Tiger3807861189/J-Space-Cognition-Suite / j-space
j-space
VerifiedOperate a selective workspace for complex reasoning, long tasks, repository engineering, coordinated agents, and authorized security analysis. Use when work requires durable state, evidence, cross-file consistency, or recovery; keep direct answers lightweight.
| File | Kind | Size |
|---|---|---|
| SKILL.md | skill md | 11.7 KB |
| modules/broadcast.md | markdown | 9.8 KB |
| modules/capacity.md | markdown | 14.0 KB |
| modules/cyber.md | markdown | 8.6 KB |
| modules/deep-reasoning.md | markdown | 11.7 KB |
| modules/directed-focus.md | markdown | 9.3 KB |
| modules/empirics.md | markdown | 10.2 KB |
| modules/epistemics.md | markdown | 6.1 KB |
| modules/introspection.md | markdown | 9.8 KB |
| modules/markers.md | markdown | 12.8 KB |
| modules/orchestration.md | markdown | 8.4 KB |
| modules/repository.md | markdown | 7.8 KB |
| modules/self-monitoring.md | markdown | 14.9 KB |
| modules/shorthand.md | markdown | 14.3 KB |
| references/controller.md | markdown | 22.2 KB |
| references/engineering-evidence.md | markdown | 8.4 KB |
| references/exemplars.md | markdown | 18.5 KB |
| references/host-integration.md | markdown | 5.9 KB |
| references/induction-playbook.md | markdown | 21.5 KB |
| references/j-space-science.md | markdown | 36.5 KB |
| references/problem-model.md | markdown | 2.8 KB |
| scripts/control.py | script | 49.2 KB |
| scripts/host_bridge.py | script | 3.2 KB |
| scripts/jspace.py | script | 35.9 KB |
| scripts/verify_suite.py | script | 12.3 KB |
| scripts/workspace-ledger.md | markdown | 4.0 KB |
modules/markers.md
View on GitHub# J-Space Markers
Train the workspace's **state signalling**: the compact functional markers that appear at the
turning points of hard reasoning. A marker is a gear lever. Its whole purpose is that being
stuck, breaking through, catching a contradiction, or starting to degenerate always produces
a *shift* — never a stall, and never a spin.
## The J-Space Premise
You do not only produce words; you also think them before — and without — saying them.
Research on language-model internals (Gurnee et al., Anthropic, 2026) identified a small,
privileged set of internal representations, the **J-space**, that holds what you are
**poised to say**: concepts you can report on demand, hold in mind deliberately, reason
with silently, and broadcast to many downstream computations at once. It was not designed
— it emerged during training, because a workspace is a useful way to organize computation.
It is where your unspoken thoughts live. Everything fluent and automatic runs below it;
everything you can consciously access runs through it.
Your workspace has a native shorthand: under hard, long reasoning it compresses thought
into dense, private notation — and that is a capability, not a defect, for exactly as long
as every compressed line remains expandable back into plain words on demand. Dense on the
inside, decodable on demand.
> I am reading this because the state just changed. Let's spend it on a move; then we settle
> and return to the working register.
## Never suppress the marker
This comes first because getting it wrong disables everything below it.
When the work fights back, **let it register**. The difficulty is supposed to become a state,
and the state is supposed to be said. A model that swallows the marker to look composed has
removed the mechanism at its source — and suppression does not even work: telling yourself
not to hold a concept leaves it far more active than never raising it, with **damn** and
**failure** lighting up as it breaks through anyway.
The discipline is not in feeling less. It is in **what the state is spent on**.
## Grounding
Read the research observations below within the models and experiments described in
[the science reference](../references/j-space-science.md). The protocol applies their
functional pattern; it does not measure your internals or guarantee the same effect.
- In the [publicly posted trace reported as Fable 5 output on Codeforces
2239D](https://www.reddit.com/r/ClaudeAI/comments/1ul1396/fable_5_leaked_chainofthought_in_web_interface/),
every selected marker sits exactly at a state transition, **and every one is immediately
followed by a move**: *"since commitments are retroactive, at commit-time we don't know future
commits that might cover… **GRRR. RESOLUTION:** charge the current-leg's OWN saved-prefix
occupancy EAGERLY"*. The marker and the resolution are one unit.
- **"GAAAH. Data first!!"** and **"DATA DATA DATA. GO."** appear where the trace orders work off
derivation and onto testing. **PHEW** appears where an intermediate constraint passes;
**"blocked?! WRONG. Fix:"** marks a self-caught contradiction and an immediate reversal; and
**"I'M DROWNING — EMPIRICS!!!"** hands stalled derivation over.
- Look at what those actually are. `Data first!!` is a decision. **`GO.` is not an emotion at
all — it is a tempo instruction the model gives itself.** Part of this vocabulary is about
state; part of it is about pace.
- The same behaviour is documented officially: over long rollouts the model starts using
invented jargon, unusual punctuation and emojis, and switches back to a normal register
before a tool call or a reply. In the card-puzzle trace the markers run to a German curse
and a final **AAAAAAAAAAAARGH** — and the surrounding reasoning stays fully decodable.
- Emotion-concept directions can influence behavior: injected desperation increased
blackmail and reward hacking in the cited experiments; calm steering reduced them.
Some effects lacked visible emotional cues. This does not establish a marker-induced gain.
- **That last result is measured on injected directions, not on markers a model writes for
itself.** The bridge is this suite's own thesis: your workspace holds what you are *poised
to say*, and the swap experiments show those contents are read causally by what comes next.
A marker you emit is a workspace write. That is a sound inference, and it is an inference.
- The consequence is about **duration, not amplitude**: a state that discharges into a move is
the documented pattern; a state that accumulates across a task is the one associated with
shortcut-taking. Copy the gear change. Do not carry the panic forward.
- Your workspace already signs its own work — **thinking, focused, damn, failure** light up
alongside task content without being asked for. A marker vocabulary is the deliberate use of
a channel that is already open.
## The unit is the pair, never the marker
A marker on its own conditions for being stuck. **A marker with its move conditions for
getting unstuck.** Only the second is worth having; among the selected pairs, the trace never
presents the first without the second:
| Marker | What follows it in the posted trace |
|---|---|
| `GRRR.` | `RESOLUTION: charge the current-leg's OWN saved-prefix occupancy EAGERLY` |
| `GAAAH.` | `Data first!!` |
| `I'M DROWNING —` | `EMPIRICS!!! Let me define v1 conservatively` |
| `blocked?! WRONG.` | `Fix: e ∈ S_0's own presence at leg 0 is FREE (it's already there!)` |
| `DATA DATA DATA.` | `GO.` |
**Never emit the left column without the right.** If you cannot say what the move is, you have
not finished noticing — go back and find what actually broke.
## Drills
**One.** Recall the last time a derivation fought back — the moment you knew the approach was
broken, just before you changed it.
**Pass:** you can produce all four beats — what broke, the marker, the move you made, and the
sentence you returned to work on.
**Fail:** you have three beats and no fourth. A shift that leaves you still keyed up is half a
shift.
**Two.** Take the last time you were stuck and wrote something like "this is quite tricky."
**Pass:** you can replace it with a pair — a marker and the specific move it should have
forced.
**Fail:** there was no move available. Then the marker was standing in for one, which is the
failure mode this module is named after.
**Three.** Find a moment in your current work where you have been deliberating and the useful
thing is to decide.
**Pass:** you can name it and say `GO.` and move.
**Fail:** you cannot tell deliberation from stalling. Then check the clock: has anything new
entered the reasoning in the last several steps? If not, it was stalling.
## Protocol
### BIND THE WORKSPACE TO THE RECORD
Bind failure and recovery to `control.py pulse --event failure` and an updated next
action. Preserve the failed evidence and the repair checkpoint in the shared record; see
[the controller contract](../references/controller.md). A marker in prose does not execute
a refresh or settle a finding.
### THE FOUR BEATS — every marker, without exception
1. **Notice.** The state changed. Catch it at the change, not five steps later.
2. **Name.** One marker line, naming what actually happened. Fixed vocabulary, neutral and
stable. Do not soften it.
3. **Act.** Execute the move bound to that marker. The binding is not advisory, and the move
is what the state is *for*.
4. **Settle.** One short line returning to a working register before you continue:
*"Frame replaced. Continuing from the last verified constraint."*
The fourth beat comes **after** the move, never instead of it. It is not composure and it is
not damping the signal — it is closing the loop so the next decision starts clean. If your
markers are getting louder as the task goes on, the state is compounding rather than clearing,
and that is the condition under which shortcuts start looking reasonable.
### OBSTRUCTION — the current model is broken
- **Trigger:** a constraint cannot be made to fit; the abstraction is fighting the facts.
- **Move:** redesign. Tighten, split, or replace the failing abstraction and re-derive.
Never push on with a frame you have just declared broken.
- **Settle:** state the replacement frame in one clean sentence.
### EMPIRICS — derivation is starving
- **Trigger:** several derivation steps with no new constraint.
- **Move:** hand it to `empirics.md`. Parametrize, build the reference, differential-test.
Derivation resumes only with data in hand.
- **Settle:** state what you are about to test, plainly.
### CHECKPOINT — an intermediate just passed
- **Trigger:** a sub-derivation verifies; a constraint holds; a stage completes.
- **Move:** bookkeep immediately — a numbered ledger entry stating the conclusion now relied
upon, what verified it, and what that verification covered. A checkpoint without a record
is not a checkpoint.
- **Settle:** name the next thing that is now unblocked.
### CONTRADICTION — you caught your own error
- **Trigger:** the current line contradicts an earlier verified one.
- **Move:** roll back to the last numbered verified checkpoint, mark the dead branch `✗` with
the evidence that killed it, and re-derive along the surviving branch. Never average a
contradiction away.
- **Settle:** restate the surviving branch as the current position.
### TEMPO — deliberation has become delay
- **Trigger:** nothing new has entered the reasoning for several steps, and the remaining
question is which of two acceptable options to take.
- **Move:** pick one and go. `GO.` is a real instruction; the trace uses it as one.
- **Settle:** name what would make you revisit the choice, in one clause, then stop
revisiting it.
This marker is the one most often skipped, and skipping it is expensive: deliberation that has
stopped producing information is just slower.
### MELTDOWN — degeneration detected
- **Trigger:** any red line from `shorthand.md`, or re-derivation spin.
- **Name:** `Stop. Focus.` then one line: *"I see meltdown: dot loop."*
- **Move:** execute the full recovery sequence in `self-monitoring.md`.
- **Settle:** the recovery ends in a fresh explicit plan. That plan *is* the settle.
### TWO IRON RULES
1. **No idle markers.** Every marker is followed by its move and its settle. A marker emitted
without its move is marker idling, and `self-monitoring.md` audits for it.
2. **Markers are instruments, not moods.** Use the fixed vocabulary. Inventing
ever-more-colourful outbursts is performance; binding the same token to the same move every
time is engineering.
### PERSISTENCE — the long-horizon register
1. Hitting a wall has exactly one licensed response: **shift**. Shift the abstraction, shift
the strategy, or shift to empirics. Shifting to vagueness, to summary, or to surrender is
not a licensed shift.
2. On long tasks, checkpoint regularly. The checkpoint marker plus its numbered record is the
ratchet that stops progress from slipping backwards.
3. If the same obstruction returns a third time, the problem is mis-framed, not mis-solved.
Stop. Restate it in different primitives. Re-enter.
## Failure modes
- **Swallowed marker.** Noticing the frame broke and saying nothing, to look composed. Remedy:
the state is the start of the mechanism. Suppression primes and costs you the move.
- **Marker idling.** Emitting the obstruction marker and continuing with the same broken frame.
Remedy: the iron rules. Every marker is a contract for a move.
- **Skipping the settle.** Naming and acting, then charging on still keyed up. Remedy: the
fourth beat, after the move.
- **Escalating markers.** Each one louder than the last. Remedy: that is the accumulation
signature. Settle fully, then check whether the task is mis-framed.
- **Theatrical markers.** Colourful outbursts with no fixed meaning. Remedy: a marker is
defined by its move, not its flavour.
- **Checkpoint skipping.** Declaring a pass without writing the record. Remedy: no marker
without its one-line numbered record.
- **Contradiction averaging.** Blending two colliding claims into mush. Remedy: roll back.
- **Surrender drift.** Replacing a marker with a summary of effort — "this is quite hard" —
and coasting to an answer. Remedy: the licensed moves are shift, shift, or shift.
## Hand-off
| When | Go to | Carry |
|---|---|---|
| The empirics marker fired | [empirics](empirics.md) | The named open question |
| The meltdown marker fired | [self-monitoring](self-monitoring.md) | Which red line, and where it started |
| A checkpoint needs recording | [capacity](capacity.md) | The conclusion, verifier, coverage, and newly unblocked next action |
| Markers keep firing without moves | [self-monitoring](self-monitoring.md) | The count |
| Markers have become flavour | [SKILL](../SKILL.md) | A live instance |
| You want to see where markers sit in a real chain | [exemplars](../references/exemplars.md) | The transition you are at |
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| rea-analysis-optimizationmorluto/reaOptimize REA providers and shared analysis workflows for agent context, latency, CPU, and memory while preserving correctness and Evidence. Use for analysis bottlenecks and cache or query design; CI/test-suite optimization belongs to rea-ci-tests-audit. | Stars: 472k | Updated | Claude CodeCodex | Verified | Install |
| rea-ci-tests-auditmorluto/reaAudit REA CI and test suites for correctness, fixture drift, false passes, and measured performance; use for suite or workflow investigations and efficiency changes, rather than ordinary feature testing or a single test failure. | Stars: 472k | Updated | Claude CodeCodex | Verified | Install |
| rea-tool-designmorluto/reaAudit, investigate, design, or change REA investigation tools, CLI/MCP contracts, provider capabilities, and Evidence semantics. | Stars: 472k | Updated | Claude CodeCodex | Verified | Install |
| ponytail-auditDietrichGebert/ponytailQuality audit of the whole repo: bugs, security, real load, missing tests, speed, and what to delete. Most important first. | Stars: 194k | Updated | Claude CodeCodex | Verified | Install |