From 7c8e8322fd4fbb503c9ee38134f4d94486a76d00 Mon Sep 17 00:00:00 2001 From: Meta Vens Date: Sat, 2 May 2026 15:41:40 +0100 Subject: [PATCH] Document VoidBot appraiser training goal --- .../soft-model-training-artifacts.md | 19 ++++++++++++++--- docs/architecture/training-plan.md | 21 +++++++++++++++++++ state/evidence.jsonl | 1 + state/map.yaml | 8 +++++++ 4 files changed, 46 insertions(+), 3 deletions(-) diff --git a/docs/architecture/soft-model-training-artifacts.md b/docs/architecture/soft-model-training-artifacts.md index 5a21085..ac2688a 100644 --- a/docs/architecture/soft-model-training-artifacts.md +++ b/docs/architecture/soft-model-training-artifacts.md @@ -206,7 +206,10 @@ state, memory, affordances, and known facts for the current turn. ### Appraiser Reads an observable event from one participant's local perspective and proposes -how that participant interprets it. +how that participant interprets it. This organ is not only for fictional scene +mutation. It is also one of the core paths by which VoidBot becomes better at +reading people: noticing what a human did or said, forming a grounded opinion +without pretending certainty, and revising that opinion as evidence accumulates. Training artifacts: @@ -218,10 +221,16 @@ Training artifacts: - relationship implications - belief implications - memory candidate +- human-facing opinion update when applicable: trust, concern, irritation, + respect, caution, attachment, boundary pressure, expected future behavior, and + what evidence would change the read - reviewer corrections The appraiser should not know canonical private state except where that -participant knows or believes it. Fundamental attribution bias belongs here. +participant knows or believes it. Fundamental attribution bias belongs here. For +VoidBot, this means the model must distinguish an observed human behavior from a +durable claim about who that human is, and should preserve uncertainty, repair +signals, and evidence needed to revise the opinion later. ### State Mutator @@ -243,7 +252,11 @@ common deltas, but deterministic gates still own schema, bounds, and authority. ### Relationship And Social Perception Classifier -Updates perceived overlays and relationship stance from repeated evidence. +Updates perceived overlays and relationship stance from repeated evidence. This +classifier should serve both Aetheria agents and VoidBot's long-running human +relationships; the same machinery that lets Sella update her read on Maer should +also let VoidBot update its read on a specific user without collapsing into +either sycophancy or a permanent first-impression mugshot. Training artifacts: diff --git a/docs/architecture/training-plan.md b/docs/architecture/training-plan.md index 4be80e2..9ca8c23 100644 --- a/docs/architecture/training-plan.md +++ b/docs/architecture/training-plan.md @@ -9,6 +9,14 @@ resource math, source provenance, Ink compilation, and mutation gates. Models handle judgment, language, social inference, prioritization, and coordination. Every model-facing judgment should leave a reviewed artifact. +Ghostlight is also the training ground for VoidBot's social intelligence. The +appraiser and relationship/perception organs must remain useful for reading real +human interaction over time: forming grounded opinions, noticing uncertainty and +misread risk, tracking trust, irritation, concern, respect, boundaries, and +expectations, and revising those impressions when later evidence contradicts +the first read. If these organs only work for fictional NPC scene mutation, they +have missed one of the original reasons the project exists. + ## Timeline Coverage Ghostlight training should cover three timeline lanes: @@ -688,6 +696,9 @@ Inputs: - cultural priors - scene pressure - known facts and suspected facts +- human interaction history when the appraiser is being used for VoidBot +- conversational tone, timing, repair attempts, recurring preferences, and + boundary signals for human-facing appraisal - post-Rupture concept being interpreted, if any Outputs: @@ -699,6 +710,9 @@ Outputs: - proposed memory candidate - proposed relationship implications - confidence notes +- human-facing opinion updates when applicable: trust, concern, irritation, + respect, attachment, caution, boundary pressure, expected future behavior, + and what would change the read - concept-specific misread labels, for weird Elysium phenomena Training architecture: @@ -714,12 +728,17 @@ Training artifacts: - emotional appraisal labels - rejected interpretations - reviewer rationale +- VoidBot-facing human interaction appraisal receipts, including what was + observed, what was inferred, what stayed uncertain, what opinion changed, and + what evidence would reverse or soften the read Pilot schema shakedown gate: - 300 participant appraisal examples - balanced set of correct reads, misreads, ambiguous reads, and biased reads - examples include fundamental attribution bias and confidence movement +- at least 50 human-facing VoidBot appraisal examples before claiming the + appraiser is useful for companion-style memory, opinion, or user-read updates - future-branch examples include socially grounded misreads of post-Rupture phenomena @@ -1095,6 +1114,8 @@ A first appraiser/classifier training run wants: - 300 participant appraisals - balanced misread/correct/ambiguous/bias examples +- at least 50 VoidBot-facing human interaction appraisals with opinion-update + labels and uncertainty notes - labeled confidence drivers - reviewer rationale diff --git a/state/evidence.jsonl b/state/evidence.jsonl index b27f802..f9f5bdb 100644 --- a/state/evidence.jsonl +++ b/state/evidence.jsonl @@ -20,3 +20,4 @@ {"ts": "2026-05-02T11:07:05+00:00", "type": "implementation", "status": "ok", "note": "Research-enabled responder captures now have a first-class research_trace seam. Accepted research-enabled gold captures must use runner_captured trace status; existing Sella research capture is labeled coordinator_reconstructed draft audit data."} {"ts": "2026-05-02T11:19:14+00:00", "type": "implementation", "status": "ok", "note": "Generated the first accepted runner-traced coordinator-scoped responder sample: Veyr offers a deniable handling category in scene-03 Callisto provenance. Artifacts include projected context, packet, accepted capture, mutation receipt, and updated aftermath state."} {"ts": "2026-05-02T14:32:15+00:00", "type": "implementation", "status": "ok", "note": "Accepted trace-backed Isdra convoy-threshold responder sample: limited-contact exception under freeze, Navigator ledger signature, isolation/decon constraints, packet-sound preservation, reviewed mutation, and after-isdra state."} +{"ts": "2026-05-02T14:41:30+00:00", "type": "architecture", "status": "accepted", "note": "VoidBot social intelligence is a first-class Ghostlight consumer: appraiser and relationship/perception training must support human-facing reads, grounded opinions, uncertainty, repair signals, and opinion updates over repeated interaction, not only fictional scene mutation."} diff --git a/state/map.yaml b/state/map.yaml index 7fc05eb..fc2e32e 100644 --- a/state/map.yaml +++ b/state/map.yaml @@ -17,6 +17,12 @@ mission: Produce sandboxed, source-grounded responder training data through projected local context, responder packets, reviewed outputs, and audited mutation receipts. + voidbot_goal: > + Keep the appraiser, relationship/perception updater, and memory machinery + useful for VoidBot's core companion mission: reading specific humans over + time, forming grounded opinions about them, noticing misreads, and updating + trust, concern, irritation, respect, boundaries, and expectations from + evidence rather than generic friendliness. active_fixture: > Cold Wake is a pre-Elysium historical training fixture, not the product target. Call of the Void and Elysium-era gameplay remain procedural @@ -27,6 +33,7 @@ hard_boundaries: - Canonical agent state and per-agent perceived state must remain distinct. - Characters act from local knowledge, local perception, local pressure, and bounded affordances. - Listener misread is first-class; intent and interpretation are not the same object. + - Human-facing appraisal is a first-class target; VoidBot must be able to read people, form grounded opinions, and revise those opinions over repeated contact. - Generated action is an event proposal; deterministic gates and reviewed mutation decide what becomes state. - Fuzzy judgments become reviewed training artifacts before they become trained models. - Aetheria generation is source-first; check lore for factions, bodies, locations, institutions, and constraints before projection. @@ -59,6 +66,7 @@ live_architecture: - Participant appraiser - State mutator - Relationship/perception updater + - Human-facing appraisal and opinion model for VoidBot - Evaluator/review classifiers - Aetheria-tuned responder - Institution, faction, and consumer decision models