Skip to content

gzkit — Big-Picture Assessment

Date: 2026-09-19 · Report: baseline (first in the big-picture series) Evidence cutoff: 2026-09-19, main at 85c857263, working tree clean Comparison points: the project's own State of gzkit reckoning (2026-06-20) and the campaign re-cut of 2026-08-17. No prior report in this series exists.

The assessment

Eight months in, gzkit has built the thing it set out to build, and has not yet shown that it works for anyone but itself.

The verification substrate is real and it bites. 10,539 unit tests pass in 96 seconds. All 62 registered quality checks pass today. The specific failure the project's June reckoning called its worst — two integrity gates that warned instead of blocking, with tests that certified the warning — is repaired; I confirmed it by resolving the live enforcement posture rather than reading the claim, and both now return BLOCK_GHI_FIX, fail-closed. Twice during this assessment a project hook refused one of my own shell commands, named the binding rule, and printed the corrected command. That is enforcement observed rather than asserted, and it is rarer than it sounds.

Against that, three numbers describe the last seven weeks better than any narrative. The governing campaign has not checked a box since 2026-08-08. The open issue queue has gone 9 → 32 → 55 since mid-August while issues close at a median age of 3.7 hours. And 453 commits sit unreleased since v0.34.7, of which 197 are fix and 9 are feat.

Those are not the marks of a stalled project. They are the marks of a project in a defect-absorption equilibrium: producing self-knowledge and repairs faster than it produces governed capability, with a queue that grows despite being drained within hours. The constraint is not repair latency. It is that self-inspection generates real findings faster than they can be discharged, and gzkit has no consumer but its own construction to arbitrate which findings matter.

The operator reached this conclusion first, on 2026-09-15, and the campaign records it verbatim:

"the rate at which new GHIs grow in gzkit far outstretches getting any feature work done. this is dismaying."

This report's contribution is not that observation. It is the measurement behind it, and the argument that the dynamic is structural rather than a matter of diligence — more diligence makes it worse, because every inspection pass finds real defects. The campaign already names the escape: an external forcing function. That remedy is currently sequenced last, and costs about a quarter.

What is being built, and why it would be worth something

gzkit is a governance meta-harness. It does not run agents; it wraps the harnesses that do — Claude Code, Codex, Copilot — with deterministic governance surfaces, evidence capture, gate enforcement, and human attestation.

The README frames the problem as a mismatch between context-bound agents and unscalable human judgment. The PRD frames it more sharply, and more usefully, as a problem about the human:

"AI-assisted development creates a drift problem: humans approve work they didn't verify, attest to artifacts they didn't read, and gradually become rubber stamps for agent-generated output."

That is the value proposition, and it is a real one. It is also unusually hard to demonstrate, because the thing being prevented is a slow decay in a person's engagement with their own codebase — not something a sprint can A/B test. The project's answer is to make the decay structurally impossible rather than merely discouraged: every completion claim witnessed by a ledger event, every "enforced" claim backed by a live negative control that constructs a real violation and proves the gate fails on it, and no agent permitted to attest on a human's behalf.

Each mechanism carries its own causal claim:

Mechanism The claim it makes
Append-only ledger (L2) Runtime truth must be witnessed, not asserted — frontmatter is a lazy mirror, the ledger wins
ADR → OBPI → REQ Intent decomposed to a unit small enough that completion is provable, with a declared path allowlist so scope creep is mechanically visible
Five gates Verification staged and non-skippable, because a single end-of-work check is exactly where rubber-stamping happens
Human attestation At closeout the agent becomes a passive presenter — paths and commands only, never outcomes or summaries
Corpus → rendition A hand-edited control surface is the same failure class as a hand-written ledger row, so the per-turn agent surface is deterministic playback of an append-only corpus
Airlock An agent cannot hold a complex system resident, so hand it a bounded seam-set on entry and make it account for exactly that set on exit

The architecture is coherent; each layer answers a question the layer below cannot. That coherence is the asset, and it is why the project deserves continued attention.

What is actually delivered

This is a real, released product, which is easy to lose sight of from inside the governance record.

  • 56 releases on PyPI as py-gzkit, first 2026-03-22, latest 0.34.7 on 2026-08-29, published through GitHub trusted publishing on tag. 574 downloads last month.
  • A frozen distribution contract: 240 canonical surfaces the wheel must deliver to a fresh adopter, enforced in both directions.
  • A substantial verification surface: 97 registered validator scopes, 62 checks in gz check, 467 BDD scenarios, 5,359 @covers annotations binding tests to 1,971 distinct requirements, 65.8% REQ coverage (1,794 / 2,728), 88% line coverage.
  • A genuinely queryable governance record: 16,970 ledger events across 66 event types, every one written through a governed verb.
  • Five feature ADRs closed and released between 2026-06-29 and 2026-07-31 (0.30.0 through 0.34.0), 27 OBPIs attested complete.

The most adopter-relevant work of the whole period is release v0.34.6 (2026-08-29): gzkit discovering that its own first hour was broken in nine distinct ways — a scaffold failing its own surface gate with 58 missing files, rules scoped to gzkit's directory layout so an adopter's first audit reported findings about gzkit rather than themselves, and the maintainer's personal name in 92 --help examples. Finding and fixing that is exactly the right work. That it was found in month eight by inspection, rather than by an adopter complaining, is the point this report keeps returning to.

The quarter: one thing repaired, one thing widened

The June 2026 reckoning — commissioned with the words "gzkit is in utter shambles, totally off the rails" — named three headline problems. Their status today, measured rather than relayed:

June 2026 finding Today Verdict
Two corpus→rendition integrity gates inert by default; tests certify the inertness Both resolve BLOCK_GHI_FIX, fail_closed=True Repaired (verified live)
A hollow "antibody" audit passed Gate 5 on a green self-check Enforcement-claim meta-validator built; a later hand audit found 32 of 47 claims still under-scoped or trivial while gz check reported 47/47 Partly repaired, honestly disclosed
Measurable accretion: 33 oversized modules, 70 validate scopes, 10,913 ledger events, 116 handoffs, a 1,589-line campaign 66 oversized modules, 97 scopes, 16,970 events, 740 handoffs, a 1,923-line campaign Widened, roughly 2×

The first row is a genuine win and should be read as one. The third is the concern, and it needs care, because "the codebase grew" is not by itself a finding. What makes it one is that reduction is a declared pre-1.0 Movement — the project ruled in July that accretion was what blocked 1.0 and moved reduction ahead of post-1.0 — and the measured surface has roughly doubled since.

One instance shows the shape of the problem rather than a moral about size. The module-size debt list reads empty today. It was emptied on 2026-08-26 not by shrinking modules but by moving the block threshold from the corpus p95 (1,032 SLOC) to p99 (3,144), after which every remaining entry fell under the new band and surrendered under the shrink-only rule. Eight modules still sit above the retired band. Nothing is in violation, the move was operator-ratified and recorded, and the ratchet worked as designed. But the number that says "debt: zero" now means something different from what it meant in July, and only the dated history inside the JSON says so.

What the architecture reveals

Three readings, in descending order of confidence.

1. The work migrated to the cheapest governed channel

Start with the throughput numbers, because they are widely misread and the correction matters.

OBPI completions per month fell from 306 in March to 4 in September. Over the same period commit volume hit its all-time high (727 in August). September's commit mix is 185 fix and 170 chore against 7 feat. ADR-level gate checks and attestations both stop on 2026-07-31.

The naive reading — "delivery collapsed" — is wrong, and the arithmetic says so. During the active ADR-0.35.0 window (2026-08-21 → 2026-09-12) the project completed 7 OBPIs in 22 days: 3.7 calendar days each, against the campaign's own planning assumption of 3.8. When the pipeline runs, it runs at precisely the rate the plan assumes. The ADR-level gates stopped firing because no ADR has reached closeout since 0.34.0, not because gates broke.

So the finding is duty cycle, not rate. There was a 20-day gap with zero completions (2026-07-31 → 2026-08-21) and a 10-day gap inside the active window. Meanwhile 1,071 commits reference a GHI, 481 of them in the last two months, and 1,001 of the project's 1,061 issues are closed at a median age of 0.154 days, with 71.2% closed the same day they are opened.

The campaign's own 2026-09-15 amendment supplies the causal step, and it is the right one:

"That amendment raised the family-closure box because it is 'drawable without the operator'. Every session could draw it, and GHI repair was the cheapest arm to draw."

Doctrine authorizes GHI-routed direct repair without an ADR, without operator initiation, without gates, and without attestation. It is the cheapest governed channel; the work is real; it absorbs the available session time. The structural consequence is the part worth holding onto: the channel that absorbs most of the work is the one channel that does not exercise the ADR → OBPI → gate → attestation spine the product exists to sell. The spine gets less exercise as the project matures — the opposite of what a system needs before asking anyone else to adopt it.

Two further measurements make this an equilibrium rather than a phase. First, the open queue is growing despite the 3.7-hour median close: 9 in mid-August, 32 on 2026-09-02, 55 today, of which 93% are labelled defect. Second, the project's own insight store shows self-observed defects quadrupling — 17 in July, 26 in August, 67 in the first 19 days of September — inside a record 208 entries for the month. More inspection produces more real findings, which produces more cheapest-channel repair, which consumes the sessions that would otherwise run a pipeline.

That is a system-dynamics problem, not a diligence problem. Diligence is what drives it.

2. The witnesses are strong artifact-to-artifact and absent prose-to-code

The project's dominant recurring defect has a name and a definition: "Layer X declares a discipline that Layer X does not mechanically enforce." Its two deepest recurrence chains in the entire corpus — depth 12 and depth 7 — are both this family, and in a 2026-09-02 measurement 19 of 32 open issues belonged to it. Instances close same-session; the class keeps producing.

The reason is stated by the project itself, in the most load-bearing sentence in the campaign:

"gzkit's witnesses compare artifact↔artifact and claim↔receipt exhaustively and doctrine↔code not at all… A general doctrine↔code witness is NOT built and is not scheduled here. The honest statement of the residual: gzkit does not today know how to mechanize 'this prose binds that code.'"

This is why the family is generative rather than finite. Every new doctrine document is a new unwitnessed claim, and doctrine is the surface growing fastest: 22.3 MB of markdown under docs/ against 6.3 MB of Python under src/. The recursion is visible in the governance scorecard — of 71 rows scored "Mechanical," 64 are frozen on a shrink-ratchet as unwitnessed, because the claim that a row is mechanically enforced was itself unwitnessed doctrine until a ruling in August forced it into a register. The instrument that measures mechanism had no mechanism. To the project's credit that is now tracked, shrink-only debt with a number on it rather than an assumption.

The project draws the right comparison itself: the one family measurably closing — layer drift, 10 → 7 → 6 → 1 across May to August — closed because it got three things: a doctrine, a regenerator, and a fail-close. The doctrine family has the first two at best.

3. A well-designed mechanism can be vacuously open

The airlock is the clearest single instance, and it is instructive because the design is sound and the input set is empty. Across 71 transits the ledger records 53 proceed decisions, and every one of the 53 computed an empty seam-map. All 18 hold decisions have non-empty maps (2 to 21 seams). The root cause is prosaic: no caller passes parent_invariants, and a leaf OBPI has no transitive dependents, so both edge sets return empty — and a fail-closed decision over an empty input is vacuously open.

The trajectory here is genuinely improving and deserves credit. September transits tripled to 34, the entry/exit accounting gap narrowed sharply (July 23 in / 5 out; September 34 in / 22 out), and the bite rate rose from 13% to 25%. The repair is designed: ADR-0.37.0 exists precisely for this, and its first proposed fix was withdrawn on its own measurement when it turned out to convert an empty seam-map into a constant one. That withdrawal is a good sign about the project's epistemics. But ADR-0.37.0 was authored 36 days ago and has 6 OBPIs at 0% REQ coverage. Nothing is built.

The same holds one ADR over: ADR-0.36.0, authored 41 days ago, carries 9 fully-drafted briefs and has landed nothing.

The central uncertainty

Everything above is second-order next to one question: does this work for anyone who is not building it?

The evidence is thin, and the project knows it.

  • Flight tests: 0 of 6 sorties flown. The manifest, templates, and skill exist; there is no flight log, no sortie record, no ledger event. The project states this itself in two separate documents.
  • rhea, the named proving ground: real, bootstrapped from the published wheel (py-gzkit>=0.34.1), 13 working sessions, 12 numbered defects filed back — then stopped. 20 ledger events total, 0 OBPIs ever completed, no attestation ever recorded, last commit 2026-07-26. Its findings were never routed back into gzkit's governance record.
  • One genuine external adopter event: a private repository filed GHI #607 from inside itself in June running gz v0.28.1 — an opinionated Pydantic rule leaking across the adopter boundary and failing their pre-existing dataclasses. Fixed. That repository has been dormant since July.
  • Community signal: 9 stars, 0 forks, and all 1,056 issues ever opened have exactly one author.
  • Evaluations: numerous, recent, methodologically careful, and including several that return honest negatives which are then acted on. But every subject is gzkit itself, a synthetic fixture, or a fresh agent reading gzkit's own instructions. No evaluation measures a real operator outcome in an adopting project.

The sequencing is the crux. External validation is a declared 1.0 gate, carried by ADR-0.38.0, which by operator ruling cannot be authored until 0.35.0, 0.36.0, and 0.37.0 land. That is 21 remaining OBPIs (6 + 9 + 6). At the observed active rate of 3.7 days each that is roughly 78 days of pipeline time; at the observed 60-day calendar rate of 6.0 days each, roughly 126 calendar days. Either way the first external signal arrives no earlier than Q1 2027 — consistent with the declared ≈2027-04 target. The plan is not lying to itself about the date.

But it means roughly 21 more governed increments will be built on internal judgment before anything outside the project pushes back. The campaign states this cost on the record: the forcing function "certifies rather than informs." That is a defensible choice, deliberately made and recorded. It is also the single largest risk in the portfolio, because every interim design decision goes unfalsified by contact with a different substrate — and the one time gzkit did meet a different substrate, the finding was a category of mistake that five months of self-inspection had not surfaced.

There is a second-order cost worth naming. The campaign is now a 1,923-line document in its fifth edition, carrying 28 amendments of which only four record a box being checked; the other 24 are sequencing rulings, count corrections, scope expansions, and corrections of the plan's own prose. Of the ten amendments written since the current edition was created on 2026-08-16, none has checked a box. The plan says this about itself, at line 49: "two amendments were spent explaining that split rather than removing it." Steering has become a work item competing with the work.

What would move this assessment

These are observations that would change the reading, in either direction — not recommendations. Work selection is the operator's.

Would strengthen it materially: any completed governed unit of work in a repository that is not gzkit — one OBPI, one attestation, one flight-test sortie. A single such event resolves more uncertainty than another quarter of self-inspection. Second: ADR-0.37.0 landing, followed by a transit computing a non-empty seam-map on a real OBPI, which would convert the airlock from designed to operating.

Would weaken it: a further quarter in which the accretion figures grow while Movement C stays open; the open queue holding above 55 through another measurement; or ADR-0.36.0 and ADR-0.37.0 still at zero landed OBPIs at 90 days from authoring.

Untested either way: whether the covenant's central value claim — that this prevents an operator's slow slide into rubber-stamping — is true. No instrument in the repository measures it, and eight months of one operator's experience is a sample of one who is also the author. That is not a criticism; it is the hardest measurement in the project and nobody has solved it. But it should not be quietly counted as demonstrated.

In summary

The craft is not in doubt. Over eight months and 3,756 commits, one operator has built a governance system with a verification substrate most funded teams never achieve, and — more unusually — a standing habit of commissioning honest diagnostics against itself and publishing the unflattering results. The June reckoning, the enforcement-claim audit, the August architecture review, the three-pillars evaluations that returned null and were acted on as null: these are the behaviors of a project that wants to be right more than it wants to look right. That is the asset that most justifies continued investment, and it is not common.

The risk is equally clear, and the project named it first. A self-governing system inspecting itself finds real defects forever, which is what makes the loop seductive rather than obviously wasteful. The machine is now excellent at proving things about itself and has never been asked to prove anything about anyone else. The escape is already identified, already designed, and sequenced last.

Evidence appendix

All figures observed 2026-09-19 against main at 85c857263 unless dated otherwise. Where a document states a number, the JSON or command authority was consulted instead and is named.

Commands run for this report

Observation Source
10,539 unit tests pass, 4 skipped, 96.2s uv run gz test
All 62 registered checks pass, exit 0 uv run gz check
REQ coverage 1,794 / 2,728 = 65.8% uv run gz covers
ADR-0.35.0 at 7/13 OBPIs, closeout BLOCKED on 6 briefs uv run gz adr status ADR-0.35.0-canon-entry-corpus-landing
Rendition gates fail-closed (BLOCK_GHI_FIX) gzkit.mx.checkpoint.resolve on rendition-freshness, rendition-floor-coherence, invariant-coherence
AGENTS.md 19,872 B against a 50,000 B budget uv run gz validate --instructions-files-budget; authority data/instructions_files_budget.json
56 PyPI releases, latest 0.34.7 2026-08-29, 574 downloads/month pypi.org and pypistats.org APIs
55 open issues; 1,001 closed, median age 0.154 d, 71.2% same-day gh issue list
453 commits since v0.34.7 — 197 fix, 183 chore, 52 docs, 9 feat git log v0.34.7..HEAD

Ledger-derived (.gzkit/ledger.jsonl, 16,970 events, 2026-01-22 → 2026-09-20)

  • OBPI completions are recorded as obpi_receipt_emitted with receipt_event: "completed" — 659 total. By month: Mar 306, Apr 109, May 134, Jun 74, Jul 28, Aug 3, Sep 4.
  • Rate windows: last 30 d, 7 completions (4.3 d each); last 60 d, 10 (6.0 d); last 90 d, 59 (1.5 d). Active ADR-0.35.0 window 2026-08-21 → 2026-09-12: 7 completions in 22 days = 3.7 d each. Longest gap 20 days (2026-07-31 → 2026-08-21).
  • gate_checked (943) and attested (103) both last fired 2026-07-31, at ADR-0.34.0 closeout.
  • Airlock: 71 airlock_in, 29 airlock_out; decisions 53 proceed / 18 hold. All 53 proceeds carry an empty unaccounted set; holds range 2–21 seams. Entries by month: Jul 23, Aug 14, Sep 34. Exits: Jul 5, Aug 2, Sep 22.
  • obpi_parked 377 against obpi_unparked 6.
  • artifact_edited 5,232 — the largest single event class.

Repository scale

  • 3,756 commits since 2026-01. By month: Mar 523, Apr 516, May 516, Jun 531, Jul 448, Aug 727, Sep 426.
  • September commit mix: 185 fix, 170 chore, 52 docs, 7 feat. 1,071 commits reference a GHI; 481 in Aug–Sep.
  • src/gzkit: 512 modules, 155,638 lines / 6.3 MB. Tests: 692 files, 226,823 lines / 9.6 MB. docs/: 1,761 markdown files / 22.3 MB. .gzkit/: 36 MB (15 MB ledger, 12 MB handoffs).
  • 66 modules over 600 lines (largest commands/content/unown.py, 2,910). 73 skills, 26 rules, 74 feature files, 467 BDD scenarios, 740 handoffs, 972 booked rulings, 367 corpus entries, 205 pool ADRs.
  • Feature ADR ages and landed counts: 0.35.0 authored 2026-07-21 (60 days, 7/13); 0.36.0 authored 2026-08-09 (41 days, 0/9); 0.37.0 authored 2026-08-14 (36 days, 0/6). Last feature ADR to reach Validated: 0.34.0, 2026-07-31.
  • Campaign: 1,923 lines, fifth edition, 7 of 25 boxes checked — all 7 between 2026-07-18 and 2026-08-08. 28 amendments, 4 recording work landing.
  • Insights store: 784 entries (510 improvement, 158 defect, 84 discovery, 32 defect-resolution). By month: May 128, Jun 147, Jul 131, Aug 150, Sep 208 in 19 days. Defects by month: Jul 17, Aug 26, Sep 67.

Document sources

  • docs/governance/state-of-gzkit-2026-06-20.md — the June reckoning; source of the three compared findings and the June baselines.
  • docs/governance/build-to-1.0-campaign-2026-08-16.md — governing plan. §1 identity and the self-consumption diagnosis; §5 the nine 1.0 gates and the ≈2027-04 target; the doctrine-declared-without-mechanism box and the doctrine↔code residual; the 2026-09-15 amendment quoted above at line 489.
  • docs/design/prd/PRD-GZKIT-1.0.0.md — the rubber-stamp framing; status Draft.
  • README.md — product framing and the meta-harness boundary.
  • docs/governance/enforcement-claim-nc-audit-2026-07-18.md — 32 of 47 enforcement claims under-scoped or trivial while gz check reported 47/47.
  • docs/evals/gzkit-campaign-architecture-review-2026-08-17.md — self-consumption finding and the August measured baseline.
  • docs/evals/three-pillars-*-2026-09-19/ — most recent evaluation set, including published null results.
  • docs/governance/external-proving-ground-note-2026-06-24.md — names rhea; "Explicitly deferred."
  • data/mechanical_witness_grandfather.json — 64 unwitnessed "Mechanical" rows, shrink-only.
  • data/module_size_grandfather.json — empty list; ceilings_tightened records the 2026-08-26 p95 → p99 band move.

Material limitations

  1. This is a baseline. No prior report in this series exists, so trend claims draw on the project's own dated diagnostics (June and August 2026) rather than an earlier report in this format. Those diagnostics were authored in-repo; I re-measured the figures I rely on and say so where I did.
  2. The rate arithmetic is elapsed time, not effort. 3.7 days per OBPI is calendar time inside an active window. It is not a measure of operator hours and must not be read as one.
  3. Two chains treated as one family. The ~19-member doctrine-declared-without-mechanism figure spans two recurrence chains no author has linked; the project records this as operator judgment rather than measurement, and the 19-of-32 ratio has not been re-measured since 2026-09-02.
  4. The failure-class index has never recorded a passing run, so the numbers it produced are the most stale in this report.
  5. Adopter evidence is partly unobservable. wtcis-pvecob/cob_reports is private; I can confirm the defect filing and its version, not that repository's current state.
  6. gz check measures the staged tree. It passed on a clean tree at 85c857263; it is not a statement about uncommitted work.
  7. Parts of the evidence were gathered by delegated read-only agents. Every figure this report's argument rests on was re-measured directly before use; figures carried from those reports without independent re-measurement are the scorecard row counts, the chore registry totals, and the eval summaries.