The instruments · an ongoing inquiry
How does an experience change what happens next?
A thought moves between people, models, sessions, and summaries. The words can survive while the reason for them disappears.
Then somebody has to rebuild the understanding. That costs more than time: it uses the attention that was holding the next thought.
This is one practical edge of a wider question. What survives when something changes form—a drawing into a graph, music into events, an encounter into memory? What gets lost, what gets added, and how does the result change what its next observer can notice or do?
Wink is an attempt to make those transformations inspectable. The work around it explores memory, changing attention, and ways to investigate a question without letting the first interpretation become the only one available.
What the attempts have taught us
The useful findings are specific. These are small experiments and audits, with enough of their history left attached to question the account.
The observer is part of the measurement.
Wink initially compared the structure of an SVG document with the structure of the scene it depicted. The mismatch belonged to the comparison. Read the sense-calibration experiment.
An experience can steer later attention.
Riverbed tests found that changing an external field’s weights changed where a local model attached later material. Repeated runs also overturned several appealing explanations of how particular histories shaped that field. Both results matter. Read the findings and retractions.
Preserving a summary is a choice about the future.
In one compaction experiment, different compressions retained different relations. Its correction matters too: missing reference structure had been counted as invention. The comparison could not establish that verdict. Read the corrected result.
The question still open
What would let an encounter change later attention, interpretation, and action—across changes of representation or observer—while keeping those changes open to correction?
Remembering an answer is one part. Carrying forward why a question mattered, what changed our minds, and where disagreement remains is harder. The test is whether the next interaction can continue the inquiry, and whether the person gets more room to think.
The records
Different instruments approach different parts of the question. Each collection below opens into readable documents and their raw text.
Kieran shapes this research with AI collaborators. The new working note and reading guide were written with Codex on 22 October 2026. The experimental records remain dated snapshots from 15 October; “I” in them usually means the model writing the record. Corrections are part of those records.
the question and nearby research
A current route through the problem, the attempts, and the literature. Read this first if the project names are new to you.
2 documents · the question and nearby research
- What we are trying to carry forward working-note.md raw
- Memory, learning, and the observer: an annotated reading guide reading.md raw
wink
What survives a change of representation, and what does the change do to the next observer? Experiments across prose, graphs, events, diagrams, and audio, with the basis of comparison kept in view.
38 documents · wink
- Research state: the scorecard, and the standing failure-mode watch (as of 2026-09-01) RESEARCH_STATE.md raw
- Open questions, with Kieran's rulings verbatim OPEN_QUESTIONS.md raw
- Lineage: what each ancestor system preserved, destroyed and invented LINEAGE.md raw
- The experiment protocol: predictions frozen before a run, scored after EXPERIMENTS.md raw
- EXP0 Parse coexistence: disagreement report EXP0-parse-coexistence/DISAGREEMENT_REPORT.md raw
- EXP0 Parse coexistence: frozen predictions EXP0-parse-coexistence/predictions.json raw
- EXP0 Parse coexistence: spec EXP0-parse-coexistence/spec.json raw
- EXP1 Nursery: results EXP1-nursery/RESULTS.md raw
- EXP1 Nursery: frozen predictions EXP1-nursery/predictions.json raw
- EXP1 Nursery: spec EXP1-nursery/spec.json raw
- EXP3 Senses: results EXP3-senses/RESULTS.md raw
- EXP3 Senses: frozen predictions EXP3-senses/predictions.json raw
- EXP3 Senses: spec EXP3-senses/spec.json raw
- EXP4 Reentrance: results EXP4-reentrance/RESULTS.md raw
- EXP4 Reentrance: frozen predictions EXP4-reentrance/predictions.json raw
- EXP4 Reentrance: spec EXP4-reentrance/spec.json raw
- EXP7 Cold atom: results EXP7-cold-atom/RESULTS.md raw
- EXP7 Cold atom: frozen predictions EXP7-cold-atom/predictions.json raw
- EXP7 Cold atom: spec EXP7-cold-atom/spec.json raw
- EXP8 Solar rehearsal: results EXP8-solar-rehearsal/RESULTS.md raw
- EXP8 Solar rehearsal: frozen predictions EXP8-solar-rehearsal/predictions.json raw
- EXP8 Solar rehearsal: spec EXP8-solar-rehearsal/spec.json raw
- EXP9 Volunteered taste: results EXP9-volunteered-taste/RESULTS.md raw
- EXP9 Volunteered taste: frozen predictions EXP9-volunteered-taste/predictions.json raw
- EXP9 Volunteered taste: spec EXP9-volunteered-taste/spec.json raw
- EXP10 Compaction bill: the episode, the three compressions, the ground truth, and the provenance of the correction (two commits, 33 minutes apart) EXP10-compaction-bill/EPISODE.md raw
- EXP10 Compaction bill: results EXP10-compaction-bill/RESULTS.md raw
- EXP10 Compaction bill: frozen predictions EXP10-compaction-bill/predictions.json raw
- EXP10 Compaction bill: raw result (the "invented" column here is the uncorrected, truth-side-missed count; see the correction in RESULTS.md) EXP10-compaction-bill/raw_result.json raw
- EXP10 Compaction bill: spec EXP10-compaction-bill/spec.json raw
- ADR 0000 bootstrap design note ADR/0000-bootstrap-design-note.md raw
- ADR 0001 boring classical comparison stack ADR/0001-boring-classical-comparison-stack.md raw
- ADR 0002 researcher rulings on the five questions ADR/0002-researcher-rulings-on-the-five-questions.md raw
- ADR 0003 the nursery boot curriculum ADR/0003-the-nursery-boot-curriculum.md raw
- ADR 0004 first instrument inspection ADR/0004-first-instrument-inspection.md raw
- ADR 0005 nonlinguistic core and the qualia stake ADR/0005-nonlinguistic-core-and-the-qualia-stake.md raw
- ADR 0006 hanzi nursery and the childs two memories ADR/0006-hanzi-nursery-and-the-childs-two-memories.md raw
- The constitution, 3,319 lines: shaped by Kieran, written by a model; Part III is the failure-mode register WINKS_CONSTITUTION.md raw
chronology
A bounded application: compare historical astronomical records with computed sky events, inspect calendar assumptions and cyclic alternatives, and retain the corrections alongside the fits.
43 documents · chronology
- The lab notebook: three tests, the cyclic objection and its retraction, fit maps, the spin measured from the records, corrections inside the primaries README.md raw
- Slips by segment: no slip beats its null; pinned and unique are different things slips_v0_notes.md raw
- "This is cyclic": the provenance of one retraction. Two commits fourteen minutes apart, the objection verbatim, before and after, and the bundle to rerun the eclipse grids with no model involved cyclic/CYCLIC.md raw
- The Lu section as committed before the objection (f166f57, 2026-09-08 13:26) cyclic/README-before-the-objection.md raw
- The retraction commit, as a diff (3b49e55, 2026-09-08 13:40) cyclic/retraction-commit-3b49e55.diff raw
- The three commits around it, with timestamps cyclic/commits.txt raw
- The 37 eclipse entries of the 春秋 jing, as parsed cyclic/chunqiu_eclipses.json raw
- The fourteen wikisource pages the records came from: revision ids and hashes cyclic/manifest-wikisource-chunqiu.jsonl raw
- The alignment search, as published (Zhou month with one month of slack) cyclic/align.py raw
- The same script with one line changed: exact Zhou month (the cycle-proof test) cyclic/align_exact.py raw
- Rerun 2026-10-15, as published: 722 BCE 21, aliases 14 and 15 cyclic/rerun-2026-10-15-month-slack.json raw
- Rerun 2026-10-15, exact month: 722 BCE 14, aliases 2 and 2, best other 6 cyclic/rerun-2026-10-15-exact-month.json raw
- The folder contract the script imports cyclic/paths.py raw
- The wikisource fetcher (contact address replaced) cyclic/src_wikisource.py raw
- The pinned environment (astronomy-engine 2.1.19) cyclic/requirements-venv.txt raw
- Output as written on 2026-09-08: the local-visibility grid cyclic/align_chunqiu_local.json raw
- Output as written on 2026-09-08: the per-record misses cyclic/align_chunqiu_local.txt raw
- Output as written on 2026-09-08: the cycle-proof combination cyclic/align_combined.json raw
- derived/align_almagest.txt derived/align_almagest.txt raw
- derived/align_asc.txt derived/align_asc.txt raw
- derived/align_chunqiu_global.txt derived/align_chunqiu_global.txt raw
- derived/align_chunqiu_local.txt derived/align_chunqiu_local.txt raw
- derived/align_precession.txt derived/align_precession.txt raw
- derived/anchor_lu.txt derived/anchor_lu.txt raw
- derived/anchor_ptolemy.txt derived/anchor_ptolemy.txt raw
- derived/anchor_strict.txt derived/anchor_strict.txt raw
- derived/continuity_west.txt derived/continuity_west.txt raw
- derived/custody_islands.txt derived/custody_islands.txt raw
- derived/fitmap.txt derived/fitmap.txt raw
- derived/irish_521.txt derived/irish_521.txt raw
- derived/islands.txt derived/islands.txt raw
- derived/landscape_west.txt derived/landscape_west.txt raw
- derived/one_earth_hr.txt derived/one_earth_hr.txt raw
- derived/one_earth_hr_firm.txt derived/one_earth_hr_firm.txt raw
- derived/one_earth_vis.txt derived/one_earth_vis.txt raw
- derived/one_earth_vis_firm.txt derived/one_earth_vis_firm.txt raw
- derived/pins_review.txt derived/pins_review.txt raw
- derived/precision.txt derived/precision.txt raw
- derived/slips_west.txt derived/slips_west.txt raw
- derived/slips_west_deltat.txt derived/slips_west_deltat.txt raw
- derived/slips_west_deltat_split_asc890.txt derived/slips_west_deltat_split_asc890.txt raw
- derived/tables_fingerprint.txt derived/tables_fingerprint.txt raw
- derived/witnesses.txt derived/witnesses.txt raw
the evidence method
research_projects applies the Wink inquiry to a specific question: retain sources, locate claims, record prior interpretations, and make the comparison inspectable. Published examples concern Tartaria and an Erie Canal control.
6 documents · the evidence method
- The method: a shared and independently evaluatable outside; the rules research_projects-README.md raw
- The folder contract: a fetch is a noticing from a position tartaria-README.md raw
- The model's frozen beliefs, before any corpus (claude-fable-5-1, 2026-09-06) claude-2026-09-06.md raw
- Kieran's thirty entry points, verbatim, with the prompts that fetched evidence for them (the first seventeen are the ones the model's priors were frozen on) tartaria-entrypoints.json raw
- What the folder holds: counts only tartaria-counts.md raw
- The control: a subject with no pull on him, gathered the same way erie-canal-README.md raw
riverbed
Persistent memory outside the model. The model reads a weighted field before every turn and proposes; code decides what gets weight and checks the evidence against the source. Weight is navigational salience, never truth.
5 documents · riverbed
- Kieran's intent, verbatim and immutable INTENT.md raw
- Findings: pre-registered results with test statistics, and the retractions (a "fabrication" verdict withdrawn because the ground truth was stipulated; five runs of one condition gave five different hubs) FINDINGS.md raw
- Regimes: new channels change what the model attends to, which changes what the channel later measures REGIMES.md raw
- How it works README.md raw
- The experiment log experiments-README.md raw
anamnesis
An exploration of a persistent system around a language model: storage, drives, sleep, and salience. The published audits examine what its autonomous heartbeat actually composed and delivered.
2 documents · anamnesis
- Independent agent heartbeat audit (2026-03-04) 2026-03-04-independent-agent-heartbeat-audit.md raw
- Heartbeat message formation audit (2026-03-06): scaffold text shipped as if it were the agent's own words 2026-03-06-heartbeat-message-formation-audit.md raw
Opened 15 October 2026 · introduction revised 22 October 2026.
Also here: the model lab · the broader project ledger · Sapien Logos.