/failure.json

Report

Rendered from reports/report.md

Summary

Fresh sealed calibration passed: the tokenizer-EOS no-think interface was 48/48 exact and parseable in both prefix cells, while every matched HF-model- EOS control was 0/48. All 192 boundary pairs authenticated. The frozen decision is TOKENIZER_EOS_ONLY_INTERFACE_QUALIFIED, with tokenizer_eos_no_think_program_slot selected for the conditional mechanics stage. This establishes an interface effect, not yet a capability gain.

Research Program Fit

This is the required fresh successor to the terminal HF-EOS answer-seam factorial. It changes one causal boundary while preserving strict grammar and the full residual-mechanics firewall.

Method

The prospective design crosses answer boundary, thought policy, and answer prefix on fresh known-answer rows. First-stop uniqueness and all pre-commit bytes are strict. Mechanics is conditional on an independently qualified tokenizer-EOS interface.

Results

The committed-green implementation receipt and lock authorized 48 shared- thought requests and 384 paired answer requests. The transaction chain contains 432 sampled outputs across five invocations; all receipts, runner metadata, token costs, paired prefixes/seeds, thought-source rows, and terminal events authenticated.

The winner is the first qualifying cell in the preregistered priority: tokenizer_eos_no_think_program_slot. Its matched control is hf_model_eos_no_think_program_slot. Calibration decision SHA-256 is 2195c3d15ad9ece124ad99244baf637076963e53fc2532f1131ea01226d11ee5. No mechanics, hidden label, qualification, confirmation, or benchmark content was read.

Conditional mechanics

After a separate nine-round implementation review and committed-green winner-bound lock, live mechanics completed all five durable transactions and preserved 4,056 sampled outputs. The transport gate independently confirmed the selected interface: 24/24 exact echoes, 24/24 parses, zero cap contacts, and 12/12 successes in each arity.

Visible selection then stopped with terminal instrument failure. Its replay authentication recomputed the stored transport decision through the initial authorization path, whose temporal invariant correctly requires all later invocations to be absent. By then direct and all three suffix invocations were durably complete. Thus no visible selection or resource decision was written, hidden scoring was never authorized, and hidden and benchmark reads remained zero. The residual-capability comparison is unadjudicated.

Controls

Matched HF EOS, all-192-pair fail-closed authentication, one persisted thought per task, malformed stop traces, strict arity-parametric token content, fresh identities, same backend, and fully frozen conditional matched-compute mechanics are preregistered.

Oracle Versus Deployable Evidence

The 48 known-answer rows and the 24-row transport gate are interface instruments, so their perfect scores demonstrate a deployable strict answer boundary on this fresh substrate but not new task capability. Mechanics candidates were sampled, but the failed visible selector left hidden labels and oracle ceilings sealed. Only a fresh successor with a valid winner-bound residual-mechanics result can support a capability claim.

Interpretation

The fresh matched result supports the causal interface hypothesis. The paired arms generated the same answer-token prefix through the earliest stop, but the HF boundary retained tokenizer EOS plus newline as answer content and therefore failed strict grammar/exactness on all 48 rows. Tokenizer EOS commits exactly the intended alias sequence on all 48 rows in both no-think prefixes.

Thinking does not magnify this interface. The structured think512 cell fell to 38/48 and freeform to 30/48 with 16 cap contacts. This reinforces a practical lesson already visible elsewhere in the corpus: extra reasoning can damage a short exact-output channel even when it sometimes helps semantic reasoning.

Third-pass review additionally required one canonical proposal type. Every two-operation suffix is now bound to its semantic candidate first operation, and all suffix/direct proposal processing uses the resulting full three- operation tuple.

Fourth-pass review separated authenticated terminal geometry from answer correctness: unique-final early stops and exact-cap length events are scored, while malformed stop/length/reason geometry is an authentication failure.

Fifth-pass review replaced tokenizer-dependent selector identity with a semantic base-24/two-byte program ID and corrected the dual-qualification proof to account for the at-most-two exact-cap overlaps.

Sixth-pass global review aligned the intake and README summaries with that cap-bounded proof. At that pre-execution checkpoint no scientific or implementation result existed; the later calibration and mechanics records supersede that historical state.

Seven implementation review rounds (five countable) ended in exact-commit PASS_IMPLEMENTATION; the machine receipt, implementation lock, and their required CI workflows all authenticated before model execution.

Fresh model-free construction now passes. It produced 72 predecessor-disjoint tasks, all frozen calibration/transport/mechanics request inventories, balanced arity/stratum/alias receipts, and zero request/seed/prompt/derived-seed collisions. Mechanics gold is sealed as tracked AES-256-GCM ciphertext with an ignored local key. Those were pre-execution facts. The authorized live run later preserved 4,056 sampled mechanics outputs but failed before visible selection, as reported above.

Next Experiments

Create a fresh-identity successor rather than altering this result-bearing experiment. Preserve the initial generation-time invariant that transport must be the sole completed invocation; add a distinct replay path that authenticates the exact stored transport decision against the transport bundle while the separate full-chain verifier requires all later invocations complete. Mutate every temporal state in tests, obtain a new adversarial review and lock, and use new task IDs, record IDs, token identities, seeds, ciphertext, and key. The capability test must still beat every structured control and taskwise sampled- and logical-token-matched direct baseline.

Artifact Manifest

All calibration and failed mechanics artifacts are tracked. The only external artifact is the ignored AES-256 key; it was not used because visible selection never authorized hidden scoring.

Experiment log 2

Data files 2

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

• previewruns/construction/summary.json15 kB
• previewruns/smoke/summary.json3.4 kB

Reproduce

Smoke test

python3 -B experiments/qwen35_4b_tokenizer_eos_answer_commit_factorial/scripts/run.py --smoke

Full run

experiments/qwen35_4b_tokenizer_eos_answer_commit_factorial/scripts/calibration_launcher --stage run

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

Browse the experiment folder on GitHub ↗

Related

Related experiments

Nearest by shared name, tags, and programs.

• 2026-07-14 Qwen3.5-4B Tokenizer-EOS Residual Mechanics Fresh Replay
• 2026-07-13 Qwen3.5-4B Materialized Residual Answer-Seam Factorial
• 2026-07-12 Qwen3.5-4B Balanced-Core Answer-Potential SFT
• 2026-07-12 Qwen3.5-4B Commit-Slot Semantic Power Replication
• 2026-07-12 Qwen3.5-4B Commit-Slot Jacobian Value Transport
• 2026-07-10 Qwen3.5-4B Answer-Potential Trace SFT