Companion to: Day 0.5 — Don’t Trust It Yet
Date: 2026-08-20
Purpose: Keep audit state, criticism trace, methodological revisions and next actions outside the rhetorical layer of the essay.
This is not a second essay. It is the inspectable companion record.
Scope
The Day 0.5 audit used a 24-object TID / adjacent-product union.
- Windowsill received a deep audit.
- Most remaining objects received breadth gates around one or a few central invariants.
- PASS does not mean “fully validated”.
- OPEN does not mean “probably broken”.
- CONFIRMED should require evidence sufficient for the defined failure.
- The audit intentionally stopped breadth expansion after the 24-object union so Day 0.5 could be completed.
Classification model
PASS — the defined invariant held under the test performed using the evidence available at that time. PASS is reopenable.
CONFIRMED — evidence is sufficient to establish the defined failure.
OPEN — a substantive question remains unresolved, evidence is insufficient, or the required test has not yet been performed.
REJECTED FINDING — a finding may only be rejected against admissible evidence independent of a later explanation. For an internally authored specification or contract to carry strong rejection weight, its exact content should be externally time-anchored before the challenged public release.
Preferred chain:
contract bytes → SHA-256 → external timestamp → public release → later finding
A post-release specification cannot retroactively redefine the earlier public contract.
Method corrections produced during Day 0.5
MC-01 — Possible source path ≠ confirmed public failure
Trigger: Eurovision Explorer.
A tolerated failed-year request path was initially treated as a confirmed completeness bug. The implementation showed a possible failure mode, but no actual omitted public year had been reproduced.
State change: CONFIRMED → OPEN.
Method change: distinguish deterministic product failure from possible runtime/source failure path.
MC-02 — Parser output ≠ rendered artifact
Trigger: 24 Doors PDF.
Parsed text indicated mojibake. Parser output alone was insufficient to establish a visible presentation defect. The PDF was rendered and visually inspected; visible corruption was then confirmed.
Method change: presentation defects in generated artifacts require artifact-level verification when parser output can diverge from reader-visible output.
MC-03 — NOT A BUG removed as a free adjudication state
Trigger: criticism of the Day 0.5 taxonomy.
The same actor could build the system, describe its intent, perform the audit and dismiss criticism as misunderstanding.
Correction: replace NOT A BUG with REJECTED FINDING, which must cite an admissible basis rather than post-hoc intent.
MC-04 — Rejection evidence needs temporal admissibility
A specification written after release but before criticism can still be post-hoc rationalisation.
Correction: internally authored contracts used for strong rejection should predate the challenged release and ideally be externally time-anchored.
Future rejection metadata should include: basis, basis version, basis date, external timestamp/anchor, public release date, and challenged behaviour date if known.
MC-05 — Criticism needs provenance
Future criticism records should distinguish, where practical: internal self-review, model-assisted adversarial review, external general criticism, external domain review, and replication/reproduction attempt.
These are not equivalent evidence.
MC-06 — PASS must remain reopenable
The Day 0.5 breadth audit did not systematically re-attack PASS findings after their first gate.
Future control: sample previous PASS findings and re-audit them using a different invariant, source path or reviewer.
MC-07 — Status labels must not become authority theatre
PASS / OPEN / CONFIRMED are metadata, not proof. They should not become repeated typographic punctuation that creates laboratory aesthetics independent of evidence.
MC-08 — Naming a weakness is not a correction
Pre-emptive self-criticism can become its own escape hatch.
A weakness named in prose is not corrected merely because it appears in the record. A meaningful correction should leave a consequence outside the sentence: changed code, method, specification, classification, public claim, test, external-review action, or explicit no-action rationale.
Selected confirmed findings used in the essay
Space Weather
Total failure of all public feeds can collapse to “Quiet signal” rather than unavailable/unknown.
Core principle: Unknown is not quiet.
Windowsill
Selected confirmed finding families include hemisphere/orientation scoring inconsistency; frost-state problems; week/month precedence and week-label issues; missing frost provenance; USDA-style hardiness based on the wrong temperature statistic; safety metadata/interaction-risk propagation gaps; documented rate-limit headers not emitted; report/PDF warning propagation failures; API/library/scoring version information dropped at client boundary; year-boundary week logic; valid zero coordinates/elevation mishandled; navigation-history-dependent report state; stale asynchronous responses contaminating newer state; and transport/API failure conflated with absent data in some client paths.
This is the audit-era finding set, not a claim that every listed manifestation was freshly reproduced against the current code. The bounded drift pass retains current support for the orientation, frost-state, week/month, week-label, frost-provenance, hardiness, structured-safety and report/PDF-warning paths. The remaining families require their own current-code or artifact pass.
Root causes must be consolidated before any public bug-count claim. Do not inflate manifestation count into root-cause count.
24 Doors
Selected confirmed findings: visibly broken Unicode/mojibake in the identified historical generated PDF; evidence/presentation contradiction around precise checked-looking result versus provisional validation status; semantic mismatch between measured language access and stronger trust framing; review-pack/native-review state mismatch with official-looking artifact.
The supplied eu27__ces-ell-eng__confident artifact was rendered and visibly broken. A different current local V2 final artifact has a different hash, language bundle and legible render. That distinction does not retroactively repair the earlier artifact, and it does not establish that the exact percentage was mathematically wrong.
bio-log
Public TID metadata says the vessel contains lake water. Current bio-log repository describes seawater.
Finding: public metadata drift.
ML Warmups Data
The primary “Open data pack” route redirects back to the same TID door.
Finding: circular primary navigation.
Anchiano
The legacy V1 TID surface says “Open source” while the local source plan remains private by design because the current product contains puzzle/secret-room mechanics. The V2 object record lacks that wording but has pre-existing local changes, so the current state is a public-surface split rather than a clean claim that the wording is already corrected everywhere.
Finding: premature public open-source claim. Private source itself is not the bug.
Important PASS findings retained
Breadth gates with no confirmed bug included: Eurovision Explorer normal country-history flow; The Other Scoreboard rank-error semantics; Route Gap Calculator accounting identity; Heat Pressure sample/proxy truth boundary; AI Labour Power Budget energy/load accounting; nearest-gravhøj nearest-point ranking; Danish/Swedish/Dutch/Belgian Politics Data correlation and exclusion contracts; nabour similarity ranking; Circuit Lab quantity conservation; Heat From Here Overpass deduplication; SINCE time/share semantics; Market Gaps evidence-maturity boundary; Objects in Time integrity-proof ≠ truth-proof.
These are not full-product validations.
OPEN items retained
Examples include Eurovision completeness/resilience under failed year fetch; nearest-gravhøj dataset-count metadata drift; nabour harmonisation/denominator review; TID Build Ledger freshness/source-of-record verification; Objects in Time JSON↔HTML parity; 24 Doors monotonicity/nested-language tests; and selected Windowsill environment-dependent verification items.
OPEN items should not be silently counted as bugs.
External Pressure 001 — Windowsill Hardiness
Objective: test whether Hedegreen Research can create a criticism request narrow enough to answer, cheap enough to inspect and consequential enough to change a system.
Packet should contain: current hardiness calculation; definition it was intended to approximate; relevant source material; identified mismatch; proposed corrected interpretation; one narrow question.
Question:
Does this calculation legitimately support the hardiness classification being claimed?
First recipient: USDA Agricultural Research Service — Plant Hardiness Zone Map contact
Channel: email
Address: phzminfo@usda.gov
Send-by date: 2026-08-27
The official USDA PHZM site currently lists phzminfo@usda.gov for Plant Hardiness Zone Map questions. Its methodology describes the map in terms of average annual extreme minimum winter temperature and the current 2023 map uses a 1991–2020 climate period.
Do not mark this action complete merely because the recipient and deadline are now specified.
Operational record must contain:
- recipient
- channel
- sent date
- packet/version sent
- response / no-response state
- resulting change, if any
A generic “eventually ask an expert” does not satisfy this item.
PASS reattack control
The Day 0.5 audit did not perform systematic second-pass attacks on prior PASS findings.
Future minimum experiment: choose at least one previous PASS; use a different invariant, reviewer or source path; attempt to falsify it; record whether it stays PASS, becomes OPEN or becomes CONFIRMED failure.
Compute / continuity hypothesis
Observed working hypothesis: greater model access appears to improve research continuity more than raw output volume.
This is not a controlled result. Variables changed during the period, including project complexity, user experience, workflow maturity, model access level, task mix and institutional infrastructure.
Safer wording: The scarce resource seemed to be continuity, not words.
Day 1.0 implications
Day 1.0 is not awarded by calendar date.
Target institutional capabilities include: article pipeline; Tool Builder pipeline; correction flow; Operations Panel continuity; recurring/annual publication handling; archive and provenance; physical source ingestion; physical field-research workflow; reopenable PASS controls; externally anchored important contracts; external-pressure experiments; and institution state stored in HR rather than only conversational context.
Working rule: 1.0 is earned, not scheduled.
Intermediate versions such as Day 0.65 remain legitimate if institutional state does not justify 1.0.
Companion-publication rule
Day 0.5 should publish as two linked objects:
- Main essay — readable institutional argument.
- Audit & Correction Record — technical companion holding findings, classification changes, criticism provenance, methodological revisions, unresolved tests and concrete next actions.
This separation prevents the correction process from swallowing the essay while preserving an inspectable trace.
Meta-rule
The central regress discovered during Day 0.5 is:
Who controls the control?
The answer is not infinite verification. The working direction is to make each important control sufficiently inspectable, temporally grounded and reopenable that the institution cannot silently place itself outside its own rules.
That boundary is not solved yet.