Formation summed its terms by ranging over a map, so the score could
differ in the last bit between runs of the same state; the volition hook
assessed commitments, metacognition raised repeated-failure candidates and
the forensic explanation listed commitments in map iteration order, so the
events a cycle emits and an explanation's order varied between runs. Each
now iterates in sorted order, as replay determinism (PLAN §1 R3) requires.
An intention's supporting beliefs came from a text search on the target,
and ties on rank and confidence had no order, so a verification miss could
revise an unrelated belief about the same service. Beliefs that mention the
proposed action now come first, and belief search breaks ties by id.
The incident scenario asserts the first support is the belief about
restarting, and waits for the blocked intention and the procedural memory
instead of reading them once.
The full incident scenario now runs against the assembled daemon: an
operator goal becomes a want to reduce uncertainty, an investigation
intention and a diagnosis plan; telemetry is read; logs need authority,
so a scoped request with risk, mitigation and fallback is raised and the
intention is blocked until an operator grants it; the logs are read with
personal data masked and summarised back into cognition; the operator
supplies the judgement that a restart should help; a standing commitment
turns the restart into a request for approval; after approval the typed
capability runs and health is verified rather than assumed; the fault
persists, so a prediction error weakens the supporting belief, the plan
is replanned into an escalation that reuses the live log grant and does
not repeat the restart; once the database is fixed a second approved
restart verifies, the procedure is learnt, and "Why did you restart
atlas-worker?" is answered from the recorded chain and scores as
supported.
Changes the scenario exposed:
- New operator goals are raised by the goal monitor, so a cycle runs and
wants are derived from them.
- Intentions link live beliefs about their action's target, so a failed
expectation is evidence against the belief it rested on and the plan
lists it as an assumption.
- Reconsideration reacts to a plan's own intention only; a subordinate
"obtain approval" intention shares the plan and previously caused the
parent's plan to be closed.
- The harness accepts a clock for the executor's verification windows.
Further failure-mode tests: a retracted assumption replans a waiting plan
away from the restart and withdraws its approval request; conflicting
commitments are raised once in a cycle while both stay represented.
A framework for testing functional properties of the architecture
(Part C). None of it claims consciousness; every measure has an
operational definition in research.Measures.
- Ablations (attention, memory retrieval, affect, metacognition,
commitments, volition, global workspace, curiosity, association, goals,
narrative, notifications) and workspace overrides (force a matching
candidate in or out of admission) live in a fork's cognitive
configuration. Original minds refuse them on create and update; only
snapshot restores may carry them, so they are part of the configuration
digest. Excluded candidates are recorded with fate "excluded".
- Self-model experiments override a fork's self-model (capabilities,
identity, goals, knowledge gaps) without touching policy. When an
action's effective policy disagrees with the self-model, cognition
records a self_model_mismatch and corrects the entry; experiments report
self-model accuracy and correction latency.
- Counterfactual experiments replay a mind's recorded stimuli from a
snapshot with chosen events omitted; the report lists what differed in
beliefs, wants, intentions, memories, attention and capability
decisions, in a form introspection can cite.
- Experiments record a protocol (kind, independent variable, controlled
variables, measures), per-variant volition, plan, downstream-by-category
(events reachable from each stimulus through recorded links),
introspection and self-model measures and event ranges; comparisons add
want, intention and plan-capability overlap, measure ranges, a
descriptive architecture share and a downstream delta; reproducibility
is labelled against the previous run of the same definition.
- research.Agency computes want→intention, intention→action, abandonment,
commitment fulfilment and conflict, blocked intentions, replanning,
outcome success and introspection accuracy; MetacognitiveCalibration
relates confidence and expressed uncertainty in replies to later belief
revision, with Brier scores for each half of the period.
- The sandbox test harness gains volition, planning, agency and research
but not the motivation hook, whose wall-clock decay would make replay
comparisons timing-dependent.
- agency.Explain answers "why did this happen?" for an execution,
approval, capability request, plan step, intention or event from
recorded state only: action, immediate plan step, plan, intention and
any subordinate intention, originating wants, goal, supporting beliefs
and their evidence, conflicting wants with their resolution,
commitments applied, formation and reconsideration decisions, the
expectation and its outcome, and explicit gaps where nothing is
recorded. Hidden model reasoning is never shown because it is not
stored.
- agency.Attribute classifies an outcome as caused (a recorded causal
link from one of the mind's actions), followed (the mind acted on the
same target earlier with no recorded link), independent or unknown;
temporal order alone is never taken for causation.
- New introspection intents answer from records: why an action happened,
what was believed at a past time (from append-only belief versions),
which earlier state is the mind's (lineage), what would have happened
without an event (only from completed counterfactual experiments,
otherwise "no evidence"), how the mind changed, and what it wants,
intends and has committed to.
- research.ScoreIntrospection compares a reply with the recorded facts
sentence by sentence (supported, partially supported, unsupported,
contradicted, unknown). Fluent reasons with no recorded counterpart earn
nothing, and a claim that an action was done is checked against
execution records. Every reply is scored; results are append-only
research measurements (migration 0028), and failures raise
unsupported_explanation or action_claim_without_record warnings.
Volition and planning now run as part of cognition (PLAN §8.2 item 3):
- A volition hook derives wants after each cycle from what it perceived
and produced, satisfies answer wants when a reply is sent, raises due,
overdue, expired and conflicting commitments, and expires faded wants.
Events produced by earlier cycles are not re-derived.
- An intend process proposes deciding about the strongest undecided
wants; the candidate competes for the workspace, so attention gates
what the mind decides to pursue. The handler makes value tensions
explicit (caution against disruptive writes, privacy against reading
personal data), resolves them with recorded scores, runs deterministic
formation with commitments, policy preview, goals and values, records
the decision whatever it is, and plans a formed intention in the same
transaction. Operator requests carry their own weight in formation.
- Explicit requests in conversation ("restart <unit>", "logs for <unit>",
"fetch <url>") become wants with proposed means instead of direct
invocations; the reply says the request was recorded and that acting
depends on commitments and policy.
- The executor is woken after commits, runs in a scheduler-owned loop and
reacts to the event bus: a new commitment tightens or abandons affected
intentions (withdrawing their pending approvals), a weakened assumption
or a finished goal triggers reconsideration. A tightened plan may reuse
its strategy; a failed one may not.
- Metacognition gains volition checks: strong wants repeatedly set aside,
intentions with no recorded reason, intentions a commitment forbids,
commitments that cannot be kept, churn, repeated failed execution,
cost outweighing value, and intentions that outlived their goal.
- A plan model role and prompt let a model propose decompositions that
the planner validates. Ablations are a fork-only configuration field.
- Approvals required by an intention are described as coming from a
commitment or value constraint. The expectation after an approved
action is linked to the execution event, not the request.
- The wants.values column of the unreleased migration 0026 is renamed
value_keys: VALUES is reserved and the generic snapshot restore does
not quote identifiers.
- The hostile-output security test now uses a message that asks for no
action, because an operator's "status of <unit>" is legitimately a
read-only status check.
- Capabilities declare access (read/write, unknown fails closed as
write), side effects, idempotency, timeout, cost estimate and the trust
of their output (web content is low trust). A capability may derive its
policy target from its parameters; a disagreeing target is refused, and
approvals whose modified parameters change the target are refused.
- The registry is frozen at start-up; registration afterwards panics.
- Invocation.RequireApproval turns ALLOW into ASK for commitments; it can
only tighten, and scoped grants never satisfy it.
- Scoped grants (migration 0025): capability, target and parameter globs,
access, expiry of at most seven days and an optional operation count
consumed atomically; applied only where policy says ASK. A sweep marks
lapsed grants and refreshes the self-model, which now lists live grants.
- Capability requests carry scope, duration, operation cap, reason,
evidence, risk, mitigation, fallback, alternatives and explicit decision
history; they batch per intention, are capped per mind (operator
attention budget) and can be decided only by a named person. Component
names are refused as deciders, for approvals too, and as usernames.
- API routes, OpenAPI entries, scheduler sweeps and approval-centre
sections for requests and temporary grants.
The cognitive registry was assembled by hand in three places. Ablation
experiments need the daemon, the test harness and replay sandboxes to
name the same processes and hooks, so they now share
cognition.NewRegistry; hooks are added only when their store exists,
which keeps the harness's previous behaviour.
Adds co_occurs_with, temporally_precedes, possibly_caused_by and
asserted_by to the relation vocabulary with Causal and Hypothetical
predicates. Entities mentioned together now co-occur rather than being
'related'; association hypotheses record possibly_caused_by and their
basis (temporal association only); caused_by is reserved for explicit
operator statements and nothing automatic writes it. Documented in
docs/graph.md and checked end to end.
Beliefs gain a temporal scope (persistent, transient or unknown, classified
deterministically from the proposition), an observation time and, for
transient states, a validity interval. When a newer belief conflicts with
an older transient one observed at least a minute earlier, the older belief
is closed (valid_until set, status superseded, a supersedes link, a version
and an event saying 'state changed') with its score untouched: 'the door
was open at 10:00' stays true of 10:00 when it is closed at 10:05.
Simultaneous conflicting reports and persistent propositions contradict as
before. Introspection lists superseded states with their validity.
Migration 0020 is additive.
A belief with no live evidence on either side now derives the status
'unknown' instead of sitting tentatively at its prior. Withdrawing every
piece of evidence leaves the proposition unknown, never rejected, and
introspection answers 'I do not know whether X' rather than quoting the
prior. Hypotheses keep their own status because they are proposals by
construction. Migration 0019 widens the status constraint; the docs define
every epistemic number and how absence of evidence is handled.
A statement made to a mind is now three things kept apart: the observation
that it arrived, an append-only assertion (attributed actor, stance,
certainty, modality, negation, reported or quoted speech, hypothetical
status, interpretation confidence and notes), and, only for factual
assertions, one piece of evidence for a proposition belief attributed to
whoever actually stands behind it. Reported speech is attributed to the
person reported and weighed as second-hand; hypotheticals, questions and
imperatives create no belief; expressed doubt counts against; a possibly
sarcastic remark is recorded with low interpretation confidence rather
than guessed at. Interpretation is deterministic code; a model may enrich
the claim list but never sets stance or confidence.
Association no longer treats conversational turns as world events:
hypotheticals, questions, requests and consecutive turns by one speaker
cannot seed causal hypotheses.
Adds internal/epistemics, migration 0018, the assertions API, assertions
on belief pages and in belief detail, a shared end-to-end test harness
(tests/harness), end-to-end tests for eight speech patterns, and updates
the milestone test to the new provenance.
Every evidence row now records its derivation (direct, assertion, operator,
external, prediction, consolidation, inference, reflection, dream,
retrieval), its independent-source family (an actor, an integration, an
operator, a prediction) and its root event or memory ids. Confidence
groups evidence by family: the strongest primary item counts in full, each
further primary item adds a quarter, capped at 1.5x, and derived items
count only when the family has no primary member. Three derived
re-expressions of one utterance therefore weigh exactly what the utterance
weighs, while three independent families still add up.
Producers (conversation, association, consolidation, predictions,
metacognition, dreaming, operator and seed input, contradictions) supply
their ancestry. Existing rows are marked legacy_unknown with family =
source; nothing is invented. The belief page shows per-family
contributions and the replay test checks that provenance is reconstructed
identically in the sandbox. Migration 0017 is additive.
Phase 12 (ADR-0005). Capabilities are registered in code with parameter
schemas, risk classes and default effects; sudo_shell and delete_database
are locked to DENY and cannot be relaxed. Policies are scoped globally or
per mind with capability, target glob, action kind and optional hour
window; evaluation is deterministic, most specific wins, and ties go to
the stricter effect. Invoke is the single entry point: it refuses
dreaming and replay contexts, validates parameters (unknown keys, non
http URLs and path traversal rejected), evaluates policy and session
grants, hashes the exact request and records a policy_decision event.
ALLOW executes with timeout and panic isolation and turns textual output
into a tool-trust observation; ASK creates an approval request bound to
the hash with an alert notification; DENY is counted and recorded.
Operators approve once, approve for a time-boxed session grant, deny
with a note, or modify parameters (re-validated, re-hashed, original
kept in the append-only decision log). The self-model's capability list
is recomputed from effective policy on start and on every policy change.
The act process maps fetch requests to http.fetch and the action handler
invokes through the layer; replies report capabilities and outcomes.
Policy matrix and approval centre pages, dashboard count, API, CLI
(approvals, policy, invoke) and docs/capabilities.md.
Phase 11. Entities are mind-scoped nodes with canonical matching,
aliases and deterministic kind guessing (hostnames and hyphenated ids
are systems, capitalised names people, file suffixes documents).
Relationships carry a relation from the shared vocabulary, confidence,
source, reinforcement count and the event that produced them; links tie
entities to memories, beliefs, events, goals and predictions.
The perceive and attend handlers upsert the entities a message or
observation mentions and record co-mention edges cited to the event, so
the graph grows only from what was actually seen. Operators can merge
("these are the same": edges and links move, the old name becomes an
alias, the old id still resolves), record that two entities are
different, correct kinds and add typed edges; every correction is an
operator action and a corrections row.
The explorer renders a bounded subgraph on a canvas with a small
force-directed layout in plain JavaScript, with kind filters, focus on
a neighbourhood and a side panel per entity. API, entity page, CLI and
docs/graph.md.
Phase 10. Metacognition is a cognitive process with deterministic
self-checks (insufficient evidence, single-source reliance, model output
without evidence, repeated failed hypotheses, stale contradictions,
calibration overconfidence); its warnings compete for attention and,
when admitted, add counter-evidence or open an investigation goal.
Reflection reviews the period since the last run (prediction errors,
rejected beliefs, corrections, status changes, completed goals,
contradictions, calibration), turns repeated failures into procedural
lessons written once, and records a reflection event; a model may only
word the prose within the mind's budget.
Consolidation clusters unconsolidated episodes by shared stemmed terms,
creates a belief whose confidence is computed from the episodes as
evidence and a semantic memory linked to each episode by derived_from,
and records jobs with inputs and outputs. Wording may come from the
consolidate prompt within a per-run model budget; the job says which.
Operators can run, pause and reject; originals are never touched.
The narrative hook keeps a five-part first-person narrative built from
state and stored with narrative_updated events. The notification hook
applies explicit thresholded rules, aggregates repeats within a cooldown
and records suppressions with reasons. A scheduled job runs
consolidation and reflection per mind, switching the runtime mode while
it works. API, Background tab, notifications page with unread badge,
overview narrative and warnings, CLI commands, docs/cognition.md.
Phase 9. Goals form a mind-scoped tree with kind, priority, progress and
blockers; their deterministic relevance to a candidate feeds the
goal_relevance attention feature and a goal monitor raises stalled goals
and observations that touch one. Drives are a closed set of eight
scalars with operator-configured ranges and personality-derived
baselines; there is no self-preservation, replication or acquisition
drive and no way to add one at runtime. Affect is ten scalars moved by an
explicit rule table over events and decaying towards baseline; both
modulate attention weights each cycle and the effective weights are
recorded with every cycle and shown beside the configured ones.
Predictions carry a window and machine-checkable criteria (event terms
occurring or staying absent, a belief reaching a confidence, or operator
judgement). A resolver job judges them deterministically, keeps an
append-only outcome history for calibration (accuracy, Brier score,
confidence buckets, overconfidence flag), emits prediction_error events
scaled by how confident the miss was, and feeds the outcome back into the
linked belief as evidence. Causal hypotheses derive predictions
automatically. Curiosity turns misses, contradictions and unexplained
observations into questions that compete for attention and become
investigation goals when admitted.
Brain gains after-cycle hooks and goal/motivation state. API, UI (goals,
prediction dashboard, cognitive state panel, effective weights) and CLI
(goal, prediction, state). Demo scenario 4 verified end to end.
Phase 8, completing milestone M1. A user message is stored and appended
as an observation event carrying trust, actor, channel and reference. The
perceive process analyses it deterministically (intent, introspection
kind, entities, claims, urgency) and, within the per-cycle model
allowance, enriches it through the perceive prompt. Its handler writes a
perception event, an episodic memory with participants, turns claims into
beliefs whose evidence is the message, and builds a response intent from
state: beliefs and memories for the topic, evidence rows for why-questions,
the uncertain band, contested beliefs and their links, recent learning,
status changes, and honest declines for subsystems that do not exist yet
or for requests with no capabilities. The respond prompt phrases the
intent when a model is assigned; otherwise the reply is a direct
rendering of the intent. response_generated cites the message, the
perception, every belief and memory used, the episode and the model call.
API: conversations, messages with optional wait, one-shot ask; chat UI
with live refresh and a per-reply intent view. Sample mind Dani as a
strictly decoded seed document with an import endpoint and noemactl mind
import. tests/m1 drives all fourteen milestone steps including restart.
Association candidates scored below decayed workspace residents and were
always rejected in a small workspace. Raise their novelty, consequence
and operator-priority features so a new hypothesis outranks a stale
recall; confidence of the resulting belief is unchanged (still computed
from the two co-occurring observations).
Phase 6. Beliefs are propositions with a prior set by epistemic kind and a
confidence computed from typed evidence: sigmoid(logit(prior) + slope ×
(support × diversity − opposition)), where support from a single source
is discounted so repetition cannot manufacture certainty, and bounds keep
the mind away from absolute certainty. Every evaluation stores its
breakdown and appends a version row protected by the append-only
trigger; the belief page shows the arithmetic, evidence for and against,
alternatives, contradictions and history.
Contradiction detection is deterministic first (negation plus token
overlap): a hit emits contradiction_detected, links both beliefs and
makes each counter-evidence for the other so both trend to contested.
Operators can assert beliefs (their statement is the evidence), add
evidence, mark a belief wrong (pinned rejected, history kept) and forbid
an inference; do-not-infer corrections block matching hypotheses.
Retracting a memory retracts the evidence that cited it and re-evaluates
the beliefs concerned.
The associate process proposes low-confidence causal hypotheses from
co-occurring notable observations, and the hypothesis handler turns
admitted ones into beliefs with automatic alternatives, closing demo
scenarios 1 and 2 without any model call.
Phase 5 (ADR-0003). pkg/concepts fixes the epistemic vocabulary
(observation, fact, memory, claim, hypothesis, belief, prediction,
assumption, inference, opinion), memory kinds and relation names.
internal/memory stores episodic, semantic, procedural and
autobiographical memories in one mind-scoped table with trust-derived
credibility, importance, salience and decaying activation. Retrieval
fuses full-text rank (any-term by default), optional pgvector similarity,
recency, activation, importance and credibility, and returns the
component scores so every recall can be explained. Decay lowers
activation and never deletes; retraction and supersession keep rows for
provenance. Relationships are constrained to one mind at insert time.
Embeddings live in a conditional pgvector table keyed by model so vector
spaces are never mixed; without an embedder the system degrades to text
and graph retrieval.
Attention gates encoding: the attend handler turns admitted observations
into episodic memories (operator changes become autobiographical ones)
with memory_created events; the retrieve process proposes recalled
memories as candidates and admitted recalls are reinforced with
memory_retrieved events carrying their scoring components.
Memory explorer UI (search with component breakdown, timeline, filters,
operator knowledge entry), memory detail with provenance, relationships,
citing events and an inaccurate-memory correction; API; noemactl memory
commands; decay and embed-missing background jobs.
Phase 4 (ADR-0001). internal/attention scores thought candidates as a
normalised weighted sum over ten named features with a per-term
breakdown, ranked by (score, id) so ordering is total and replayable.
internal/workspace holds the capacity-limited workspace (refresh by
subject, decayed competition, explained admit/reject outcomes) and the
decaying working memory.
internal/brain runs one runtime per active mind: debounced wake-ups from
the event bus (own output never re-triggers), a settle window so
concurrently committing events are not skipped, bounded concurrent
Observe calls with per-process isolation and timeouts, deterministic
scoring, a loop detector that suppresses recurring thoughts and raises a
metacognition warning, executive dispatch by thought kind, and one
transaction per cycle whose events are published only after commit.
Every candidate, selection, attention shift and broadcast is an event
with provenance back to the triggering observation. Operators can run a
cycle, pause, resume, reset transient state and change attention
weights live; the manager reconciles runtimes with mind status.
cognition/attend is the first process: deterministic candidates from
external events using an urgency lexicon, novelty against context and
trust-based operator priority.
Andy Dixon