rea/docs/specs/agent-environments/write-rules/verification.md

Verification

Executed on macOS with Node v22.23.1, one change at a time on the state-targets branch, over the 50 published capture-built environments (each app's latest capture). The served smoke rebuilds each manifest with the tree under test and applies every served write tool. The serving check answers every recorded read made before the capture's first write through state serving and compares it with the recording; it reads no calls from the five browser-scenario captures. The session replay answers every recorded call in order, writes included, through state serving, reading records with REA's own capture parser, and compares the reads; it covers all 50 captures.

Gates

• npm run -s check:fast, npm run -s format:check, npx knip, npm run -s scan:todos, npm run -s docs:check, npm run -s verify:test-discovery and npm run -s mcp-catalog:check: passed.
• npm run jscpd: 3.57% duplicated lines over 4235 files.
• npx vitest run --project domain --project services: 9545 passed, 6 pre-existing skips. The 10 failures (JavaScript Map endpoints, a returned invocation's alias graph, Frida cleanup) fail the same way on main.
• npm run -s test:prepare, then npm run -s test:focused -- the environment serve and export CLI boundaries, the serve and served-writes MCP boundaries, the analysis-view CLI and MCP boundaries, the compact input schema boundary and tool schema validity: 8 files, 91 passed.

Measurements

Step	Served	No state target	State-backed read views	Reads before first write	Session reads
state targets	62	59	1,151	2,085 of 2,155	2,764 of 3,022
blame and constant guard	63	59	1,151	2,085 of 2,155	2,764 of 3,022
update rules	68	59	1,151	2,085 of 2,155	2,763 of 3,022
view order	69	59	1,151	2,085 of 2,155	2,763 of 3,022
bulk writes and item actions	71	48	1,151	2,085 of 2,155	2,763 of 3,022
per-response members	71	46	1,153	2,087 of 2,162	2,766 of 3,031
nested counters	71	46	1,153	2,087 of 2,162	2,766 of 3,031
created nested clocks	71	46	1,153	2,087 of 2,162	2,766 of 3,031
unshown writes refused	69	46	1,153	2,087 of 2,162	2,766 of 3,031
catalog values, atomic bulk	69	46	1,153	2,087 of 2,162	2,766 of 3,031
review fixes	69	46	1,153	2,087 of 2,162	2,766 of 3,031
scoped envelopes, bulk rows	69	48	1,151	2,085 of 2,155	2,763 of 3,022

All 187 write tools are counted at every step, and the served smoke applies every served write tool.

• Scoped envelopes, bulk rows: a shared envelope masks only members minted on every path, so Vault's identity entity and group item views, which matched one of four and one of three recorded reads from the state, replay their recordings, and Vault's entity update and delete lose their state target. Serving bulk responses that list each written entity changes no capture's verdict.

• Review fixes: literal bulk ids, per-entity bulk values, pinned bulk delete members, comparison ids per collection, echoed creation clocks, required held clocks, nested clocks checked against every call, learned sorts kept open to contrary reads and limited to the read's own entities, one query check per recorded call, guards relaxed only by successful calls with the exemplar's query, and slash-named counters. No tool's verdict and no session read changed; each fix has a test that fails without it.

• Newly served: Ghost's post create and update, Metabase's card create and update, n8n's workflow update, Navidrome's playlist create, and the NocoDB and Mailpit bulk deletes.

• No longer served: Gitea's issue update. Its create, update and read share one second, so the evidence cannot show the update stamping updated_at, and serving it left the clock as it was.

• Refused that were briefly served on this branch: EspoCRM's task update, for the same reason, with modifiedById absent from its create response.

• Session reads: one read is lost at Ghost #74, a 404 right after the served create whose delete the recording lacks; Vault #80 answers 404 after an unserved create, as a served read on an entity only a replayed write made does.

• Ghost's create and update accept only the recorded html, and Metabase's create only the recorded collection, display, type and query; the served catalog shows those values.

Every change had an independent review that rebuilt it and tried to refute it; the corrections they found are part of the change. A final review of the whole branch found one wrongly served update (EspoCRM) and the hidden pinned values, both fixed.

Cost

Over the 50 captures, one Node v24.22.1 process per sweep builds the environment-state and environment-manifest views of every capture (macOS arm64, warm medians of three runs; the cold first run is within 2% of them):

Tree	Wall	CPU	Peak RSS	Environment-state view	Its results
main d87e73f2	2.61 s	3.41 s	516 MiB	1,791 ms	3.69 MB
state targets (#6385)	2.96 s	3.82 s	540 MiB	2,093 ms	3.84 MB
this branch	3.06 s	4.07 s	601 MiB	2,224 ms	3.85 MB

The write rules add 0.10 s wall, 0.25 s CPU and 61 MiB of peak memory to the sweep while serving 7 more write tools. Agent tokens were not measured; the views' results grow by 0.3%.

Not included

• Templating literal ids by value: an opt-in rule served Grafana's folder delete and Vault's policy delete, but both targets came from unserved writes, so serving them answered their own recorded calls wrongly.
• Completing a created entity from a later read: it served Bookstack's page create wrongly.
• Soft deletes, appended relation rows, rendered markup, reverse embeds, identity aliases and resource names.

Linux integration and review corrections, 2026-10-24

The following is separate from the historical 50-capture measurements above. Node 24.22.1, baseline 3ce311757, reviewed implementation 352c0278d:

• 1,236 selected source tests passed. Typecheck, built lint/module boundaries, format, knip, metadata, architecture guards and documentation checks passed.
• Seven review regression checks failed before correction and passed after it. Earlier structured update values now retain canonical equality digests within the history budget. A shared envelope masks only members independently proved minted on every concrete path. Empty advertised literals are widened while execution enforces their values, and pinned queries are optional so their execution defaults can run. Numeric and string bulk id arrays stay distinct.
• Nineteen compiled SDK/schema checks passed, including real HTTP MCP discovery, empty fixed values, omitted empty and nonempty queries, and unchanged state after refused writes. Live Gemini API execution was not performed.
• Four public datasets were downloaded at immutable Hugging Face revisions: NocoDB d0f7e040, Ghost 6d5b2d1c, Metabase fa2aa0fd, and Mailpit 416b29be, under earakely-scale/*Env-REA. Capture hashes, full calls, independent recorded JSON comparisons, runtime identities and resource receipts were retained locally. These were recorded real applications; the original applications were not re-recorded in this verification.

Capture	Eligible write tools before / after	Initial reads matching	Session reads matching
NocoDB	0 / 3	20 / 20	80 / 92
Ghost	1 / 1	19 / 20	72 / 88
Metabase	1 / 2	19 / 22	99 / 122
Mailpit	0 / 1	51 / 52	51 / 56

Every previously matching read remains matching, record by record. Real CLI listeners and the MCP SDK reproduce the shared application's outcomes. The feature session applies six NocoDB mutations, one Ghost deletion and one Metabase update; the baseline applies only the Ghost deletion. A state answer with view missing is not counted as an applied write. Mailpit exchange 65 is refused because its target was introduced only by an unserved create. Eligibility does not establish that every recorded write can run on the initial state, and the remaining read mismatches are preserved coverage gaps.

Three alternating fresh producer sweeps over all four captures had median wall/CPU/cgroup peak of 5.360 s / 4.263 s / 327.3 MiB before and 4.960 s / 4.357 s / 323.1 MiB after. Ranges overlap. One complete SDK session per app/revision, 403 versus 413 calls, took 20.905 s / 19.039 s / 508.7 MiB before and 23.426 s / 19.396 s / 462.8 MiB after; these are single observations with different write coverage, not a speed or memory improvement claim.

Catalog estimates (cl100k_base) were 8,630 / 9,408 tokens for NocoDB, 11,355 / 11,355 for Ghost, 12,162 / 12,648 for Metabase and 2,955 / 3,138 for Mailpit. Arguments, instructions and text replies were also counted in retained receipts; actual model usage, prompts and client rendering were not measured. No token saving is claimed.

Processes were fresh, OS caches warm or uncontrolled, on a shared Linux host. Each task used one CPU, a 3 GiB early stop / 4 GiB hard limit and zero swap. All eight recorded-data task groups exited zero, verified payload cleanup, and observed no OOM or limit events. Payload-exit counters were verified; final whole-scope counters were unavailable. Downloads, builds and runtime identity preparation are excluded from task measurements. The current four-capture verification does not requalify the earlier 50-capture totals.

Bulk-response follow-up qualification

The follow-up at f4bc513dd classifies every returned bulk element and checks all recorded calls with the shared serving renderer. Nested parsed-body leases retain their enclosing scratch charge. Literal JSON IDs never pass through a synthetic URL. The author's subsequent bulk regressions and measurement note were retained when integrating a319ba1af; their historical census does not measure this follow-up.

A fresh four-capture producer sweep and real CLI/MCP replay retained every previously matching read (109/114 initial, 302/358 session). NocoDB, Ghost and Metabase plans and complete SDK calls equal 9942863b2. Mailpit's text/plain ok deletion acknowledgement is not parsed by the state JSON owner, so its contents remain unknown and its write tool is now withheld as body-unusable. The prior session already refused this deletion for a missing target; no previously applied mutation was lost. Plain-text acknowledgement support remains unverified rather than treating unknown contents as empty. The three other captures retain the same eight applied mutations.

This qualification-only sweep took 5.857 s wall / 5.413 s CPU / 329.9 MiB peak; the complete SDK replay took 23.264 s / 22.298 s / 474.0 MiB. These single observations are not pooled with the earlier performance cohort. Payload cleanup and payload-exit resource counters were verified, with no OOM/limit events; final whole-scope counters and actual model token usage remain unmeasured.

Review round on main 3d84215b, 2026-10-24

Three automated review findings were fixed in d0198389f and 7c25ad4dc: a create compares the other calls' request bodies one at a time inside their parsing reservations, an echoed reference to a nested object stays pinned when the plan cannot rebuild the object, and a sorted read view a later read contradicts is recorded and loses the confirmations it gave. On macOS with Node v24.22.1 at 05566419e, the static gates passed, npx vitest run --project domain --project services passed (10,164, 6 skipped; the reference-lifetimes file whose worker crashes on main was excluded) and the focused build, export, serve and analysis-view boundaries passed (10 files, 93). Over the same 50 captures every write plan, session replay and served read equals 0895c89bc's: 187 write tools with 68 served, 68 of 68 smoke calls, 1,151 views, and 2,085 of 2,155 initial and 2,763 of 3,022 session reads matching.