Research
Research status
Ordo is a methodology under active research, not a validated discipline. This page is the whole map: what is supported, what is provisional, what has been refuted, and what is simply unknown.
Two labels, never merged
A rule can be required by the methodology while the benefit it predicts remains unproven. Those are different statements and Ordo keeps them apart with two vocabularies.
Confidence class
How strongly the rule is held
REQUIRED — strong enough and fundamental enough that the methodology
depends on it.
RECOMMENDED — evidence-backed or strongly justified, with contextual
exceptions.
EXPERIMENTAL — promising, needs validation.
RESEARCH ONLY — not yet suitable as normal guidance.
Evidence state
What the evidence currently says
Supported — evidence establishes it within its stated scope.
Provisional — some evidence, or an adopted contract whose predicted
benefit is untested.
Contradicted — measurement refuted it. Kept permanently.
Open — asked, not answered.
An adopted decision can make a contract REQUIRED. It cannot make the outcome that contract predicts Supported. Several rows below are exactly that pairing, and they are the rows most likely to be misread.
The map
| Proposition | Evidence state | Confidence class | Records | Do not conclude | Source |
|---|---|---|---|---|---|
| Change semantic authority before propagating outward: a semantic change starts at the model, not at the edge. | Supported | REQUIRED | EV-HN-2026-0003 EV-HN-2026-0004 EV-HN-2026-0005 TH-SDE-2026-0001 | No counterexample found, but this has not been tested outside one application and one language. | Sourcekemiller2002/state-directed-engineering / doctrine/EVIDENCE-TO-ENGINEERING-MAP.md — Evidence-to-Engineering Map |
| Compile or build before searching the repository broadly, where the architecture provides exhaustiveness checking. | Provisional | RECOMMENDED | EV-HN-2026-0005 HY-SDE-2026-0006 | Substitutes for search only in tiers the hardening actually covers — the API tier in Experiment 3 was unguarded — and an agent that reads broadly first under-realises the benefit. | Sourcekemiller2002/state-directed-engineering / doctrine/EVIDENCE-TO-ENGINEERING-MAP.md — Evidence-to-Engineering Map |
| A boundary fact needs one authority, a derivation from it, or a mechanical agreement check between its copies. | Supported | REQUIRED | EV-HN-2026-0002 TH-SDE-2026-0002 | Not tested on failure shapes beyond the four reproduced in the source investigation. | Sourcekemiller2002/state-directed-engineering / doctrine/EVIDENCE-TO-ENGINEERING-MAP.md — Evidence-to-Engineering Map |
| Behavioural and present-but-wrong implementations require behavioural verification. Compile, architecture and contract checks cannot substitute. | Supported | REQUIRED | EV-HN-2026-0004 EV-HN-2026-0005 HY-SDE-2026-0003 HY-SDE-2026-0005 | Replicated twice and the strongest-replicated finding in this map — but still within one application. The greenfield experiments corroborate the shape independently, not the same measurement. | Sourcekemiller2002/state-directed-engineering / doctrine/EVIDENCE-TO-ENGINEERING-MAP.md — Evidence-to-Engineering Map |
| Boundary hardening reduces search operations and repair loops. | Provisional | RECOMMENDED | EV-HN-2026-0005 HY-SDE-2026-0001 HY-SDE-2026-0002 | One paired trial. Unique files inspected did not move the same way search operations did, and the two agents' investigation strategies differ enough to confound the comparison. | Sourcekemiller2002/state-directed-engineering / doctrine/EVIDENCE-TO-ENGINEERING-MAP.md — Evidence-to-Engineering Map |
| Boundary hardening reduces total boundary propagation. | Contradicted | No class | EV-HN-2026-0003 EV-HN-2026-0004 EV-HN-2026-0005 TH-SDE-2026-0004 | Refuted. Boundary Change Amplification was 4.0, identically, three times across two architectures. Do not state that boundary hardening reduces the number of boundary files a change touches. | Sourcekemiller2002/state-directed-engineering / doctrine/EVIDENCE-TO-ENGINEERING-MAP.md — Evidence-to-Engineering Map |
| A higher mechanical discovery rate means lower engineering cost. | Contradicted | No class | EV-HN-2026-0005 TH-SDE-2026-0003 | Refuted within a single trial: the hardened condition had a lower as-experienced discovery rate and lower search and repair-loop cost at the same time. The two proxies moved in opposite directions. | Sourcekemiller2002/state-directed-engineering / doctrine/EVIDENCE-TO-ENGINEERING-MAP.md — Evidence-to-Engineering Map |
| Exhaustiveness, architecture and contract checks are sufficient for behavioural correctness. | Contradicted | No class | EV-HN-2026-0004 EV-HN-2026-0005 HY-SDE-2026-0005 | Refuted. The present-but-inert fan-out arm passed every mechanical check identically on both architectures. | Sourcekemiller2002/state-directed-engineering / doctrine/EVIDENCE-TO-ENGINEERING-MAP.md — Evidence-to-Engineering Map |
| API and HTTP boundaries are sufficiently protected by the mechanisms these experiments tested. | Contradicted | No class | EV-HN-2026-0005 | Refuted, and open. Zero mechanical guard exists for this tier on either architecture tested. A candidate mechanism is sketched in the source report; it has not been built or tested. | Sourcekemiller2002/state-directed-engineering / doctrine/EVIDENCE-TO-ENGINEERING-MAP.md — Evidence-to-Engineering Map |
| Explicit state modelling reduces AI-agent token or wall-clock cost. | Open | No class | EV-HN-2026-0005 | Open, not answered. The one trial that measured it has incomplete telemetry for the treatment condition: the recorded figures cover a resumed partial run only. Never cite the raw 59%, 68% and 59% reductions as a confirmed result. | Sourcekemiller2002/state-directed-engineering / doctrine/EVIDENCE-TO-ENGINEERING-MAP.md — Evidence-to-Engineering Map |
| A shared host-language contracts assembly is sufficient boundary protection for a same-language pair. | Contradicted | No class | EV-HN-2026-0002 TH-SDE-2026-0002 | Refuted by direct reproduction: the identical shared type produced three different wire behaviours depending on call path, inside one process. | Sourcekemiller2002/state-directed-engineering / doctrine/EVIDENCE-TO-ENGINEERING-MAP.md — Evidence-to-Engineering Map |
| A public integration contract must not require adopting this methodology in order to integrate with a system built under it. | Provisional | RECOMMENDED | EV-HN-2026-0002 TH-SDE-2026-0002 | Plausible but not directly tested. The source investigation gives it exactly that rating and this map does not upgrade it. | Sourcekemiller2002/state-directed-engineering / doctrine/EVIDENCE-TO-ENGINEERING-MAP.md — Evidence-to-Engineering Map |
| The construction method's fixed ordering — compile, architecture, boundary and contract, behaviour, search, integration. | Provisional | EXPERIMENTAL | EV-HN-2026-0005 | Assembled from the supported rows above rather than trial-tested as one ordered workflow. Not yet validated as a sequence on a real project. Reorder only with recorded rationale. | Sourcekemiller2002/state-directed-engineering / doctrine/EVIDENCE-TO-ENGINEERING-MAP.md — Evidence-to-Engineering Map |
| A semantic area may span several files and clusters while its semantic authority remains singular. | Provisional | REQUIRED | EV-SDE-2026-0006 DF-SDE-2026-0004 TH-SDE-2026-0005 HY-SDE-2026-0007 | This is an adopted contract, not a measured outcome. No controlled comparison of a monolithic against a decomposed treatment exists, and physical splitting can worsen navigation or duplicate rules. | Sourcekemiller2002/state-directed-engineering / doctrine/EVIDENCE-TO-ENGINEERING-MAP.md — Evidence-to-Engineering Map |
| Use a repository semantic map and feature manifests for deterministic local-first routing in a non-trivial project. | Provisional | REQUIRED | EV-SDE-2026-0006 DF-SDE-2026-0004 HY-SDE-2026-0008 | An adopted execution contract with no outcome trial behind it. Manifest maintenance and staleness may offset the benefit, and a small repository may not need per-feature manifests at all. | Sourcekemiller2002/state-directed-engineering / doctrine/EVIDENCE-TO-ENGINEERING-MAP.md — Evidence-to-Engineering Map |
| Bounded reasoning scope — limiting what must be inferred at once to make a safe change — reduces engineering cost or defects. | Open | RESEARCH ONLY | EV-SDE-2026-0006 TH-SDE-2026-0005 HY-SDE-2026-0007 HY-SDE-2026-0008 | This is the methodology's central unifying idea and it is the least evidenced thing on this site. No causal comparison exists. Task, model, tooling, familiarity and verification differences may dominate any effect. | Sourcekemiller2002/state-directed-engineering / doctrine/EVIDENCE-TO-ENGINEERING-MAP.md — Evidence-to-Engineering Map |
| Treat physical line count as a configurable review signal, not as semantic truth. | Provisional | RECOMMENDED | EV-SDE-2026-0006 DF-SDE-2026-0004 HY-SDE-2026-0007 | The specific thresholds are provisional and sensitive to language and generated code. Line count cannot detect duplicated authority, which is the thing it is often used as a proxy for. | Sourcekemiller2002/state-directed-engineering / doctrine/EVIDENCE-TO-ENGINEERING-MAP.md — Evidence-to-Engineering Map |
| Context Surface, Context Expansion Ratio and Discovery Expansion diagnose unexpected context growth. | Open | EXPERIMENTAL | EV-SDE-2026-0006 HY-SDE-2026-0007 HY-SDE-2026-0008 | Proposed, not validated. The counting unit and denominator must be preregistered or the numbers mean nothing, and cross-cutting changes legitimately expand context. Never gate conformance on these. | Sourcekemiller2002/state-directed-engineering / doctrine/EVIDENCE-TO-ENGINEERING-MAP.md — Evidence-to-Engineering Map |
| A serious experiment captures durable telemetry at fixed checkpoints and leaves genuinely unavailable metrics missing rather than estimated. | Supported | REQUIRED | EV-HN-2026-0005 EV-SDE-2026-0006 DF-SDE-2026-0004 | The no-invention rule is evidenced; the specific checkpoint schedule is an adopted practice for future trials. Checkpoints cannot retroactively repair telemetry that was never captured. | Sourcekemiller2002/state-directed-engineering / doctrine/EVIDENCE-TO-ENGINEERING-MAP.md — Evidence-to-Engineering Map |
| Heterogeneous verification mechanisms find defects that earlier mechanisms — including careful, mutation-tested self-review — did not. | Supported | RECOMMENDED | EV-HN-2026-0004 EV-HN-2026-0005 | Supported by four independent observations in three codebases: two deliberate no-op probes and one replication where an independent reviewer's five findings had zero overlap with a 90-test suite and a self-review. All four involve agents from one model family, so a shared blind spot cannot be ruled out. | Sourcekemiller2002/state-directed-engineering / research/evidence/EV-HN-2026-0004--experiment-2-boundary-authority.md — Residual finding |
Why the refuted rows stay
five propositions here are classified Contradicted. They are not deleted, and they are not moved to a footnote. They stay in the map permanently, cross-linked to the record that refuted them, so that a future engineer or agent cannot silently re-propose them without first seeing why they were rejected.
Two of them are the sort of thing a methodology would normally quietly stop mentioning. Boundary hardening was expected to reduce how much boundary work a change requires; measured three times across two architectures, it did not. A higher mechanical discovery rate was expected to mean lower cost; in one paired trial the two moved in opposite directions.
The least evidenced thing here
Ordo's unifying idea — that bounding how much must be inferred at once to make a safe change is what produces the benefit — is classified RESEARCH ONLY with an evidence state of Open (CL-016).
There is no causal comparison behind it. Task, model, tooling, familiarity and verification differences may dominate any effect entirely. It is the organising principle of the whole methodology and it is the weakest claim on this site, and putting those two sentences next to each other is the point of this page.
Open questions
- Does explicit state reduce agent cost? Re-measured, and still not demonstrated. The original trial was interrupted and its treatment-condition telemetry covers a partial run (CL-010). Two further experiments ran eleven fresh trials, each in an isolated session so the two conditions could not share telemetry. Neither reached its pre-registered sample size. Across every completed, independently verified run the conditions are separated by $2.61 against within-condition spreads of $9.78 and $15.15 — and the final pair reversed, with the hardened run costing more. The effect is not shown to exist and is not shown to be absent; what is now measured is that it cannot be separated from run-to-run noise at this sample size.
- Can the HTTP and API tier be mechanically guarded? A candidate mechanism is sketched in one source report. It has not been built or tested (CL-009).
- Does the construction ordering work as a sequence? Its individual steps are supported; the ordering has never been trial-tested as one workflow on a real project (CL-013).
- Do navigation artifacts pay for their maintenance? No outcome trial exists, and staleness may offset the benefit entirely (CL-015).
- Is agreement between two agents evidence, or a shared blind spot? Every agent in every experiment here came from one model family. Two systems that fail the same way are not two mechanisms, and this evidence base cannot currently tell those cases apart.
- Does any of this transfer? Four languages, one organisation, no team-scale observation, no counterfactual arm.
The standing rule
The repository that develops this methodology carries its own warning, and it applies to this site too: do not infer that State-Directed Engineering is a validated discipline, method or product merely because the process around it is rigorous. A careful process is evidence about the process.