Qwen trajectory snapshot dogfood — 2026-08-27

Status: VERIFIED after #9702 and #9703; the historical pre-fix capture remains REFUTED below.

This is the public-safe readout for #9697 and advances #8929. It exercised the snapshot spine at 1f7aebc59e7cafaaa55ffe2bf6a8b60357756181 from origin/main at e2db29f84963679e574a9174dbf08ef09ae2c857. Raw transcripts, manifests, and full reports stayed in private allocated scratch. This note contains no transcript text, tool arguments, tool outputs, or machine-local paths.

Commands and result

The private scratch variable below intentionally replaces the machine-local absolute path. Both capture attempts used this command shape with distinct empty snapshot targets:

PRIVATE_SCRATCH=<allocated-private-scratch>
go run ./cmd/fak trajectory audit \
  --since 7d \
  --user-contains qwen \
  --snapshot-out "$PRIVATE_SCRATCH/snapshot-a" \
  --jsonl "$PRIVATE_SCRATCH/capture-a.jsonl" \
  --md "$PRIVATE_SCRATCH/capture-a.md"

Both independent attempts copied the selected inputs into private staging and then failed their built-in verification with the same typed result:

TRAJECTORY_SNAPSHOT_REFUSED SNAPSHOT_OUTPUT_CHANGED
trajectory audit snapshot: SNAPSHOT_OUTPUT_CHANGED: current audit output does not match the captured audit schema

The fail-closed cleanup removed each staging snapshot. Consequently no manifest or corpus digest was published, and the required replay command could not safely run:

go run ./cmd/fak trajectory audit \
  --snapshot "$PRIVATE_SCRATCH/snapshot-a" \
  --jsonl "$PRIVATE_SCRATCH/replay-a.jsonl" \
  --md "$PRIVATE_SCRATCH/replay-a.md"

Capture success was 0/2, replay success was 0/0, and capture/replay byte identity is not established. This does not satisfy the issue’s 100% capture/replay and zero-drift operating envelope.

The content-free live fallback was run separately:

go run ./cmd/fak trajectory audit \
  --since 7d \
  --user-contains qwen \
  --jsonl "$PRIVATE_SCRATCH/qwen-live-final.jsonl" \
  --md "$PRIVATE_SCRATCH/qwen-live-final.md"

It completed with 430 canonical sessions from 432 raw fragments, 311,053 records, 48,033 exact usage records, and zero parse refusals. Accounted usage was 259,086,472 input, 4,884,324,480 cache-read, and 16,874,503 output tokens. The top ten held 2,240,250,063 tokens, or 43.41% of the 5,160,285,455-token cohort. Private output identities were:

JSONL SHA-256: 9528208ddf7e13fd530dce1794452f16e06b54ee6df2493e5e2a2f03797480f3
Markdown SHA-256: 0c8148fdc992a38fa75f5f52a0d5f7c7613dcf0c02fdb7260b323ef285827b07

Top-ten attribution

Turns counts Codex task-start events. Errors is the audit’s tool-error count; expected bounded waits are listed separately because they are not failures. No top-ten session had mutation churn. A public commit is named only when its subject and ancestry on current origin/main independently witness the effect.

Rank	Session	Input / cache-read / output	Turns	Tools / errors / expected waits	Classification	Public landed effect	Conservative counterfactual
1	`01a03b0a-0981-7ba3-ae80-7fd5402239cd`	5,400,182 / 373,675,776 / 458,884	2	3,002 / 44 / 1,100	Productive long Qwen work; one repeated failed action, zero churn	Q8 Metal residency `84760f5a25a8`; mapped Q4_K owner `f732527d4daf`	0 auditable runtime tokens; the duplicate action has no attributed token cost
2	`01a03b37-7f33-7050-9fd7-0cf89564dd56`	5,953,198 / 316,454,144 / 815,441	35	2,354 / 72 / 47	Productive long Qwen native-profile work; zero replay/churn	Metal session counters `ca8993c3a913`; control profile `425a71bb5616`	0
3	`01a03c91-d142-7133-9fab-3260bc53bf21`	5,701,131 / 271,320,576 / 772,657	4	2,126 / 128 / 97	Productive Qwen closure and Mac-evidence work; zero replay/churn	Streamed Q4_K load ablation `4212905035a4`; slab rejection witness `3e5891679d4b`	0
4	`01a03b13-e4d3-7021-8364-b09832f30fe1`	5,346,077 / 222,344,448 / 690,023	41	1,713 / 52 / 4	Productive Qwen GDN work; zero replay/churn	Resident GDN primitive `dfa6ea83b927`; hybrid GDN route `254fb97a797f`	0
5	`01a0400d-a03f-7650-86a0-ca090f569732`	30,143,680 / 191,682,048 / 853,910	25	2,680 / 104 / 0	Cohort defect: productive research work, not Qwen execution	Research envelopes #9362–#9387, including `255b34e33060` and `b3d717c8412a`	0 runtime; remove 222,679,638 from the Qwen denominator
6	`01a0400c-33a1-7240-a36c-96180b009ad7`	15,337,949 / 167,598,976 / 411,740	7	2,059 / 169 / 0	Cohort defect: productive study-forge work, not Qwen execution	Disabled-discussions terminal state `1056b5c744c2`; delta policy `3752e60701b0`	0 runtime; remove 183,348,665 from the Qwen denominator
7	`01a03b37-4676-7c62-9e60-d5083d511db9`	4,425,036 / 166,486,528 / 491,948	22	1,301 / 28 / 33	Productive long Qwen runtime work; zero replay/churn	Measured Qwen3.8 cold-arm runner `6abd002e0dec`	0
8	`01a04401-0fe0-7c23-aa3c-ec8bbf47cbbc`	3,208,196 / 166,559,488 / 328,865	2	1,301 / 71 / 177	Productive Qwen batch/GDN wave; zero replay/churn	Qwen native batch repair `d7ce989e497d`; duplicate encoder removal `2f0fcd0f5353`	0
9	`01a04403-68e1-7173-b48c-b2ba235e7af8`	1,791,649 / 141,595,392 / 189,943	2	1,124 / 39 / 168	Cohort defect: productive code-slop work, not Qwen execution	Benchmark helper refactors `b1539ebc2050` and `e73e7065a868`	0 runtime; remove 143,576,984 from the Qwen denominator
10	`01a04401-b3ff-7fb2-831d-6f9509c59265`	2,404,652 / 137,517,568 / 289,958	11	1,112 / 64 / 87	Cohort defect: productive guard work, not Qwen execution	Guard-route witnesses `67a5dd2e2d18`, `534c481d422d`; provider fold `13d33528b9f6`	0 runtime; remove 140,212,178 from the Qwen denominator

The operational saving lower bound is zero: the evidence shows durable landed effects, the audit assigns zero tokens to its one repeated-failure event, and tool-error counts alone do not prove avoidable work. The exact cohort correction is different: excluding the four proven false positives removes 689,817,465 accounted tokens, 13.37% of the current selected denominator, without claiming that their useful execution should not have happened.

Defects filed

#9702 — snapshot output digest changes between two audits of the same staged bytes. Marker: fanout-trajectory-spine-00cac68f178d67cb-dogfood-self-run-output-drift.
#9703 — injected repository guidance can satisfy --user-contains and contaminate a topical cohort. Marker: fanout-trajectory-spine-00cac68f178d67cb-dogfood-self-run-topical-injected-context.

The smallest terminal path for #9697 is #9702 first, then repeat the two capture/replay cycles and append the corpus/output digests and zero-byte drift result here. #9703 is separately required before #8929 can treat this Qwen cohort as a defensible optimization denominator.

Post-fix retained rerun — 2026-08-28

The combined #9681, #9702, and #9703 base completed one capture and one replay over the same private corpus. Both passes reported verified=true, 152 files, 150 canonical sessions, and zero refused records. The corpus digest was 89e1a483b036bd734af91d7bf23d3a15b784a495cd04725626b120d3f9091542. Capture and replay were byte-identical:

JSONL SHA-256: 33bd683c18c7268b464e0a5909261d0a2b6aa8f978c8b7a23874f66e76cd02ca
Markdown SHA-256: 701280bdb98277523a9f1533cdf8a0a75796bc38c9bf78ce0b2f622cd81da961

The private snapshot used restricted permissions: 0700 for the snapshot directory and 0600 for its manifest. No raw corpus bytes are retained here.

The public-safe readback immediately preceding that capture selected the same 150 canonical sessions from 152 fragments. It accounted for 108,393,944 input, 2,618,842,368 cache-read, zero cache-create, and 8,310,616 output tokens: 2,735,546,928 total. The top ten held 1,820,248,183 tokens, or 66.5405577%. Continued writes to those live sessions advanced the later pinned capture to 24,011 exact usage records and 66.63808296%; this is expected sliding-corpus movement, not capture/replay drift. The two outputs from the pinned corpus remained byte-identical.

Corrected top-ten attribution

Turns counts Codex task-start events. Every current top-ten session had an operator-authored Qwen objective. All had zero mutation churn; only rank 1 had one repeated failed action. Tool-error counts alone do not prove avoidable work.

Rank	Session	Input / cache-read / output	Turns / tools / errors / churn / repeated	Classification	Public landed effect	Conservative counterfactual
1	`01a03b0a-0981-7ba3-ae80-7fd5402239cd`	5,400,182 / 373,675,776 / 458,884	2 / 3,002 / 44 / 0 / 1	Useful Qwen Metal residency/runtime work	`84760f5a25a8`, `f732527d4daf`	0 auditable tokens; the repeat has no token attribution
2	`01a03b37-7f33-7050-9fd7-0cf89564dd56`	5,953,198 / 316,454,144 / 815,441	35 / 2,354 / 72 / 0 / 0	Useful Qwen native profiling	`425a71bb5616`	0
3	`01a03c91-d142-7133-9fab-3260bc53bf21`	5,701,131 / 271,320,576 / 772,657	4 / 2,126 / 128 / 0 / 0	Useful Qwen Mac closure/evidence work	`4212905035a4`, `3e5891679d4b`	0
4	`01a03b13-e4d3-7021-8364-b09832f30fe1`	5,346,077 / 222,344,448 / 690,023	41 / 1,713 / 52 / 0 / 0	Useful Qwen GDN work	`dfa6ea83b927`, `254fb97a797f`	0
5	`01a03b37-4676-7c62-9e60-d5083d511db9`	4,425,036 / 166,486,528 / 491,948	22 / 1,301 / 28 / 0 / 0	Useful Qwen runtime/cold-arm work	`6abd002e0dec`	0
6	`01a04401-0fe0-7c23-aa3c-ec8bbf47cbbc`	3,208,196 / 166,559,488 / 328,865	2 / 1,301 / 71 / 0 / 0	Useful Qwen batch/GDN wave	`d7ce989e497d`, `2f0fcd0f5353`	0
7	`01a04678-03a4-7840-9f61-1d9a861d7662`	1,623,761 / 97,444,608 / 180,423	1 / 797 / 38 / 0 / 0	Useful Qwen issue/closure wave	`1f7aebc59e7c`, `8f070581cb4b`	0
8	`01a03c93-fe4a-7cf1-85ee-1b6f26129431`	2,628,481 / 67,498,496 / 320,549	22 / 513 / 11 / 0 / 0	Useful Mac/Qwen performance wave; some work remained unlanded	`3d90878da`, `4e3d17b92`, `bc88a3817`	0
9	`01a03eec-abcf-7eb0-9ae7-1c63f48ffb5b`	1,327,250 / 50,507,776 / 130,146	1 / 404 / 7 / 0 / 0	Useful but incomplete exact Qwen3.8 decode-parity campaign	None independently verified	0
10	`01a0451f-5952-7643-8108-90808bd16212`	9,693,054 / 38,136,576 / 324,465	8 / 1,325 / 83 / 0 / 0	Useful Qwen/GLM hardware and long-context economics	`02ed431a6604`, `49433e5b5f81`, `6667e4e2e66e`	0

The operational saving lower bound remains zero. The current top ten is productive work, not demonstrated replay, churn, or error-loop waste. The concentration percentage is therefore diagnostic, not an optimization target.

Concentration reconciliation

Historical 35.47%: 450 sessions accounted for 2,864,577,361 tokens. The rounded fraction implies approximately 1,016,065,590 top-ten tokens, but the exact numerator was not retained and two malformed duplicate-usage records blocked broad conclusions.
Pre-fix 43.41%: the 2026-08-27 live fallback counted 430 canonical sessions, 5,160,285,455 total tokens, and 2,240,250,063 top-ten tokens. It preceded #9703 and included injected-guidance matches. Four proven false positives alone contributed 689,817,465 tokens, or 13.37% of that denominator.
Post-fix 66.5405577%: #9703 reduced the selected cohort to 150 canonical sessions and removed all four known top-ten contaminants. Relative to the pre-fix run, the denominator fell 46.99% while the top-ten numerator fell 18.75%. The resulting increase is a purified topical denominator concentrated around legitimate long Qwen sessions, not increased waste.

Mitigations and remaining gaps

#8714 is closed by f6d64aa37: select user-authored topical content. #9703, closed by 8f070581cb, completes the later-discovered injected-guidance exclusion.
#8715 is closed by bd638fc69: report deterministic top-contributor concentration.
#8716 is closed by cab2b8c00, 257e3f3bcb, and 3e27e1e4b9: stop a third identical tool failure while allowing changed attempts and recovery.
#8700 is closed by d4cb83dcdb: rank tool-error families with repeated-failure and churn attribution.
#8702 is closed by 86db073803: parse Codex hook durations with typed missing and malformed states.
#8703 is closed by 1b982cc711: enforce canonical Qwen input-amplification budgets.
#8717 is closed by b7930a79cc: detect and bound mutation churn.
#8718 is closed by 26fe15c497: fail closed on unsupported evidence shapes.

No original #8929 gap remains open. The highest-confidence improvement is the combined measurement spine: #9681 (1f7aebc59) pins and verifies the private corpus, #9702 (1ae91f630) makes its output deterministic, and #9703 (8f070581cb) excludes injected guidance from topical admission. It corrects what the audit measures without inventing a performance saving or suppressing productive long work.