The record · iterations i01–i11
Research timeline
Every iteration, experiment, gate, and milestone — and, most tellingly, the 41 self-corrections where the refutation pass overturned the program's own conclusions before they stood. The honest record is the method.
Dots are colored by kind — ● experiment, ● milestone, ● gate, ● retraction. Entries marked ↺ self-correction are where a first-pass claim was narrowed, walked back, or retracted. Grades are A–D (glossary). A community research track is planned next — see the repo to contribute.
Decision locks signed off
Initial program decisions D1–D8, D10, D11 locked as L11–L20; firewall, harness, and stratification rules set.
Foundational program (Phase 0–3)
Built the corpus pipeline, validation harness, and annotation dataset; ran baseline morphology/topic/anchor studies and the first living synthesis.
Replication gate passed
35/38 scored published-baseline statistics reproduced (h2, MZ peak, glyph rules within tolerance); the replication firewall gate passed.
Annotation gate passed
227-page bulk annotation at schema v0.2-fine accepted; QA passed retuned thresholds; root_type flagged low-confidence.
Synthesis gate passed
First living-synthesis flagship accepted with graded claims.
Final i01 gate passed
Post-adversarial-review honest revision accepted; i02 agenda approved; i01 complete.
Sharpening the envelope by subtraction (E1–E5)
grade CWord-order/meaning-detector claims tested and corrected; the i01 'conlang best fit' was withdrawn under fair encoding-bracket tuning.
Is ΔI a meaning detector?
grade CΔI value is not a meaning certificate; a meaningless drift-null reaches the same ΔI at the wrong scale, so ΔI cannot settle meaningful-vs-meaningless.
Word-order signal confound
grade BReordering signal is intrinsic (block scale 812), not an artifact; the anti-cipher point was corrected — a deterministic type-preserving verbose cipher retains ΔI, reopening that cipher family.
Anchor-hunt power curve
grade BThe 'null informative' claim was oversold; the null excludes only strong balanced anchors, weakening the flagship anchor leg.
root↔leaf masked positive
grade CAn apparent cross-organ visual bundle; the coded 'confirmed' verdict was overclaimed and corrected to suggestive-but-unresolved after a cross-model test.
root↔leaf same-source confound test
grade COpus-4.8 re-annotation showed the association rides on Sonnet's root label; graded down B→C — leaned toward a labelling artifact.
Fair encoding bracket
grade CUnder equal tuning and held-out scoring, no encoding family is robustly distinguished; i01's 'conlang best fit' withdrawn, two bugs caught en route.
Reconstruction + variable-introduction agenda (E6–E10)
grade CHardened negatives (no family distinguished, no statistic localizes meaning) while reopening the root↔leaf candidate; excluded verbose cipher but revived the abjad class.
Verbose/nomenclator cipher reconstruction
grade CVerbose cipher deductively excluded on the ED1 morphology network; an abjad of real Latin reaches it, refuting the first-pass 'needs constructed morphology' and reviving the abjad class.
Fine-granularity anchor hunt
grade C18 raw anchor 'discoveries' were caught by the permutation null (p=0.48); no anchor above chance at higher power. Anti-labelled-herbal null holds.
Whitened encoding bracket
grade CThe whitened ranking is a regularisation artifact (ill-conditioned covariance); no family stably distinguished. The 'whitening confirms E5' claim was withdrawn.
VMS-coordinate meaning localization
grade DAfter two refutations of trivial endpoints, no structural axis localizes meaning — the structured-meaningless adversary out-scores meaningful text and the VMS sits below both; a pre-registered commitment was retracted.
root↔leaf third vision-model rater
grade CHaiku 4.5 also reproduced the leaf association, overturning E4b's 'Sonnet-specific artifact'; root↔leaf reopened as strongest referential candidate but unconfirmed.
Post-i02/i03 synthesis accepted
Tim signed off the honest post-review synthesis (E1–E10 folded in); L37 locked and forward direction set.
E10 confirm-or-kill (E11–E12)
grade CThe root↔leaf bundle survived a style control but the independence test showed model annotation is underpowered/untestable; E10's positive candidate withdrawn to UNRESOLVED.
Illustration-style control for root↔leaf
grade CThe bundle survives conditioning on palette-richness style for both models; not a palette-style artifact (partial control only).
Independent-lineage (GPT-5.1) root rater
grade CA non-Anthropic rater does not reproduce the association; a first-pass 'KILLED' was corrected to UNRESOLVED-underpowered — model annotation cannot adjudicate, only a human panel can.
Mid-level linguistic program (E13–E17)
grade CBuilt a null-correction framework; found the VMS lacks the surface content>function collocation gap and has weak word-class structure, with A/B differing lexically not grammatically.
Function/content bimodality probe
grade DThree global operationalisations failed calibration; no VMS verdict issued (harness-first). Prompted the E13b redesign.
Function/content, order-shuffle null
grade CCalibrated; narrowed to a surface finding — the VMS has near-chance collocation and its most-frequent words are template-like, the opposite of flat function words. First substantive i05 result.
Distributional word-class induction
grade CNull-corrected POS induction shows the VMS has only weak word-class structure (~0.13–0.19× real language), calibrated on real languages.
Content-controlled A-vs-B contrast
grade CThe apparent A/B word-class difference vanishes within the herbal section — it is a section/content confound, not dialect; corrects E14b as E12 corrected E10.
Paper v1 preprint
First firewall-sourced MS408 preprint packaged via the package-paper skill (through E12).
Paper v2 (folds i05)
Preprint updated to fold in the i05 mid-level linguistic program.
Cryptanalytic direction (E18–E20)
grade BConcluded the cipher-of-real-prose class is EXCLUDED via the joint low-entropy + retained-ΔI + weak-syntax signature; this universal headline was later retracted by i11.
Joint-signature cipher test
grade BNo cipher of real prose matches the VMS's low-h2 + weak-syntax combination; the negative is grade B and the circular 'favours generation' positive was dropped.
Language-universality control
grade BThe exclusion holds across typologically diverse languages including Hebrew (a native abjad); an order-scrambling transposition cipher is the only surviving weak-syntax lead.
Transposition closure
grade BRetained-ΔI and weak-syntax are mutually exclusive under any word-reordering cipher, yet the VMS has both — closing the cipher-of-real-prose class (headline later retracted).
Paper v3 (folds i06)
Preprint updated to fold in the i06 cryptanalytic direction (carried the later-retracted cipher exclusion headline).
Characterising the generative class (E21–E22)
grade CA context-free positional/template generator matches entropy + block-ΔI + weak-positive syntax but never morphology connectivity, lexical reuse, or frequency slope — a partial account only.
Minimal positional generator + ablation
grade CA first-pass 'class sufficiency [B]' was overturned by the refutation pass — constants were grid-selected to the VMS bands and the weak-syntax leg was shuffle-passable — narrowed to an informative negative.
Genericity/coupling sweep
grade CA broad a-priori grid never reaches the VMS's ED1, TTR, or Zipf; the context-free bag-of-slots is INSUFFICIENT and demands an added reuse/smaller-lexicon mechanism.
Adding frequency concentration (E23–E24)
grade CAcross three generative families, none reproduces the full 8-axis signature over the swept ranges; the summary statistics are mutually coupled. This hard constraint was later walked back by i09.
Paper v4 (folds i07–i08)
Preprint folds the generative-class characterisation with refutation-corrected wording.
Symbols-as-values direction (E27–E28)
grade DThe quantitative-register hypothesis finds no support and closes at grade D — a positional-numeral sub-type is shape-excluded and no robust ordinal anchor appears in the zodiac rings.
Symbol quantification
grade DThe VMS's strong positional specialisation (0.74) shape-excludes a positional-numeral sub-type; a non-positional value scheme remains untouched (→ E28).
Angular/ordinal anchor in circular diagrams
grade DA sensitivity-confirmed Mantel test on all 12 zodiac rings finds no robust ordinal structure; the lone f73r hit is not reproduced — register stays D.
Is the i08 coupling real? (E25–E26)
grade CWith ED1 made an independent knob, i08's gross incompatibility collapses to a shallow ~0.03 near-miss frontier — a retraction of the i08 hard constraint, not a promotion to 'no constraint'.
Decoupled-ED1 type-lexicon generator
grade CMaking ED1 an independent knob shows E24's saturation was a small-char-space artifact; the coupling largely dissolves to a shallow h2↔ED1 near-miss (initial over-read later curbed).
Word-length variance vs the h2↔ED1 frontier
grade CWord-length variance lands ED1 in-band jointly with ΔI/TTR/Zipf; the residual is a ~0.03 h2 near-miss plus a fixable length artifact — the frontier is all-but-crossed.
Paper v5 then v5b
Preprint folds i09; v5's stronger 'signature doesn't constrain mechanism' phrasing was corrected by a refutation of the walk-back itself to v5b.
Methods paper v1 then v2
Packaged the transferable adversarial-self-correction methodology paper; v1's inferential over-claims were corrected to v2 by its own refutation pass, and the refutation archive was created.
Naibbe cipher vs the i06 discriminators
grade CA first pass read 'i06 confirmed [B]'; the refutation killed it — word-boundary Latin sits in the VMS ΔI band and ~82% of Naibbe's ΔI loss is respacing, exposing that ΔI is a homophony detector and weakening i06's ΔI leg.
Cipher exclusion re-examination (multi-seed)
grade COn word-boundary Latin: order-preserving ciphers robustly EXCLUDED (strong syntax), but verbose+homophonic (≈ Naibbe) NOT excluded — i06's universal cipher headline RETRACTED, robust core kept.
Harden the fc_z/wc_z syntax discriminators
grade BDeconfounded syntax measures firm the order-preserving exclusion (6.8σ/8.0σ after v6b correction of a fabricated ~30σ); the homophonic class stays inconclusive. A recalled-number firewall slip was caught and corrected.
i06 universal cipher headline retracted
The 'cipher-of-real-prose EXCLUDED' headline (papers v3–v5b) is retracted to 'order-preserving ciphers excluded; verbose+homophonic (Naibbe-class) not excluded' — the VMS-as-cipher hypothesis is viable on the program's own analysis.
Paper v6 then v6b + methods v3
Paper v6 retracts the i06 cipher exclusion; its own refutation caught a fabricated ~30σ and over-correction → v6b, and methods v3 adds the i06/Naibbe self-correction example.
Open-source release Tier 0
Repo packaged as a public evaluator: ms408.evaluate() API, firewall-clean reference bands, Apache-2.0 license, docs; 174 tests pass. A subsample-without-replacement bug fix put the VMS back inside its own bands.
Engaging concurrent work (Naibbe/Parisel) + release
grade BThe make-or-break Naibbe engagement exposed and corrected i06's ΔI-leg confound and retracted the universal cipher headline; also shipped the open-source release tiers and arXiv bundles.
Block-scale like-for-like ΔI
grade CA first pass wrongly found a homophonic cipher reaching the VMS corner ('ΔI leg dead'); the refutation showed the corner was a homophony-marker h2 artifact — under a fair model no config reaches it, so block-scale ΔI weakly separates verbose+homophonic ciphers.
Open-source release Tier 1
Value-pinning tests, a Naibbe worked example, a reproduce-the-paper --verify path, and CI added to the public evaluator.
Open-source release Tier 2
Hardened CI, CONTRIBUTING, tutorial, and wired docs; Tier-2 packaging marked done.
arXiv submission bundles prepared
Build/check arXiv submission bundles prepared for paper v6b and methods v3 (no submission yet).
Paper v7
Preprint folds the E33 block-scale ΔI result and announces the open-source release.
Public GitHub Pages site
Scoped and shipped the Astro v1 public site (docs + papers), later extended with a library/education portal and media feed.
H2-2026 publication planning package
Venue landscape, scoring rubric, and four pathways for placing the constraint-envelope (Paper A, v7) and methods (Paper B, v3) papers; every venue rejects perceived 'solutions' and welcomes the hypothesis-shrinking posture.