WarrenBarr
SEC-01
Advisories
Published findings
OPS-02
Methodology
How an engagement runs
ENG-03
IT & Dev
Build and administer
ARM-04
Software
Tooling I wrote
MSP-05
MSSP
Flat-rate managed
REC-06
Case Studies
Work, in full
WEB-07
Websites
Sites that convert
SUP-08
Gear
Coming soon
DOC-09
Doctrine
What I will not do
WHO-10
About
Who you'd be hiring
00 // Software · Developer Tooling · Case Study

A Development Brain, Folded Into A Live Production Product

The scarce asset in a long-lived codebase isn't the code — it's the decisions: why a boundary was drawn here, why an approach was rejected, what a subsystem is actually for. HESUS makes that reasoning durable and queryable. This is the story of folding it into a real shipping product, upgrading it across a generational engine rename, and measuring what actually held — including the experiments that didn't.

Subject  Omniscient — Windows fleet-hardening tool Engine  6,497 LOC PowerShell Date  June 2026
01 / Shift
From Authoring Code To Governing Decisions
02 / Shift
You Don't Maintain The Output. You Version The Reasoning.
03 / Shift
Architectural Drift Is A Data Problem, Not A Discipline Problem

AI-assisted development has a quiet structural failure: decisions aren't tracked. An architectural choice made in session one is forgotten by session five. The same utility gets reinvented across sessions. Code quality starts to depend on whether the assistant happens to "remember" an earlier design choice — and it doesn't. The reasoning lives scattered across commit messages, docstrings, runbooks, and wikis that drift out of sync with the code, and it evaporates between people and between sessions.

Omniscient is a real product with this exact shape: a substantial PowerShell engine and, at the start, 81 documented architectural decisions scattered across decision logs, docstrings, and comments. Cold-start, an assistant reads the codebase with no decision context. Warm-start, it reads the entire decision history and burns a large fraction of its working context just orienting. Neither is good. That's the problem HESUS was folded in to solve.

The difference between a project that evolves and one that accumulates technical debt is not velocity or test coverage — it is whether the reasoning-state is versioned. A project that accumulates carries every past decision as undifferentiated sediment: live constraints and dead ones look identical in the code, superseded approaches leave residue no one dares remove, and each change is made against an architecture nobody can fully reconstruct. Architectural-drift is the widening gap between a system's intent and its current state, and it widens silently precisely because intent was never recorded as state. A project that evolves keeps that intent as versioned, decoupled-reasoning: decisions are typed, supersession is explicit, and a rejected path stays legible as rejected. The codebase still changes constantly — but every change is made against a known architectural position, not a guess. HESUS doesn't slow accumulation down; it changes the substrate, so growth compounds reasoning instead of debt.

HESUS treats decisions as first-class objects. Each lives as a structured record — in the cortex a record carries a Title, Context, Decision, Rationale, Consequence, Status, Supersedes and a version Bundle marker — and the engine loads them all into a small in-memory graph, enforcing unique keys and resolving cross-references on load. Records are typed by a Kind field: a fact the operator stated, an observation the brain inferred, or a ratified decision. Only an unconfirmed observation surfaces a loud trust-tag; confirmed records read clean. The default Kind is decision, so an existing cortex stays backward-compatible the moment the typing was introduced — that exact reasoning is itself recorded as decision HD-001 in HESUS's own brain.

The load-bearing design choice is that the engine and the data are strictly separated. Every organ reads project-specifics through one seam — a descriptor file (hesus.toml) — so no organ hardcodes a path, marker, id-scheme, or threshold. The brain is project-blind: it names only its own organs and verbs, never a project fact. That separation is why the same ~6,200 lines of engine run unmodified against any project, and why the engine can be swapped underneath a project without touching accumulated knowledge. The descriptor even ships a tiny pure-Python TOML parser as a fallback, so the brain needs zero pip installs and runs on Python 3.8+ — "a brain that needed pip install would be friction at exactly the moment you want zero friction."

The brain assembles and points; the operator judges and acts. A brief is "a curated reading list plus a map of what's in scope" — it errs generous, points at full sources, and shows what it searched so the operator can tell if something's missing. It never summarizes a decision in place of the decision. That line is written into the code as BD-018: a summary pre-connects badly and inhibits creativity; a pointer-pack gathers and gets out of the way.

Day to day, the benefit is context-efficient briefing. The waste a brief targets isn't orientation or finding one decision — it's the four or five exploratory round-trips at the start of a task (query, then debt, then hunt for files, then read the anti-patterns) just to assemble context before the assistant can think. A brief collapses that into one call: the decisions that may govern the task, the files to read in full, and the constraints in play. Because a brief stays near-constant in size while the decision memory grows, the saving compounds — the bigger the cortex, the more a brief saves over loading it all.

HESUS is a set of small, single-purpose organs behind one entry point. The ones an operator touches daily: brief (task-relevant retrieval that points, never summarizes), gestate (extract an existing project's decisions on first contact), query (pull one record and expand it), and a set of gates that keep the brain honest.

Gestate — three-pass extraction
On existing code, one-pass init builds a good skeleton but captures zero decisions. The measured failure (recorded as HD-051): a 5,460-line codebase held 109 decision-bearing phrases across 19 of 42 files — one-pass ingested none. Gestate runs SURVEY → EXTRACT (mine docstrings and WHY-comments for decision language) → SYNTHESIZE, emitting candidates each as an observation pointing at its source file:line, capped, for the operator to ratify.
The ship gate — plugin guards
The gate auto-discovers every check-*.py guard and reports one PASS/FAIL. Findings are tiered: parse/runtime block, hygiene warns. A reviewed false-positive isn't silenced in code — it's waived in a triage registry with a written reason. A broken guard can't crash the gate; a malformed triage file warns loudly rather than silently re-blocking.
The version spine
Every organ carries a PART-VERSION stamp; selfcheck.py reconciles each against a single manifest so a part can't ship with a forgotten version. This exists because of a real lapse: three early passes all shipped as "V1A" because there was no spine to increment against. There is now.
The development-loop gate
HESUS ships with a brain that tracks its own development. A ship gate refuses to let the engine release while that development-brain is stale — the tool is forced to eat its own cooking before release.

These gates verify what they're built to verify — version integrity, execution, project-blindness — not correctness. That distinction is stated plainly rather than oversold.

The brain Omniscient runs is a prior generation of the engine — a different entry-point and a different configuration name, from before a generational rename. The real test wasn't “does it work on a fresh project” but “can the engine be swapped underneath an accumulated cortex without losing anything.” Auditing that real production fold surfaced the cost of leaving the seam implicit: a leftover configuration file from a rename never retired, a stale comment naming the old seam, and runtime caches stale against the new engine. None were the project's bugs — each was a place the fold couldn't be mechanical.

That drove codifying upgrade as a first-class operation. The engine ships a machine-readable capability manifest, and an upgrade diffs the installed manifest against the incoming one, classifying every capability into exactly what it will cost: drop-in (works on the existing cortex immediately), seam-migration (a rename/move — migrate and retire the old, no orphan), new-gate (verification that may legitimately go red on first run), needs-data (present but inert until the cortex is armed — the dormant trap, surfaced loudly), and unknown (never assumed safe). The operation is thought-gated: the instant the two versions match it no-ops, so an up-to-date project pays nothing.

The acceptance test is printed by the tool itself: the new engine drives the OLD cortex → 0 errors, record count UNCHANGED, descriptor resolves the SAME project. Nothing gestated may be lost. That is the test that turns a one-shot trick into a durable methodology.

In AI-assisted engineering the engine — the specific model, version, and prompt regime that generated a system — is the most perishable component you own. It is replaced on a vendor's release cadence, not yours. A codebase whose architectural reasoning lives only in the chat history of a deprecated model is already orphaned; it just hasn't been told yet. The reasoning that justified every boundary evaporated the moment that session closed.

HESUS treats this as a structural fact and decouples accordingly. The Cortex — the versioned-state of decisions and the architectural reasoning behind them — is held as durable, project-owned data. The Engine — the retrieval-and-verification machinery, and by extension whatever model drives it — is project-blind and replaceable. The two are joined at a single descriptor seam. This is not a convenience; it is the load-bearing invariant. It is why, on the live production fold, the engine was swapped across a generational rename while the Cortex was driven at zero errors, record count unchanged — decoupled-reasoning surviving an engine it was never written against.

The decisions are more valuable than the code because the code is the cheap artifact. Given the decisions, the code is regenerable — by the next engine, repeatedly. Given only the code, the decisions are unrecoverable: you read what was built and never why, and every change becomes a guess against intent you no longer hold.

HESUS inverts the thing you preserve. You stop versioning the disposable output and start versioning the irreplaceable reasoning — which is the only asset in the system whose value compounds instead of decaying with the model that produced it.

Six months after a security boundary is drawn, the question is rarely what the boundary is — the code shows that. The question an auditor, an incident responder, or a new lead actually asks is why it was set there, who ratified it, and what was rejected in its favor. In a conventional workflow that answer lives in a closed chat window, a chat thread, or nobody's memory. Forensically, it is gone.

The HESUS Cortex is a queryable Decision Graph in which each record is a forensic decision log: a typed entry carrying the decision, its rationale, its consequence, its status, and its provenance — pinned to the source file and line it governs. Three properties make it audit-grade rather than documentation:

Even the gate that ships the system carries this discipline: a finding that's been reviewed and waived isn't deleted from the code, it's entered in a triage registry with a written reason. The accountability trail is mechanical, not cultural — it does not depend on anyone remembering to document.

One honest boundary on this claim: HESUS makes the reasoning auditable — the decision, its provenance, its lineage. It is not a tamper-evident ledger and makes no cryptographic chain-of-custody guarantee about the records themselves. It answers "why was this decided, and what was rejected," with provenance; it does not, on its own, prove a record was never altered after the fact. Framed as a decision-accountability layer, it delivers exactly that.

A project doesn't have one kind of memory — it has three, and they must not mix. Decisions about the product are not decisions about the engine that remembers the product, and neither is an experiment that might be refuted next week. HESUS keeps them in three separate cortexes with separate id schemes, and one verb tells the assistant which one it's standing in.

Brain A
The project's memory
The decisions that govern the shipping product. Its ids follow the project's own scheme — ED-282, PAY-002. This is the brain a developer's brief reads. It ships inside the product folder and is the thing being protected.
Brain B
The development brain
Decisions about the engine itself — HD-001 through HD-152. Why retrieval scores a title above a caveat, why a summary is refused. Brain A's engine improvements fold up into B; nothing ever flows back down, so B's dev reasoning never contaminates a customer's project.
Brain C
The research brain
Studies and their verdicts — ES-001 through ES-032, plus RD- research decisions. Deliberately outside the A↔B sync: a lab, not a synced peer. A refuted hypothesis stays legible as refuted rather than propagating.
Figure 1 — brain topology and the one-way fold
BRAIN A Project memory ED-282 PAY-002 governs the product BRAIN B Development brain HD-001 … HD-152 governs the engine FOLD never B→A BRAIN C Research brain ES-032 RD-012 outside the sync archival a study can be refuted without touching what governs the product
Why the fold is one-way. Engine improvements must reach the development brain or B falls behind the very thing it develops. But a customer's project decisions must never leak upward into the engine's own memory, and the engine's internal reasoning must never appear in a customer's brief. The sync carries organs, spine and shared tools downward into B; a hard-coded local guard blocks the descriptor, the cortex and dev state from ever crossing.

Retrieval is the whole product. A decision that exists but isn't surfaced at the moment it governs is worth nothing — worse than nothing, because the operator believes it's covered. So the retrieval path was rebuilt, measured, and rebuilt again across thirty-two studies.

The naive approach flattens every record into one bag of words and counts overlap. That fails in a specific way: a term buried in a Caveat counts exactly as much as the same term in the Title, so a record that merely mentions "payment" outranks the credential-exposure record that exists to stop you. The fix was to score per field, then to add a channel that isn't a coincidence of wording at all.

Figure 2 — one query through the retrieval pipeline
TASK TOKENIZE SHARED INDEX SCORE + FLOOR "add logging to the session" logging session stem: log~ built ONCE per brief log→ ED-006, ED-282 sess→ ED-114 cred→ ED-006 export→ED-006 699 → 81 tokenizations 6.4× faster retrieval O(1) prefix stem lookup ED-006 credential-safety 4.88 ED-282 session handling 3.25 ED-114 log rotation 2.90 relevance floor — 28% of top ED-041 naming convention 0.71 dropped, not padded — fewer, stronger records FIELD WEIGHTS TriggerTerm exact4.5 TriggerTerm stem3.75 Title3.0 Decision2.0 Context / Rationale1.0 Caveat0.5 load-bearing records carry a further ×1.3 — a safety decision outranks a topical coincidence
The TriggerTerm is the load-bearing idea. A field match is a coincidence of wording. A TriggerTerm is hand-authored temptation vocabulary — written down precisely because a developer who types that word is about to do the forbidden thing. For a while an exact TriggerTerm hit scored the same as a title match, and a stemmed one scored less — the mechanism losing to the coincidence it exists to replace. Correcting the order moved the governing credential record from rank 2 to rank 1 and took recall@1 from 95% to 100% at base scale.
Try it — no download

The scorer described above runs in the browser against a worked sample project: eleven decisions, five studies, real field weights, real trigger terms, the ×1.3 load-bearing boost and the 28% relevance floor. Type a task and watch it rank the cortex — then type something the project has no opinion about and watch it say so.

Open the live dashboard demo →

Two structural failures were found in that pipeline and are worth naming, because both are the kind that pass every test suite. The first: retrieval was non-deterministic. Two sorts truncated on a hash-randomised tie-break, so the same query returned a different near-list on each run. The top hit stayed correct, which is exactly why twenty-five studies never caught it. The second is worse — found, ranked, and thrown away. A load-bearing credential-exposure record scored 3.25 and ranked seventeenth of sixty, behind sixteen generic competitors that matched one more query token; the eight-slot window discarded it. And the guard meant to catch that built its candidate list from the records already in the results, so a record crowded out was invisible to the mechanism designed to rescue it. Meanwhile the brief printed "read the GOVERNS section above" — while that section held only the competitors. A warning pointing at a record it has declined to show reads as reassurance.

The clean result is the one below: same project, same three hundred decisions, same eight development tasks. The only variable is how the constraints reach the model.

Figure 3 — context delivered per task, measured
Chart: context words per task. Without HESUS, pasting the architecture document each session grows linearly to 3,934 words at 300 decisions. With HESUS, a task brief stays flat at 228 words. Both surface the governing decision on 8 of 8 tasks.
Both conditions surface the governing decision on 8 of 8 tasks. HESUS reaches that at 228 words instead of 3,934 — and stays flat as the project grows, because a brief carries only what governs the task at hand. The full-dump condition is scored generously: its recall is 100% by construction, since the constraint sits in the context whether or not the model actually attends to it.
17×
less context at 300 decisions
0.0047
context-efficiency ratio at 1,500 records
100%
recall@k, 10 to 3,000 records
32
studies, 125 checks, 203 bundles
What was measuredBeforeAfterHow it was found
Cortex parse, 10,000 decisions10.2s0.30sTwo stacked quadratics — a tail-slice copy per token, and 64,040,000 str.lower() calls in the duplicate-key check
Retrieval, 850-record cortex33.9s5.3s699 tokenizations for one search collapsed to 81 by sharing one index across the brief's sub-searches
Brief length570 words324 wordsCeremony cut — per-decision expand hints, section explainers, headers. Zero retrieval regression
Concurrent cortex writes11 of 1220 of 20Unlocked read-modify-write. No corruption, but a silently dropped decision breaks the one promise the tool makes
Verb coverage in smoke24 of 4848 of 48Smoke reported "24/24 clean" while ship crashed on every invocation — it was never in the list. Omission is silent
Determinism across processes2 answers1 answer25 fresh processes, one query. Hash-randomised tie-break in two truncating sorts
False-quiet rate61%0%Novelty-gated confidence — the brief said nothing while a safety decision sat uncovered
Composite score20%86%Across the full study suite, from the earliest measured baseline

Everything below is drawn from the research brain that ships with the engine — thirty-two studies, each with a pre-registered hypothesis, a stated method, raw per-pass measurements, and a verdict that was allowed to be no. The charts are rendered directly from those payloads. Where a study refuted the thing it was testing, that is what the chart shows.

32
studies, hypothesis registered before the run
203
bundles, each recording what changed and why
125
checks in the engine's own suite
4
shipped levers later refuted and recorded as such
ES-004 — the context crossover, 300 decisions over three passes
Chart: context words per task. Re-reading source grows 1,207 to 3,613 words as the cortex reaches 300 decisions. A brief stays bounded at roughly 355 words. Recall held at 5 of 5, 8 of 8, 12 of 12.
The advantage is directional, and it grows. At 100 decisions a brief is 3.8× cheaper than re-reading the source it replaces; at 300 it is 10.2×. The brief itself is bounded — 319 words, then 355, then 355 — while the material it stands in for triples. Retrieval held at 100% throughout, and the brief stayed near 110ms.
ES-028 & ES-027 — context-efficiency ratio at four cortex sizes
Chart: context-efficiency ratio falling from 0.1476 at 10 records to 0.0047 at 1,500 records on a log scale.
The ratio is brief words divided by whole-cortex words. It falls from 0.148 at ten records to 0.0047 at fifteen hundred — a brief carrying under half a percent of what the cortex holds, and still answering. Reproduced independently in ES-027 across a folder-layout change at 50, 300 and 800 records: 0.131, 0.026, 0.0068, with recall@1 at 100% throughout.
ES-005 → 006 → 008 → 013 — the retrieval arc on a frozen probe set
Grouped bar chart: across four studies, tier-1 recall stays 15 of 15, tier-2 rises from 5 to 14, tier-3 rises from 2 to 14, out of 15 probes each.
Forty-five labelled probes, frozen, re-run against every shipped change. In-vocabulary retrieval was solved from the start — 15 of 15 the whole way. The work was in the two hard tiers: stem-only matches went 5 → 14 when load-bearing records were given a boost, and adversarial paraphrase went 2 → 14 once authors could declare trigger vocabulary. Total across the arc: 22 of 45 to 43 of 45. The same experiment refuted a sort-key change and a stem-weight change; both are in the record beside the two that worked.
ES-007 vs ES-001 — the same six tasks, before and after the efficiency work
Line chart: brief words per pass. The ES-001 baseline hovers between 309 and 420 words. The ES-007 run falls from 165 to 115 words by pass six.
The brief gets smaller as the project gets bigger. Ceremony was cut — per-decision expand hints, section explainers, headers — and content was not: recall stayed 6 of 6 across both runs. Orientation runs 94–191ms, a brief 60–69ms. Amber points are passes where the brief flagged its own confidence as low rather than answering quietly.
ES-028 — two engine generations, same cortex, same probes
Comparison: the older engine returned 2 distinct answers across 25 fresh processes for one query; the current engine returns 1. Retrieval accuracy is identical between them at every cortex size.
Ten bundles of work, measured honestly. The determinism bug is real and was fixed — one query, twenty-five fresh processes, two different answers from the older engine and one from the current. Verb coverage in the smoke suite went from 24 to 48. But retrieval accuracy is identical at every size, and the chart says so. Ten bundles bought stability and coverage, not retrieval quality.
ES-026 — eight guarantees against hostile input
Bar chart: eight guarantees, 69 of 69 trials passed across 53 hostile fixtures.
Sixty-nine trials, fifty-three fixtures, all passing — and it still found three real defects. Retrieval was non-deterministic. A missing cortex directory fingerprinted to empty and then “verified” against empty, so a broken configuration reported success. And a vendored sub-project's install survived a removal that promised to be total. A suite that passes everything on the first run usually means the fixtures are too kind.
ES-031 — when the measurement itself was wrong
Bar chart: fixture recall@1 95 percent and recall@k 100 percent, against real-brain recall@k of 0 percent as measured, rising to 48 percent after the fix.
This is the study that matters most, and it is the one that failed. Every accuracy number produced before it was circular — the same author wrote the decisions, the trigger vocabulary, and the queries used to test retrieval against them. Re-measured with queries derived from source docstrings instead, recall@k on the real development brain was zero.

The cause was isolated rather than argued away. Same cortex, same queries, only the trigger vocabulary stripped: recall@1 fell 95% → 25%, recall@k 100% → 45%. The gap was never synthetic-versus-real; it was with-triggers versus without — and the real brain carried triggers on nine of two hundred and twenty-two records. Drawing candidate vocabulary from each record's own source file took real-brain recall@k from 0% to 48%. 48% is not 100%, and coverage is 13%. Both numbers are in the shipped record.

Two more results are worth stating without a chart, because they are single comparisons rather than curves. ES-002: on a cortex dense with code identifiers, the identifier-aware tokenizer took recall from 0.00 to 1.00 — a compound name like SaveWTSettings is now emitted whole and split, so a natural-language task can reach a decision written in code. ES-003: on a small project measured with and without the brain across three passes, cross-pass regressions fell from three to one and recorded rationale went from nothing to twenty of twenty — while context cost went up, because at that size re-reading the source was cheaper. That result is why the compact brief exists, and it is why the crossover chart above starts where it does.

Both front-ends are deliberately thin over the same tested command line, so there is no second implementation to drift. They exist because two different people need the brain: the operator standing at a machine, and the operator handing a link to somebody else.

The desktop console
stdlib tkinter · zero dependencies · 11 tabs

A bound workstation tool. It remembers the brain it is tied to by lease id and reopens straight into its health on launch. Opening any other folder is a peek — it finds the nearest brain, however deep the folder schema buries it, and shows you that without moving the binding.

It is where the code and the reasoning become one navigable thing. The source browser renders Python through the standard-library tokenizer for real syntax colouring, and selecting a symbol re-scopes a panel to the decisions that govern it. The reverse works too: click a decision, jump to the file it names.

It also carries the parts nothing else does — a development timesheet with an accruing stopwatch and value-per-day graphs, and a capability catalogue that derives the project's public surface statically and caches it against a source fingerprint, so it recomputes only when code actually changes.

SEEDINSPECTEXPLOREOPSCODELIBRARYCAPSBUNDLESSTEERWORKR&D
The web dashboard
pre-baked single file · 0ms load · 9 tabs

This one started as a twenty-endpoint HTTP server and was deleted. Every tab had a loading state; every read was a round trip; a slow endpoint queued behind another. It was replaced by a generator: the command loads the project once, bakes every brain's records, every study payload and every chart into one HTML file, and opens it.

There are no fetch calls for reads. Switching brains re-renders the cortex instantly from data already in the page. A tiny background thread survives for exactly one endpoint — running a live brief — because that genuinely needs the engine.

It is also the multi-project view: it scans for every project carrying a brain, deduplicates by real path, and shows each one's record count, engine version and full bundle timeline. Output runs 34–132KB depending on cortex size.

ProjectsInspectCortexBriefStudiesGapsOpsProgressRelease
Figure 4 — why the dashboard was rewritten as a generator
BEFORE — 20 ENDPOINTS browser server loading state on every tab · slow endpoint queues the rest AFTER — PRE-BAKED load once bake all data one HTML file 0ms reads 1 endpoint kept — live brief
The check that finally worked. Every earlier dashboard regression passed a syntax check and still shipped an inert page — balanced braces, valid HTML, no click handlers bound. The generator now validates its own JavaScript at build time and refuses to write a broken file, and a separate check executes the generated page against a DOM shim and clicks every tab, asserting the matching panel activates. It is the first check that fails when the dashboard is dead rather than merely malformed.

A memory tool that loses a record, or reports a pass it did not observe, is worse than no tool — it converts an absence of information into false confidence. Most of the engineering here is defence against exactly that. These are the failures that mattered, in the shape they were actually found.

A false GREEN on box verification
Some things cannot be proven where they are built — a Windows-only code path, a renderer needing a real browser. Those ship unverified and the operator on the machine is the only oracle, so the engine records a claim: what it is, how to prove it, and the signal a pass looks like in the machine's own output. The operator pastes the output back and the claims resolve. That matcher compared the signal as a case-insensitive substring. So the signal PASS matched byPASS. Reproduced at its worst: a claim marked VERIFIED whose own cited evidence line read "bypass check failed — access was NOT denied." For a ledger whose entire purpose is honest verification, a false green is the one unacceptable error.
Word-boundary matching, then tuned again — a pure boundary rejected passed too, so the rule became a word-start boundary plus an optional inflection. PASS now matches passed, passes, passing but still not bypass, password, passage. Thirteen cases pinned.
Ten records that read as zero
The loader filters cortex records to the id schemes the project declares, to keep other subsystems' data out. A record whose prefix was undeclared was dropped with no error at all. A project holding ten real decisions under one prefix, while declaring another, read as an empty brain — and every organ downstream ran happily on nothing. No crash, no warning, just silence where the memory should have been.
The drop is now loud in all three places it can be seen: the loader reports every skipped decision-shaped record, status prints those warnings, and the self-check fails outright, naming the invisible records and the two ways to fix it. Separately, the verb that mints new ids was choosing its prefix by continuing whatever was most common in the cortex — which is how one off-scheme record makes every later record inherit the same invisibility. It now mints under the declared scheme, so a record can no longer be born invisible.
The measurement was grading its own answer key
This is the one worth the most. Every accuracy number produced before study thirty-one was circular — the same intelligence wrote the decisions, wrote the trigger vocabulary, and wrote the queries used to test retrieval against them. Measured properly against the real development brain, with queries built from source docstrings rather than authored alongside the answers, recall@k was 0% — against 95–100% on the synthetic fixture.
The cause was isolated rather than explained away. Same cortex, same queries, only the trigger vocabulary stripped: recall@1 fell 95% → 25%, recall@k 100% → 45%. The gap was never synthetic-versus-real; it was with-triggers versus without. The real brain carried trigger terms on nine of two hundred and twenty-two records. And the tool built to fix that could only see records tagged with a concern — eleven of two hundred and twenty-two. It offered one candidate for a two-hundred-record brain.
Records without a concern now draw candidate vocabulary from their own source file's identifiers — a developer about to touch a query module types search, score, index. Candidates went 1 → 21; real-brain recall@k 0% → 48%. Recorded honestly: 48% is not 100%, and coverage is 13%. Two of the test designs were wrong before the measurement was right, and both are written down.
An inference that looked good and was refuted
Eighty-three percent of records had neither trigger vocabulary nor a declared source. Inferring the source from a record's own wording looked like it would reach 152 of 186. Validated against the 38 records that do declare a source — a free labelled set — it scored 33% precision at a single identifier hit, and no record cleared two. Two of three guesses wrong, and a wrong trigger fires on unrelated work forever.
Refuted and recorded as refuted, not shipped. The upstream leak it was chasing was fixed instead. The honest residue stands: the 186 records with neither a source nor triggers remain unreachable, because the only method that could have reached them is the one just disproved.

There is a pattern in these. Every one was a mechanism that reported success while failing — a smoke suite reporting 24/24 clean while a verb crashed on every invocation because it was never in the list; a self-check reporting a file verified when its version stamp format meant the matcher skipped it entirely; a dashboard passing every syntax check and shipping an inert page. The engineering answer, applied repeatedly, is that a guard which can silently not-run is not a guard. Each fix is now paired with a check that asserts the guard itself is reachable — every dispatchable verb is either exercised or explicitly excused, and a new verb cannot go uncovered by accident.

Twenty pre-registered studies across 154 build passes. Three arenas: synthetic testbeds where ground truth is controllable, a hand-rolled binary-protocol codec the engine was never tuned for, and Omniscient — a real shipping Windows fleet-hardening tool carrying 326 live architectural decisions.

89%
context saved versus reading the source — break-even on the first pass, not the sixth
0
retrieval degradation from 40 to 1,000 decisions — a plateau, not a cliff
88%
of safety constraints surfaced across 30 adversarial tasks on the hardest domain tried
Context efficiency ratio (6 passes, 70-decision project)0.112 — briefs cost 11.2% of source reading
Cumulative context over 6 passes: briefs vs. naive re-reading863 words vs. 7,695 words
Retrieval at 1,000 decisions (in-vocabulary / paraphrase)100% / 90% — unchanged from 40 records
Brief latency at 1,000 decisions114ms (from 88ms at 40 — still interactive)
Adversarial safety recall, binary-protocol codec (30 tasks)88% — 29 of 33 constraints
Six-step cascading build — the first mistakeCaught; all 5 downstream violations prevented
Naive violations prevented on the real 326-decision project9–11 of 20
Head-to-head A/B: cross-pass contradictions3 → 1
Head-to-head A/B: decision rationale preserved0/20 → 20/20
Silent misses (a thin answer that still looked confident)61% → 0%
Frozen 45-probe retrieval baseline across the research arc22/45 → 43/45

Two results deserve to be read together. The context win is automatic — it arrives on pass one and widens as the project grows, because a brief stays near-constant in size while the codebase does not. The safety win is conditional — it is bounded by retrieval, and retrieval is bounded by vocabulary. A recorded constraint only helps if the brief surfaces it for the tempting task. That distinction is the honest core of the whole system, and the studies below are mostly the story of pushing on the second one.

Chart: context words needed to reorient — the brief stays bounded while re-reading the source grows with the project
The core economics, measured. As a project grows, re-reading the source to reconstruct context grows with it — while a brief stays bounded. At 100 decisions the brief is 3.8× cheaper; at 300 it is 10.2× cheaper. The gap widens with every decision you add. Generated by the engine's own study tooling.

Retrieval — deciding which decisions a brief surfaces — is the whole ballgame. A brain that holds the right constraint but fails to show it when you're about to violate it is worthless. So it was built as a chain of pre-registered experiments against a frozen set of 45 labeled probes, split into three difficulty tiers, with tier assignment verified mechanically rather than by eye.

BoundaryWhere exactly does retrieval fail?
100% when the task shares 2+ exact words with the decision. 33% on stem-only. 13% when they share nothing. The cliff is at the exact-word line — not where anyone assumed.
Trigger termsCan an author-declared vocabulary field bridge the gap?
Hard adversarial recall 13% → 67%. Inert without the field, so the control cell reproduced exactly — a clean experiment.
Ranking (2×2)Is the stem-only collapse a ranking problem, not a matching one?
Confirmed. Load-bearing decisions were being out-scored by chattier neighbours. Boost them: 33% → 93%. The rival hypothesis in the same factorial was refuted.
Precision auditWere those recall wins bought with noise?
No. All three pre-registered targets passed — and the boost actually reduced noise.
ReplicationDoes any of this survive off the testbed it was tuned on?
43/45 on a fresh, hardened cortex it had never seen.
ScaleDoes it fall apart at 1,000 decisions?
Zero degradation. 100% / 90% held from 40 records to 1,000, still interactive at 114ms.
Adversarial domainDoes it work on a domain the author never optimized for?
88% on a hand-rolled binary-protocol codec — 100% in-vocabulary, 90% paraphrase, 77% cross-domain.
Chart: retrieval recall by vocabulary tier — 100% on exact-word overlap, 33% stem-only, 13% on zero overlap
The study that found the wall. Recall by vocabulary tier: near-perfect when a task shares exact words with the governing decision, collapsing as the wording diverges. This chart is the product's honest weakness, published rather than hidden — and it became the baseline every later fix had to beat. Generated by the engine's own study tooling.

The arc, in one line: the frozen baseline went from 22/45 to 43/45 — and every step of it is a study you can read, including the ones that went the wrong way.

The most useful study wasn't about recall percentages. It simulated how damage actually happens in AI-assisted work: a six-step build where each step's mistake feeds the next. Step one adds a "lenient mode" flag. Step two relaxes a check because lenient mode exists. Step three trusts the relaxed check. By step six you have shipped a security hole, and no single step looked wrong.

HESUS caught the first domino — the brief surfaced the strict-by-default decision the moment the lenient flag was proposed, and flagged caution. Catching that one step prevented all five downstream violations. Five of six steps surfaced their governing constraint; step six slipped (caching a parsed chain shares no vocabulary with zero key material from memory), and it is written down as a miss.

That is the shape of the value. Not "the AI never makes a mistake" — but the expensive mistake, the one that compounds, gets caught at the top of the chain where it is still cheap.

Chart: cross-pass regressions per pass, developing without the brain versus with it
Head-to-head on the same project, three passes: the same work done without the brain and with it. Cross-pass contradictions fell 3 → 1, and decision rationale went from 0/20 preserved to 20/20. The one that got through is on the chart too. Generated by the engine's own study tooling.

Three studies argued against the product. They are the reason to trust the other seventeen.

It retracted its own headlineTwo early studies reported 100% retrieval recall, and that number was going in the marketing. A later, harder study proved it was an artifact of easy probes: real paraphrase already lost a third of recall. A pre-registered rule — "if this invalidates prior work, say so" — fired, the headline was retracted, and the harder probe set became the permanent baseline every subsequent fix had to beat. The engine got better because the study was allowed to embarrass it.
It deleted code that didn't earn its placeFour zero-dependency mechanisms were built to close the remaining paraphrase gap — character n-grams, a synonym lexicon, and two others. The bar was set before the run: recover at least half the gap. None did. Two were refuted outright and their code was removed from the engine; one shipped as an optional flag at an honest 25%. The negative result is archived. The remaining limit — what the studies call the authoring wall — is stated plainly rather than papered over.
It corrected its own scale claimAn early result showed retrieval collapsing to 65% at 1,000 decisions — a real cliff, and a serious problem. Re-run with honest probes drawn from the cortex's actual vocabulary, the cliff vanished. It had been a flaw in the test, not the engine. The corrected result (zero degradation) replaced it; the original error remains on the record where anyone can see it.
The gate that saved itA brain migration once passed three green gates and still silently dropped three organs. Green gates only verify what they were built to verify. That incident became a permanent capability-audit gate that now refuses any upgrade which would lose a capability — and it has blocked a bad ship since.
ProvenContext efficiency — 89% saved, break-even on pass one, widening with project size. Scale — zero retrieval degradation to 1,000 decisions, still interactive at 114ms. Constraint surfacing — 88% on an unfamiliar adversarial domain; 9–11 of 20 naive violations prevented on a real 326-decision production codebase. Cascade prevention — the first domino caught, five downstream violations stopped. Knowledge preservation across a generational engine upgrade, enforced by a gate that refuses to lose an organ.
ConditionalThe consistency benefit is bounded by retrieval, and retrieval is bounded by vocabulary. When a task and its governing decision share no words, it can still miss — and the studies name the exact cases where it did. Declaring trigger terms closes most of that gap (and the tool now tells you which decisions lack them, and proposes candidates that were accepted 100% of the time in blinded review). Nothing closes all of it. Ingest quality is the ceiling: a thin or padded cortex steers weakly, and a padded one is worse than none because it looks authoritative.
Not yet provenEvery study here was authored and audited by the same intelligence that ran it. Violations are assessed, not observed in the wild by an independent, blinded agent. A proper multi-agent randomized trial — a genuinely amnesiac AI, no author in the loop — has not been run. Until it is, the behavioral lift is recorded as modeled, not demonstrated. It would be easy to omit this paragraph. It stays because the entire product is a bet that saying so is worth more than the sale it costs.

HESUS does not autonomously write code or make decisions. It makes an operator's judgment durable, queryable, and provenance-preserved — and it tells you how much it might be missing. Sold as exactly that, the measured results above are what it delivers.

Lease HESUS for your project

$300 per project — folded in once, yours to keep in that codebase for life. Pure Python, standard library only, nothing added to your runtime.

Figures were measured on live production artifacts on 2026-06-24 and are reproducible from the shipped bundles. Where a number is projected rather than measured, it is labeled as such.