dctx1.3.1

STAB-01: byte stability on 0495f5f

View source · Edit on GitHub · Markdown · docs/research/stability-20261105/REPORT.md · 1,063 words · updated

On this page

Measured 2026-11-06 UTC on Linux from commit 0495f5f87717df5385209d306b92328a8cd96cf6. The directory retains its original evidence date. No engine changes were made.

Measurement Text JSON
Answers identical after unrelated edit 33/33 33/33
Answers identical over five separate invocations 33/33 33/33
Baseline post-edit common-prefix ratio, minimum / mean 1.0 / 1.0 1.0 / 1.0
Consecutive same-request transitions, five per case 165 165
Same-request M7, minimum / mean 1.0 / 1.0 1.0 / 1.0
Consecutive changing-question transitions 195 195
Changing-question M7, minimum / median / mean 0.007795889 / 0.107142857 / 0.214830257 0.008776458 / 0.067594433 / 0.182446794
Changing-question M7, total prefix bytes / total current bytes 0.061846978 0.079892615

All 405 CLI invocations exited 0, with empty stderr: 396 measurement invocations (66 cases × six runs), three target-resolution invocations, and six snapshot controls. All 33 default text headers were exactly # <place> <verb>; all 33 default JSON answers omitted snapshot, generation and freshness.age_s. --meta --json controls confirmed that each corpus's snapshot changed after the edit.

Protocol and pins#

The original producer is phase_query.py, with byte comparisons in aggregate.py, section 7. It runs five warm queries, appends a comment to an unrelated file in another directory, reruns queries, and restores the file. Its original evidence remains at data/stability.csv.

The original /path/to/workdir/dctx-vs-trail host tree is absent. The repository contains the scripts and historical results, but no original seven-corpus checkout set or the host's scripts/corpora.json. Checked alternate bench3 and perf-20261101 corpora paths were also absent. This run uses the documented fallback: three corpora extracted exclusively from tracked files at the pinned current commit. No fetch or network access was used. The new denominator differs from the historical 0/63; Trail-Code's historical 59/70 was not rerun.

Corpus Exact tracked prefix extracted Files Symbol / file Unrelated file edited
parser-mixed crates/dctx-map-parse/tests/fixtures/ 17 helper, rust/src/support.rs ts/module.ts
lexer-go crates/dctx-lex-go/ 19 Add, tests/fixtures/basic.go src/lib.rs
map-query crates/dctx-map-query/src/ 30 render_answer_at, adapters/output.rs domain/ranking.rs

Each prefix is copied in full using git ls-tree -r --name-only <pin> -- <prefix> and git show <pin>:<path>, retaining paths relative to the prefix. Every file's size and SHA-256, and the CLI binary SHA-256, are in corpora.json. Work copies and isolated cache/config directories live in a disposable /tmp directory. The live worktree's corpus files are never edited.

The original applicable map query shapes are retained: find --limit 10, inbound neighbors --edge calls at depths 1 and 2, outbound neighbors, map --depth 2 --limit 40, read, and explain. Each now runs in both formats. The current bounded skeleton, grep, hubs, and path verbs add four shapes per corpus, again in both formats: 11 × 2 × 3 = 66 cases. Stable target IDs are resolved by find --json --limit 40, matching exact path and symbol. path tests a file's containment edge to that symbol. grep is confined to the target file.

Default budgets and default output are used, without --meta in the measured answers. There are five warm sweeps, separated by 1.05 seconds, then a sixth sweep after appending exactly \n// STAB-01 unrelated comment\n to the named unrelated file. The comment is outside declarations and contains none of the queried names. It changes snapshot identity without changing the selected declarations, calls, search hits or hub counts. Original bytes are restored in a finally block and verified. Configuration and caches are isolated using XDG_CONFIG_HOME, XDG_CACHE_HOME and DCTX_CACHE_DIR; inherited DCTX_* overrides are removed.

Every JSON case has records except outbound neighbors for the two leaf functions helper and Add. Their empty answers are included. All three inbound-neighbor cases have records, all paths are found, and render_answer_at has nine returned outbound neighbors plus eleven more in the bounded page. No cursor pages were followed. Whole-graph export has no unrelated graph input and is outside this bounded-answer protocol. The historical legacy snapshot overview, Trail-only grep and no-refresh shapes are outside the map-answer comparison.

M7 and acceptance limits#

For each transition, M7 = longest common byte prefix of previous and current stdout / current stdout byte length. Both block lengths and the prefix length are stored in turn-prefixes.csv; there are no empty blocks. No whitespace, JSON fields, paths, timestamps or newlines are normalized.

The same-request trace is each case's five warm answers followed by its post-edit answer: five consecutive transitions per case, 330 total. The additional changing-question trace uses actual invocation order, separately per corpus and format: the 11 query shapes in the order above, repeated over six sweeps; 65 transitions per stream, 390 total. Text and JSON are separate context-block streams. These are recorded tool-answer traces, with one answer treated as one injected block; there is no provider cache-hit or task-acceptance measurement.

The local unrelated-edit and five-run JSON gates pass. The low changing-question ratios show why this result cannot establish §12.1's T0–T3 invariant or its recorded tier-context M7 ≥ 0.9 gate. Tier layering remains target. No cross-platform, thread-count, changed-answer-content, or full original benchmark gate is claimed. No unrelated-edit instability was observed, so there is no instability reproducer.

The three text read pages actually include their selected source bodies. Text explain still consists of header and id/kind/path, so §11's richer text-page contract remains a target. Other text fields, JSON compaction and provenance-trailer contracts were not promoted to landed by this measurement.

Exact commands and exits#

Run from /path/to/workdir/codex/STAB-01R:

export PATH=$HOME/.cargo/bin:$PATH
export CARGO_TARGET_DIR=$PWD/.target-stab
cargo --offline build --locked -p dctx-cli > /tmp/stab-01-build.log 2>&1

Exit 101: configured sccache failed with Operation not permitted, before compilation. A retry after unset RUSTC_WRAPPER, with output redirected to /tmp/stab-01-build-no-wrapper.log, also exited 101: Cargo's user config still specified build.rustc-wrapper = "sccache".

export PATH=$HOME/.cargo/bin:$PATH
export CARGO_TARGET_DIR=$PWD/.target-stab
export RUSTC_WRAPPER=
cargo --offline build --locked -p dctx-cli > /tmp/stab-01-build-direct.log 2>&1
python3 docs/research/stability-20261105/measure.py > /tmp/stab-01-measure.log 2>&1
python3 docs/research/stability-20261105/verify.py
git diff --check

Exits: build 0; measurement 0; independent raw-data audit 0; git diff --check 0. Build logs and measurement.txt are preserved here. commands.md lists every query shape and setup/control command; runs.jsonl records all 405 exact argv arrays, shell-quoted commands, working directories, environment overrides, exits, full stdout/stderr, byte counts and SHA-256 hashes. measure.py records a fresh run only when no runs.jsonl exists; use a fresh copy of the evidence directory to repeat it.

stability.csv is the new measurement. The adjacent stability-original.csv is a byte-for-byte copy of the historical CSV, SHA-256 46b54db2fb469dc4df70eaa0bf22b458a8089f437cdee3294efe071314280266. Both files are retained to distinguish the new measurement from the unchanged historical source.