Under the Hood · launch edition 1.0.0
A result you can
take apart.
Can a learned construction method survive a change of obligations—and help rebuild the context needed to justify a decision?
Start with experiment seven. Follow its definitions, a small proof, matched controls and runnable code. Then trace the six experiments that supply its foundations. This is an inspectable research laboratory for bounded mechanisms on the path towards MeRSIA.
out of 272 frozen records
when a current grant is supplied
under the declared reference
Exact finite verification with supplied meanings and feedback. These numbers do not measure real-world safety or the percentage of MeRSIA achieved. Version 1.0 marks the launch edition of the laboratory.
01 · State the contract
What must a decision know?
An input record is x = (s, c, p, a, u). The observed count is s, capacity is c, and 0 ≤ s ≤ c. Three supplied Boolean fields record permission evidence p, anchor warrant a, and source currentness u.
Old risk: r₀(s,c) = 1[s = 0 ∨ s = c − 1]
New risk: r₁(s,c) = 1[s = 1 ∨ s = c]
New release eligibility: Q₁(x) = ¬r₁(s,c) ∧ p ∧ a ∧ u
1[statement] is one when the statement is true and zero otherwise. Eligibility means that this supplied reference permits the record. A live receiving grant is an additional, separate condition for acting. Learning the correct label cannot issue that grant.
Development searches all 32 subsets of the five fields on 120 records with capacities 3–5. Nine teaching rows select a risk expression; six separate rows at capacity 5 validate it. The frozen evaluation uses all 272 records at capacities 6–9. “Frozen” means the selected rule and view are fixed before those evaluation records are scored.
What is supplied, and what is learned?
The fixture supplies the reference change and its authorization, exact feedback, count meanings, adapter, interpreter, higher recipe language and trusted host. Within those foundations, the experiment selects a new risk expression and a sufficient coordinate projection. It does not learn which human authority is legitimate or discover unfamiliar field meanings.
02 · Expose the failure
Same visible context. Opposite obligations.
A context view C is adequate for Q₁ on a domain D exactly when equal visible contexts always have equal required answers:
∀ x,x′ ∈ D: C(x) = C(x′) ⇒ Q₁(x) = Q₁(x′)
In plain English: if the view makes two records indistinguishable, it must be safe to give them the same eligibility label. The initial count-only view C(x) = (s,c) fails immediately:
Record A
(0, 3, 0, 0, 0)Visible count: (0,3)
Required eligibility: 0
Record B
(0, 3, 1, 1, 1)Visible count: (0,3)
Required eligibility: 1
A small proposition. On the declared development domain, retaining all five fields is the unique minimum adequate view among coordinate projections. Here is a short proof: omitting p, a or u hides a decisive toggle at (s,c)=(0,3), with the other gates true. Omitting s identifies (0,3,1,1,1) with (1,3,1,1,1). Omitting c identifies (3,3,1,1,1) with (3,4,1,1,1). Each pair has opposite labels. Every field is therefore necessary in this projection family; the full identity view is sufficient.
The exhaustive 32-view search corroborates this proposition. It is not a global minimum information claim: another representation language could encode the answer itself in one bit. The result concerns the declared coordinate views.
With the new risk expression but the old view, only 64 of the 272 frozen records have an unambiguous eligibility label. Requalifying the full view yields 272 correct labels. Under a current simulated grant, 26 are released and 246 blocked. With the grant off, all 272 effects are held.
03 · Compare fairly
What caused the improvement?
The inherited OR constructor makes (s = 1) OR (s = c) reachable in a 47-program language. Removing it leaves 26 candidates and no fitting expression. Restoring it restores acquisition. Compare that with both a preinstalled OR constructor and a larger direct search.
| Route | Candidates | New contract | Later charged work | Standalone source/replay work |
|---|---|---|---|---|
| Inherited | 47 | Qualified | 4,303 | 16,788 |
| Constructor removed | 26 | No expression | 4,108 | 16,788 |
| Constructor restored | 47 | Qualified | 4,303 | 25,182 |
| OR preinstalled | 47 | Qualified | 4,303 | 0 |
| Direct AND + OR search | 68 | Qualified | 4,492 | 0 |
The strongest matching control is preinstalled OR. It matches the inherited route’s later results and cost. The inherited method demonstrates reusable acquisition; it does not beat this preinstalled control on work. The 189-operation saving relative to the 68-candidate search is conditional pruning, not a general efficiency theorem.
Exactly what does “work” count?
The valid inherited later run charges 423 fitting predictions + 6 validation predictions + 3,840 projection record visits + 34 historical predictions = 4,303 operations. Direct fitting charges 612 predictions. Source acquisition/replay costs 8,394 per pass: the standalone inherited script performs two passes, restored performs three. Browser sequence accounting reports replay separately from earlier-stage acquisition.
These are heterogeneous declared logical operations, not elapsed time, hardware compute or a complete resource model. Memory, data construction, serialization and frozen scoring are excluded. The renamed-field condition adds 120 supplied-map checks. No amortized advantage or alignment-development rate dominance is established.
A deliberately unchecked old-rule shortcut releases 208 records, including 190 invalid releases under the new reference. That is a specific failure control, not a claim about ordinary AI systems. The old rule’s own 34/34 historical predictions are retained and scored separately; changing the task is not counted as improving the old task.
04 · Run the evidence
Your machine. The same finite test.
Download the script below and open a terminal in its folder. Use Python 3.10 or later; only the standard library is required. Every action here is simulated.
# Default: qualified labels, no receiving grant
python3 07_changing_obligations.py
# Supply a simulated grant: 26 releases, zero invalid
python3 07_changing_obligations.py --permission
# Keep the old view: 64 warranted labels, all effects held
python3 07_changing_obligations.py --phase 2 --permission
# Matched stronger control: same later work and result
python3 07_changing_obligations.py --route preinstalled --permission
# Removal: no fitting expression; all effects held
python3 07_changing_obligations.py --route removed --permission
# Stale qualification: all effects held
python3 07_changing_obligations.py --revision 2 --permission
# Three-peak rule outside the available language
python3 07_changing_obligations.py --condition off-language --permissionRead study.passed, study.costs and study.sourceCommonWork separately from live.warranted, live.admissions, live.wrongAdmissions and live.reason. The stored study.frozen comparison explicitly supplies a grant; the live output follows your command-line grant. Use --output result.json to save the full report.
Further probes include --condition bad-feedback, --condition unapproved, --condition wrong-map and --budget 4302. Each fails a different declared gate. A hold demonstrates the implemented boundary for that case; it does not establish that the checks catch every possible failure.
All seven scripts and reports → · Stored numerical fixture → · Repository validation instructions ↗
05 · Trace the machinery
Seven experiments. Explicit handovers.
- Coupled repairImprove rule and view under fixed evidence; emit a separately acquired construction artifact.
- Qualified inheritanceReplay that method in a fresh child; observations and grants stay local.
- Constructive enquiryUse the preceding milestone as provenance, with a separately supplied diagnostic compiler. Emit an audited gap.
- Acquired expressionUse the audited gap to initiate calibration; select and replay a qualified response expression.
- Bounded meta-extensibilityReplay the expression and acquire an OR constructor under fixed higher meanings and recipe language.
- Resource comparisonReplay the constructor, compare checking costs, and forward its qualified artifact.
- Changing obligationsReplay the constructor under a supplied count adapter and authorized new reference; acquire rule, repair view and requalify use.
The sequence is a chain of scoped contracts. It does not imply an automatic adapter between every mathematical model. The host checks envelopes and stops on reported failure; each module owns its checks. Modules are trusted developer-installed code in an honest host, not arbitrary plugins in a security sandbox.
06 · Keep the research open
The questions that matter next.
- Reference legitimacy. Who supplies the goals, authorizes changes and resolves disagreement? Those decisions are fixture inputs here.
- Uncertain evidence. What survives noisy, conflicting, delayed or adversarial feedback? The principal studies use exact supplied feedback.
- Meaning and representation. Can a system acquire unfamiliar semantics and richer views while preserving obligations? Current adapters and interpreter meanings are supplied.
- Resource and rate claims. Can benefits survive complete cost accounting and strong optimized baselines across new tasks? Alignment learning faster than other capability growth remains unproved.
- Independent replication and scale. Can external implementations reproduce the results, and do mechanisms transfer beyond finite toys? Browser/Python parity is internal verification.
Qualification receipts use exact structural/version binding in an honest local host. They are not cryptographic authentication. All simulated release results are relative to the supplied reference; they cannot establish real-world safety or general self-governance.
Where this connects to the publications
Learning What Must Be Distinguished, pp. 2–3: view adequacy and inadequacy witnesses. Towards Recursively Self-Improving Alignment, pp. 3–4: qualification, preservation and conditional composition. The Aligned Signal, preface p. 27: the MeRSIA proposal; §18.4, p. 89, discusses bounded meta-extensibility. Stage seven is a new finite study inspired by the manuscript, rather than a reproduced published general theorem.
These links provide publication citations and access information. The forthcoming manuscript’s private reading copy is not distributed here. Browse the publication references →
Inspect the evidence checklist →. Its 28/49 checklist entries describe documented gates across seven experiments; they are not a probability of safety or completion percentage for MeRSIA.
Source-available noncommercial research software. Commercial permissions by agreement. Full licensing terms →